> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
Fatal misunderstanding.
Datacenters are not warehouses. Amazon.com being borderline-profitable owing to massive ongoing capex on warehouse expansion is not at all like AWS becoming borderline-profitable owing to massive ongoing capex on LLM datacenter build-out and SOTA model training (it's CAPEX, but carries OPEX affect because of GPU depreciation schedules and competitive hardware upgrade cycles), and opex of delivering LLM inference services, and being able to price it all competitively.
(edit: expand explanation of why it's a misunderstanding)
There's also a failure to misunderstand the arms race nature of it all and the need to be in the lead since everyone is assuming it's a winner take all market. That's probably close to correct.
A data center is a warehouse for sure but the building is only a few percent of the cost. The GPUs are probably 80% of the total cost or more. And they probably only have a 3-5 year primary lifecycle.
Pay the building off over 20 years? Sure but 80% of your capex problems remain. 2 trillion already spent? Just those GPUs need to generate 500 billion a year over the next 5 years.
> everyone is assuming it's a winner take all market.
I guess this is the big gamble, the idea that if you just spend enough money as fast as you can you can become the gods of AI forever and can just charge whatever you want for your magic beans.
But what is the moat? If in reality these things become commodities and just compete with each other to spit out cheap tokens then there's no way the margins are there.
> everyone is assuming it's a winner take all market
I don't think anyone serious believes this any more. The tech is thoroughly commoditized. There are a lot of plays to be made, like OpenAI cornering memory, or spending heavily on unique data for RL, if you can stop distilling, but there is no proprietary core IP at all.
There is no way to monopolize a technology this big. It would be like monopolizing literacy.
It's not the AI technology itself that is being monopolized, it's the data necessary to train it.
More specifically: when everything you could previously have earned money with is subsumed by AI, the "frontier" of profitable knowledge exists solely with the highest concentration of capital. Because that's where the highest amount of compute and consequently valuable data is.
Already now the idea of the tech being "commoditized" is true only if you exclusively look at the bubble of rich countries. AI-worthy compute is prohibitively expensive otherwise for individuals.
Power, natural gas, gasoline, water, mobile contracts, railways, internet access are probably better analogies. They have tendencies to (local) monopolies but also a lot of precedent and knowledge on how to regulate and break them, how to introduce competition.
> and can just charge whatever you want for your magic beans.
'Winner take all' doesn't mean convincing others to buy your product. If others have something to buy your product with, you've not taken all, fundamentally.
'Winner take all' means having the most powerful army. The assumption is that once AI crosses a certain threshold then there won't be anything the rest of us can do to stop the machine of whomever gets there first, so they will truly be able to take all. Hence why the AI labs sometimes hint about the risk of human destruction.
The army interpretation exists, and might sadly be relevant here, but it's not the only one. Google, for instance, won a nonviolent winner-take-all competition for search.
You're right that Google has been greatly helped by having what has been considered the most powerful army (the US army) backing them. They wouldn't have had a hope in hell of surpassing the "mom and pop" stage without that.
But I am not sure you are right that it is powerful enough to secure all. I search daily, but cannot remember the last time I used Google. The other AI companies do it better. And I'm not alone in that sentiment. Google's stock price recently took a nosedive because investors have grown quite worried about the competition. Without a sufficiently powerful army stopping them, there will always be someone else ready to try to outdo you.
> since everyone is assuming it's a winner take all market. That's probably close to correct.
why? I keep seeing this over and over again but I've not seen an explanation. I would really appreciate some insight into this line of thinking, since it's not obvious to me. What is the benefit of holding hundreds of billions of computer hardware when it's not proven that it's even needed.
Lets say the total market for an arbitrary winner-take-all industry is $100 Billion/year in profits. Just so that we have a round number. And for another round number, lets assume the fair market PE ratio is 20x. This over-simplification would then have a fair market valuation of $2 Trillion total.
Just simple napkin math.
Now, lets say there are 5 companies competing to be the winner. Only one can be the winner, but all 5 are trying to be the winner. How much should each company individually spend?
Well... starting-bid is $2 Trillion, or ~$10 Trillion in the aggregate. Its basically the waste of competition. Because everyone is vying for monopoly power, the only fair for them to all invest over $2 Trillion in the hopes of being the dominant one at the end. Even if $8 Trillion is wasted, that's "other companies", your competitors, who lose out.
But wait: if one company spends $2.1 Trillion, that one company likely becomes the winner and the other 4 companies lose out. So then another company spends $2.2 Trillion, to get ahead of that war. Etc. etc.
The idea is, innovation will concentrate where the necessary ingredients are.
When AI replaces human workers and researchers, the "frontier", with which you can earn money and gain an edge over your competition, resides in those data centers with the highest concentration of computational power.
If you own those data centers, you control the frontier of human knowledge, the edge where money and power is being carved out.
You got me on that one, geez! I'm not going to edit it though because I think you're right. Nobody really knows if there's going to be a winner-take-all or not. During the arms-race portion though people are definitely acting like it will be that way. But watching how fast models catch up to each other it's not obvious it will play out that way.
The first one says that networking infrastructure costs 13% and the other says 1.5%. Not a great start.
Let's imagine that a traditional full rack can host 80ish H100s and that each one costs $25k. That's 2 million right there. There's probably only 10x cases in that rack so that's 10x dual core motherboards, RAM, network, etc. Call them 20k each. Maybe you need a 32 port 100gig fiber switch, that's 8k. Let's say it's 2.3 million per rack and 2 million of it is just the GPUs. We're at 85% of the rack cost right there in GPU, but the other stuff in the rack isn't durable either.
I suspect that it would take more than just another 100k per rack to finish out a data center so my "80% of the cost is in the GPUs" is definitely wrong. But I didn't see anything in these articles that indicated to me that 50% of the cost of a datacenter is anything really durable like buildings or electrical infrastructure or cooling. One rack might only take up 20ish sqft itself another 100sqft of isle. These buildings don't cost $10k/sqft to build the envelope. There's still an enormous amount of cost in stuff whose performance follows more of an exponential performance curve, not linear. It's just an absolute nightmare to try and amortize exponential performance curves with linear repayment.
These days there's usually two networks. Multi-trillion parameter models are too big to fit on one box of 8x H100's, so many of these are running RDMA (800Gb/sec on GB300 series) for rack-scale/many row-scale inference.
For GB300 series, pricing along the lines of $500k/server, 9 servers/rack ($4.5M/rack) before looking at water cooling, 800Gb RDMA network, and your 100Gb IP network
but is it a winner takes all market? currently we have models that are close-but-not-quite SOTA which are still heavily used and very profitable because they are _good enough_ (i'm talking about deepseek, of course).
or is the "winner takes all" mentality based on the assumption of being the first to achieve the technological singularity?
I read somewhere that the expected 3-5 year lifespan of inference GPUs was not holding up to scaring: in practice, the GPUs have the same lifespan as any datacenter CPU or RAM. I think that estimate was based on Bitcoin mining ASICs, which wasn't an apples-to-apples comparison.
Older GPUs still have an active secondary market but they are probably not profitable at market rates for inference because newer chips are so much more efficient.
If Semi-Analyses is correct [1], even H100s would lose money at current prices - I doubt that is actually true of H100s in practice because from what I can tell it is based on a 3-year depreciation from the new price, but the point is that there is a point at which old hardware's operating inefficiencies dominate.
I'm not certain it's the longevity of the actual hardware so much as there is a huge incentive to replace your hardware with faster, more capable equipment. Who wants to invest in the AI company that's using 10 year old GPUs when the next company is using brand new?
Old GPUs today are selling and renting for more than they were new which suggests the typical depreciation pattern isn’t holding.
There’s a risk that likely will happen that once demand is met capacity will overshoot significantly. However, when will that occur is very up in the air. Everyone is training every larger models and there’s more AI usage than ever
There’s still a cost to underestimating asset depreciation from an accounting perspective. Imagine being left holding the balance sheet when you budgeted for X amount of depreciation but the demand drops significantly more than you expect during the next fiscal year. I guess my point is that artificially constrained supply is not a free lunch.
Power generation is also extremely expensive ongoing - either through gas turbines that eat fuel or on-grid power that datacenters will need to pay to capitalize (depending on ratepayer agreements of course - but clearly people aren’t stoked about subsidizing these)
In aggregate it’s a lot of electricity, enough to impact local grid markets.
But as a fraction of inference revenue, there is good reason to estimate that the direct electricity costs are perhaps 1-5% of inference revenue.
That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
> That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
But that’s a feedback loop that could lead to more supply being built. Or it could lead to regulations that put enormous surcharges and low limits on data center power consumers, which effectively shield consumers directly from the lack of supply. Which in turn removes the incentive for consumers to expand power generation capacity to keep up (or to allow data centers to expand it themselves).
There is enormous opposition to a new data center project 30 miles outside my city because they want to build a 1000MW natural gas power plant and a 500MW biomass plant (that will take millions of acres of forest to sustain).
If that opposition continues, I see little chance of anything but electricity cost rising as a percent of inference.
The problem is likely the GPU price, NVidia has a very large margin - but it doesn't seem like the the use case supports it. If NVidia cut the GPU price by 2-4x then these problems likely go away.
I wouldn't be surprised if most hyperscalers are betting on cheaper GPUs on their next upgrade through either pricing pressure on NVidia, competition from AMD/Broadcom or in-house chips.
Normally you'd replace the device because there is more power efficient one on the market so running one at current energy prices costs more than buying new, more power efficient one and saving power.
IF prices of GPU/RAM come back to pre bubble costs, we might see the 5 year cycle of replacement that's relatively common in normal datacenters pre-ai bubble.
At current prices ? those GPUs that are there now will be running there in 10 years still.
I'm of two minds on this issue. Off the cuff I agree, but then again where are all the used "enterprise" GPUs coming from right now? There's plenty of secondary market for used Telas K, M and V series GPUs right now and the V100 is only 9ish years old, and that only if you bought it on the release day.
I do agree that 5 years isn't the hard cutoff I was suggesting but I'm not convinced that 10 years is a reasonable timeframe to consider paying the debt down when Moore's Law is involved. I know it's no longer the 12-18mo doubling it used to be but even if it's 2ish years then over 6 years you get 3 doublings or 8x performance gains. At 10 years it's another 2 doublings or 32x performance gains. It's hard for linear payback curves to deal with exponential performance curves past 2-3 doublings.
> At current prices ? those GPUs that are there now will be running there in 10 years still.
If they don't burn out after two or three. Annualized failure rates are about 9% according to Meta.
A quarter will have failed within three years. Less than half of the original hardware will be running after ten years. In both cases if you assume that rate does not increase over time. Do H100s come with replacement guarantees?
The gamble: gunning for market supremacy before investor cash dries up.
The real gamble: betting there is a large enough pool of c-suite that doesn't fundamentally grasp the concept of vertical integration that they're willing to pay to train their competition. Because if you think AI companies aren't training models how to do everything your company does with your data you are deeply delusional.
The realreal gamble: praying sovereign governments stay anesthitized to the national security implications of a handful of extra-national organizations operating the conputer security equivalent of WMD with no oversight.
> And they probably only have a 3-5 year primary lifecycle.
I think in 3-5 years close to 100% of current GPUs/RAM will still be in service. New hardware matters only if gains in efficiency are large enough to put you out of business because the competition gets them. Extreme case of that are bitcoin miners. With AI inference there are way too many factors that influence the profitability. It's not like you can't run new model without the new hardware. It can just run on more of the old one. 3-5 years for GPU are estimates when the main application was gaming with new games with new requirements for running on a single GPU being constantly released.
Thanks for pointing that out. That has been my concern as well. Sure the post says "they make $2.08 for every $1 of compute" but that seems to only be accounting for OPEX. If we start accounting for CAPEX and variations with interest rates things look a lot less rosy. I mean a lot of folks are suspecting Anthropic will get a junk bond rating when it rolls out to public markets.
> A typical one-gigawatt AI data center requires $38 billion in up-front capital expenditure (CapEx) and $0.9 billion in annual operating expenses (OpEx). If CapEx is annualized over each asset’s lifespan, the total cost of ownership equates to $8.5 billion per year. Servers dominate this cost at $5 billion per year, or 60% of the total. Operating costs are small by comparison: even energy, the largest OpEx category, costs only $0.6 billion per year.
This is my heuristic for the capex split and effective lifetime of the asset it creates (I'm no expert of course, this is based on various reading):
- 20% Capex has ~20-year lifetime: The shell (land, building, cooling gear, rack mounts, power lines etc.).
- 80% Capex has ~3-year lifetime: The computers, including bulk commitments to power and maintain them.
And interest rates are why this split is a ticking time bomb. The sheer quantum of investment to build one is almost entirely debt funded (~90% of total outlay). I would guess the endgame would be bad even if the split for these entirely different asset classes is 50/50, which would be very generous, based on what I've understood.
Datacenter funding is extremely leveraged. For comparison, only some banks are allowed that much leverage for that size of capital, that too only if they have federal reserve backing. Even in such a case, the retail account holder (pseudo-investor, because our deposits are loans to someone else) has practically no downside protection. If the bank collapses you will lose everything.
They're paying way under 20 cents a kWh. They're paying less per kWh than residential households. If the residential rate in an area was 20 cents, I'd guess the datacenter is paying 40-60% of that rate
They get bulk discounts because their load is 24x7, so the fixed costs covered by a per kWh charge is spread over many more kWh than a residential customer
Even if we allow for the strange “it’s only a one-time cost” hypothesis, the author’s next paragraph outlines another actual bubble, just 5-10 years from now:
> Once all the data centres are built, AI spending is going to plummet.
Please, Giancarlos. Let's all behave with the times... I think it's culture-appropriate to say "You're absolutely right. Datacenters aren't more clients. We build. What if they don't come? It's not just a one-off. It's a massive gamble."
To be clear: it's a tongue-in-cheek riff, macroexpanding Giancarlos's (I think humorous) comment. I don't mind the downvotes---acerbic / black comedy is risky business :sweat-smile:
To be fair, I could be wrong, we could all be wrong, we don't know what we don't know. On the other hand, I do think they should come up with a better plan than just trying to retrofit them everywhere they can find. I do think some states make more sense than others.
We will know once it all unfolds, on the other hand, at least there's people who are getting paid to build these.
Worse case, we have acccess to (hopefully) really well-built data centers that can be used for anything hosting related in the future.
I, for one, welcome the coming age of the post-LLM-datacenter-overinvestment-bust-fueled backyard GPU supercomputer revolution.
(edit: in which case, I'd say more people would get paid to build more of these, except not as obnoxious township-destroying behemoths ... the community library, as opposed to the giant book seller in the cloud)
(edit 2: And I'd go as far as to say that it is anti-employment-generation to allocate economy-scale capital to a very small number of very large scale industrial installations... You wanna incentivize a large number of small-to-medium ones instead, that employ locally in many many more places. That is how you win back industrial capacity.)
this also misunderstands the data center rollout itself. openai and anthropic etc aren't building any data centers. they're promising to buy billions of compute and 3rd party companies are taking those promises and using them to fund the construction of data centers which will then lease out that compute back to them. so it's not actually a capital construction project at all. CoreWeave or whoever is making the capital investment. OpenAI is just promising to make that a good investment.
Yeah, Open AI and Anthropic may be buying the bulk of their compute from other providers but they are now moving into building their own infrastructure directly. There is a scale beyond which it makes more sense to build vs. rent
Zero cash flow is not the same as zero earnings. They may end up with negative earnings sometime in the future depending on how much additional capex they spend, how long they take to depreciate the capex, how fast revenue grows, etc
GPU seem to last lot longer than expected.. And have good resale value for other use cases. The $2 for every $1 inference cost quote from Anthropic I presume includes all opex.
I reckon it depends on how hot they run GPUs for how long. Running any chip at clock or overclocked (likely) will burn it out faster than "business as usual" workloads.
Defraying of capex also depends on how the hyperscalers scrap used hardware.
AFAIK, all hyperscaler hardware is effectively a consumable. Disks, memory, boards, chips; all the high-value pieces do not get refurbished into new computers for sale into secondary markets. The commercial and legal liability is too high---leaking their own proprietary IP as well as leaking customer IP, not to mention secrets and hardening configs and so forth.
I can't recall where I read it, but if a datacenter is modeled as a bunch of black-box hot-swappable shipping containers, then for pre-LLM workloads, per-container unit economics works only as long as a container delivers over 30% capacity. After that you got to swap it out, and send the black box into a furnace to recover them rare earth metals for the next generation of chips.
To a first approximation, high-throughput, gpu-maxxing math workloads (LLM inference) are surely harder on chips and memory, than are bread-and-butter SaaS / data lake / ETL / RDBMS workloads. Speculating further, I'd guess that the ~40% operating margin of datacenters is due to the relative compute-friendliness of non-math workloads for lack of better words.
> Running any chip at clock or overclocked (likely) will burn it out faster than "business as usual" workloads.
Could it be that they _have_ to run them like that because of the shortage of compute volume and once there are sufficient data centers, they can be run relatively cooler and thereby extending their lifespans? That doesn't bode well for NVIDIA/AMD.
> AFAIK, all hyperscaler hardware is effectively a consumable. Disks, memory, boards, chips; all the high-value pieces do not get refurbished into new computers for sale into secondary markets. The commercial and legal liability is too high---leaking their own proprietary IP as well as leaking customer IP, not to mention secrets and hardening configs and so forth.
Maybe that could be a catalyst for going for more "commodity" approach to hardware instead of highly spcialized (atleast for some sections of the hardware) so that the IP is not the hardware but the model/software and they can re-coup the costs by selling the hardware to "tier-2" providers that might emerge based on using cheaper second-hand hardware?
I suspect GPU and CPU failure rates have more to do with stress from thermal cycling, if they can keep them loaded and running without heating/cooling cycles, they should last pretty long. Their lifetime is probably dictated more from when do they stop being competitive on a flop/watt basis than by their failure rate.
Interesting... Upon reflection, I feel like it's a "yes, and" proposition coming down to failure mode.
[Type A] Constant use a constant power level: Quantum world rules would dominate.
When run hot, electron tunneling will kill off individual transistors, and AFAIK that is a cascading failure mode once it takes root. A big enough failed spot, even if localised, could effectively kill a core (if not the whole chip---too many errors). A local hotspot will directly affect neighbours, and indirectly the chip as a whole because the same work is being pushed on fewer cores.
[Type B] Thermal cycled: Mechanical and electrical factors would dominate...
Microfractures / mechanical damage in weak areas of whole chips, and motherboards. In a hyperscaler server rack, if a board dies that part of the rack is dead---micro-maintenance is probably too expensive.
And speculating further, it stands to reason that Type B failures would likely cause Type A failures, and vice-versa.
That only explains why the price of those GPUs at resale is high, but it doesn't invalidate the fact that those GPUs are still very useful way past a 2 to 3 years depreciation schedule advocated by Michael Burry et al.
Older GPUs can still be used to run model inference, albeit more slowly than newer GPUs. We already see frontier labs turning this insight into an opportunity to charge a premium on token delivery speed, when models run on newer hardware. As AI becomes more widely spread throughout the economy, not all workloads will require fast inference, and there will be many use cases for previous generation hardware. The only exception is if the latest batch of specialised chips become ubiquitous very quickly, but then price of compute AND intelligence per watt will both go down dramatically, which would reduce the need for sustained high capex and opex
Yes it is a supply-side crunch (chip manufacturing), but IMHO, it is not temporary. Only two things will remove the supply constraint.
- Demand drops like a stone: raw TFLOPS / per model. This will happen if someone figures out insanely fast and power-efficient math (specifically, multiplication) for existing chip architectures, that makes it dirt-cheap to keep re-computing ever-better SOTA LLMs, as well as to on-line compute inference with ever-bigger models (scaling laws rule). These are different workloads, AFAIK.
- Moore's law re-starts, because someone figures out entirely new transistor designs that can deliver at least up to 100x per-core frequency gains over Si transistors (which would be nothing considering that Si transistor tech delivered ~1Mx gains since invention) and which can be manufactured at similar bin sizes, at similar scale (i.e. equivalent per-die yield), price-competitively (no more 10x the equivalent in SOTA Silicon compute).
I'd say that a math breakthrough might be more accessible than the transistor-manufacturing breakthrough.
IDK if Moore's law restarting is possible. But I wouldn't rule out some sort of non-silicon compute breakthrough that functionally jumps us a couple orders of magnitude. The entire world economy being aggressively focused on compute seems like a good environment to drive investment into figuring out photonic computers or whatever. We do know at least one other non silicon highly energy efficient compute technology (biological brain) exists and works very well
> Moore's law re-starts, because someone figures out entirely new transistor designs that can deliver at least up to 100x per-core frequency gains over Si transistors (which would be nothing considering that Si transistor tech delivered ~1Mx gains since invention)
That assumes that there is no hard, physical limit on Moore's Law, which is obviously false. There is a limit, and as long as we depend on mass flow of charge through conductors, it's nowhere near quantum sizes.
Where that limit is has not yet been determined, but at this point the support structure to write transistors is the size of a house trailer, using a staggering amount of toxic chemicals (just to run the exotic lasers). So, advancements today may scale very nonlinearly in cost.
The very phrase "Moore's law restarts" acknowledges that we are no longer on-track.
Yep, I mean to imply practically impossible, in context of bubble-timelines. Theoretically anything is possible given enough decades. Practically, AGI has thus far been 20 years out, as has "clean" fusion power, 1 gee EM drives etc.
because of GPU depreciation schedules and competitive hardware upgrade cycles), and opex of delivering LLM inference services, and being able to price it all competitively.
Hacker News increasingly feels like being a horse in 1916, reading in the Horse Bulletin that Ford is a scam and a bubble because the mass-market automobile will soon be commoditized.
There's an object-level claim to consider, but behind it the more baffling implication that this would be good for horses.
No horsing around here, just money math that says interest rates on debt sunk into datacenters will far outstrip the most optimistic revenue projections as of 2028-29-ish, when the bulk of the debt will come a-calling.
Also, inverse Pascal's Wager, because horse v/s motorcar is a false dichotomy... they are bound by the same laws of physics (faster horse, less shit, more noise and more smoke), merely different economic realities.
Hyperscaler bet is "Superintelligence" and/or "AGI", or we will make a God. That is a technological singularity. There will be a "before God" and "after God" world, and we will very likely never be able to understand what the AGI/Superintelligence's world looks like. (Because to understand what it is like to be a bat, you have to become the bat, but at that point your meaning of "understanding" itself changes entirely).
So I believe ten trillion dollars and a bunch of pirated books and source repos, and well pirated everything, mostly of very dubious quality and provenance, will not result in a God. I've been acting accordingly.
I mean Ford didn't get a trillion dollars because he is to be only manufacturer of any and all transportation. There are still horses, and Ford and Dodge(Anthropic) are not even the biggest manufacturers.
Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
Everybody knew it.
But some investors don't have the patience for these "sure things" to mature into their commercial destiny and others know they can siphon off a killing by taking advantage of them as they speculate.
These bubbles aren't about potential, they're about timing.
And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
Most of the top players involved in AI right now are going to be rich. And in 20 years, the world is going to involve a lot of AI in soooo many things. But a lot of shirts are going to be lost below those top players and a lot of companies are going to fold because infrastructure, refinement, and integration take time while the FOMO speculation starts as soon as the big discovery is recognized.
> Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
Canals... where a repeating pattern was the company underestimating difficulty/costs, and going unrecoverably into debt. Another company, which then bought the largely finished canal, cheap, they had potential profit... until tech change (railroads).
> Another company, which then bought the largely finished canal, cheap, they had potential profit... until tech change (railroads).
Absolutely right. Somebody may buy the DCs. The tech changes very very fast, the DCs may be entirely worthless. US has high energy costs, it can never be cost competitive.
Useful (like solar, airlines, fiber), but it won't be profitable. High capital cost, low marginal cost, subject to competition and with technical improvement --> these four factors guarantee businesses will never be profitable, and they go bust.
> US has high energy costs, it can never be cost competitive.
A non-trivial percentage of that is artificial.
When push comes to shove and enormous profits (and anything remotely resembling national security) are at stake, the US is decent at removing artificial problems.
You've already seen a ton of that related to chips and data center build outs.
Why is solar more expensive in the US than China? Mainly tariffs. Same for batteries.
Why is nuclear more expensive in the US than South Korea? The government.
Why is SMR never going to beat conventional nuclear if you remove artificial problems? Physics. Though specialized applications will be viable with SMR.
The UK has even higher energy costs and people seem to want to build data centers here? (Or at least they say they do - looks more like land speculation to me):
The UK has got a huge amount of wind power that can be purchased ahead of time at better rates than listed in that table, it's also got some bottlenecks with power distribution (especially between Scotland and the rest of the UK) which will take years to fix so there area areas where decent power contracts can be had for customers who can build out fairly quickly (datacentre operators with large power requirements).
Both our high energy prices and the current capabilities of our energy distribution network relative to the renewable generation and storage capacity we are trying to move towards are significant barriers.
People want to do all kinds of idiotic things which make no sense. We've had excess compute/gpu since 2010. People didn't know what do with all that excess and started bitcoin and tens of thousands of crypto, followed by NFT. The claim was crypto will solve all problems, which hasn't happened. Now, everyone moved on to AI, AI will do this, that and everything else and also take away all your white collar and creative jobs.
Even third world countries like India which can't provide basic services to their citizens, have no water, no power, no sanitation or cleanliness are giving hundreds of billions of dollars in incentives (land, power, water, tax subsidies) to hyperscalers, to create at most 2.6 jobs. Google clearing land to build India's largest data centre: https://www.youtube.com/watch?v=dBJPf6Pxfvs
> The UK has even higher energy costs and people seem to want to build data centers here? (Or at least they say they do - looks more like land speculation to me):
There is no meaning to what politicians and crony capitalists do. It is all lies, theft and grift.
I agree with your point about rule of law, but let’s not gloat too quickly about nasty SA getting its comeuppance. It’s going to look pretty ugly in Europe when Russia follows through on its potential to disrupt or destroy critical infrastructure. We are pretty vulnerable too.
“When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds. I'm not going to pretend that I'm an expert on financial bubbles, but I've never heard of a bubble where the companies involved had such a modest valuation based on their earnings, even if their earnings (and therefore their evaluations) are enormous.”
Railroads had some of the same issue encountered by the dotcom bubble, and in particular dark fiber: Overcapacity.
Once railroads where demonstrated to be (extremely) profitable, people kept building them even if the existing demand was already met. The exact same thing happened with dark fiber (looking at you Enron).
AI is a bit different in the sense that the capacity can be used for research, and more importantly is not bound to a specific connection (between locations), and therefore a lot more fungible.
Not sure about canals, but while I think there are issue with current AI investments, I would not conflate them with the railroad bubble on this basis.
AI is a bit different in that unlike the rail tracks and the dark fiber the data centers arent actually going to be finished. It's much more of a stranded asset - there will be piles of chips and 1/4 finished construction projects, not usable assets.
No doubt somebody will use the assets for something but it's still wildly more wasteful than either railroads or dark fiber.
Also, (setting aside Standard Oil’s business practices), this is a common meme in oil exploration historically, and even today. Early prospectors were honestly pretty stupid, and their principal differentiation was risk tolerance.
There a lot of parallels with AI…it doesn’t take a genius to see the long-term potential, but the risk is in implementation, so it’s a mad rush to the finish line by a wide variety of less scrupulous early players. Rockefeller got his big start by sweeping up failing producers who were all capital weak, and I’m sure something similar will happen here during the inevitable bust.
I can do the hard, risky work of drilling wells / deploying AI, or I can sell you a vision of that future, today. The most dangerous technologies in terms of grifting are the ones that everyone can understand.
You should check out the fundamentals of all those businesses; quite different. AI might be closest to a railroad based on these Capex outlays, and like railroads and other utility businesses the margins are thin and paybacks long. We wouldn't have Microsofts and Googles at 3% a year, so OpenAI and Anthropic are going to need something else to support this level of investment and speculation.
A large number of companies failed in the dotcom crash, but the founders did pretty well if they'd made it to an exit event. Don't forget that it was a stock market bubble imploding. The founders and early employees had already sold shares in the IPOs that put the companies on exchanges. It was the investors who bought stock that lost money.
The after effects were that it was virtually impossible to raise capital for a while afterwards, so the next round of founders struggled. But then interest rates went to practically zero and money got cheap so it was all good.
The founders doing well is not mutually exclusive to the overall crash as experienced by the public. Also, a lot of founders are doing well even before they take on a venture. 8/10 founders from successful high growth startups self report as middle to upper-middle class background and 50%+ have elite education from top 10 schools.
You're right, the broad market drags down all retirement savings and those who are about to hit the age of 60+ pay most dearly.
Wait. Are you talking companies, or people? Companies come and go. They are replaceable virtual containers for people. I don’t think it’s appropriate to apply the term “poverty” to companies.
The Yahoo founders left with never-fly-commercial money before re-joining the billionaire club by investing the money they had. Others that survived post-IPO include eBay and Dell.
There were high profile failures but we have had those recently too eg. pet.com (dot com) and color.com (post-dot com, domain has since been repurposed)
> The Yahoo founders left with never-fly-commercial money before re-joining the billionaire club by investing the money they had.
I think you're confusing people cashing out with actually creating a successful company. Yahoo is not what people would describe as flagship company, or pets.com or Ask Jeeves or myspace or Voat or whatever. The dotcom bubble made a lot of people rich, but most examples were at expense of bag holders.
If you go by current standings in the market, yes, Yahoo is no longer at the top. Yet Yahoo was one of the most valuable and influential tech companies in the late 90s, early 2000s.
Yup, if you google what founders in those 'failures' did was that they pretty much jumped vc side or founded more companies and have had very successful career. Not mind boggling like Bezos or Brin but still very cool career.
For a VC "running" would mean re-invest in the next thing to get more money. What alternative would you prefer? Keep doing the same thing without investing in new stuff?
They don't "take the money", there are other people that believe some things (profits/growth in a field) will last forever and will bring 10x return, so they are willing to give them the money.
> AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
That's an interesting claim but it needs support. How are you comparing their maturity rate? AI is arguably maturing very fast.
Another argument in favor of the AI build out, usage of the data centers is not geographically bound like the other technologies considered. A datacenter in Cheyenne WY can be used by the entire continent. They are not even bound to AI usage.
>But a lot of shirts are going to be lost below those top players and a lot of companies are going to fold because infrastructure, refinement, and integration take time while the FOMO speculation starts as soon as the big discovery is recognized.
Another reason is that the way we find out what works and what not is through experimentation. Especially in a transformation like this it’s impossible to predict which configurations will work and which won’t.
Yes. And valuation. I think the leading frontier labs will be fine in the long-run and NVidia will be fine. Some of their investors who bought at too high of a valuation, expecting to flip it in too short a time frame will lose a chunk of their money when they have to sell for less.
The hyper-scalers in the middle are the most vulnerable. They borrowed billions at increasingly high rates to lock up GPU and power capacity. They are 'ground zero' of a lot of the 'circular money' deals. Big banks who financed hyper-scalers initially, expecting they could package that debt up and sell it on to diversify, are already struggling to find buyers for the debt, causing its value to drop, which then forces the banks to cover with more collateral, and that's how the snowball starts down the mountain.
NVIDIA will take a huge hit and we will probably end up with anti circular financing laws.
Frontier labs are losing money and are most likely to go bankrupt and be absorded by bigger companies.
Hyperscalers will take a huge beating but they are ultimately profitable and they already control the frontier labs. So after a hard crash they will be fine. A lot of their debt is actually offloaded onto pension funds, insurance companies, etc. Those risk bankruptcy.
They will be just fine. They might not see the astronomical price margins of the current bubble, but there is a lot of untapped demand that is curbed by current mem/GPU prices and if they go down the demand will rapidy grow.
Wait... with hyper-scalers, we are talking about Google, Meta, Amazon, Microsoft, are we not?
And do these not all have established business models that are ridiculously profitable, and huge cash reserves?
How the heck are THEY vulnerable, and the "bare" frontier labs like Anthropic and OpenAI with no preexisting business models or profits, but equally ridiculous borrowing are supposedly not?
I don't think they are talking about the ones you listed, but companies like CoreWeave.
Google, Meta, Amazon, MS have huge cash-producing businesses they can just fall back on. Of course the stocks will get hammered in the short term, but they are still sitting on cash volcanos.
Pure AI plays are what are most at risk. In the bad case, all the CoreWeave type companies have their assets acquired on the cheap by the existing big players. In the armageddon case, companies like Anthropic and OpenAI also end up in the discount bin.
I believe AI is here to stay as general concept turned into practically working implementation.
What that practicality means is we all who consume AI are learning but we're still in the early adopters phase of all that.
Of course AI companies invest a lot and try to stay on top of the traction but I also believe not all of them will survive test of time.
I personally believe that a lot of over investment in data centers and supply of energy will be burnt and will have to be reinvested somewhere else when technology advances enough to "miniaturize" AI enough so it won't require couple of football fields to provide what it does today.
I do believe cloud based AI will be always attractive to small business if price is right and to big enterprises so they don't need to invest inhouse for the scale they need, but that will be sort of alliance based enterprise B2B market.
anything between (including me for the scale I represent business wise), I can't wait until I will be able to buy appliance from 12 to 48U maximum that can do all what me and my company needs AI wise as local AI.
moment subscriptions become equally expensive as pay as you go models, I switch to mac studios with 96GB and reasonable weighted models that will be specialized in something instead of one LLM/AI to prompt everything to it and hoping it can address all sort of problems as one solution.
AI one day must fly to space, be called Mother and discover aliens. it can't be 6 football fields as physics won't let it take off. It must be one day small enough so we all can have ChatGPT 6.1 medium model at home without subscription, just as fixed price with TCO over lifecycle of the whole appliance (software + hardware together)
that may become a drama to many companies that apparently may not be worth trillions unless they diversify enough its AI offering.
> Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
I don't see all these things as being comparable to AI... At the bottom, they were "100%" technologies. Trains were just able to transport commodities, and the internet was able to reliably connect computers. So their path to success was about gaining market.
Since they were 100% technologies, they disappeared. One could build solutions using them and forget they are building blocks, because they just worked.
LLM AI is different at the fundamental level, because it's non-deterministic. It needs human-in-the-loop. It won't ever disappear. Gaining market share doesn't fix anything.
Their point is (and I tend to agree), that LLM AI is not different at the fundamental level. The idea is that e.g. coding capability will be a standard feature of future computing hardware, not a very expensive service you rent. Anthropic is the current-gen mainframe operator selling time-shared access. In the future, that capability, too, will just be a commodity, i.e. a part of the silicon in your notebook or smartphone or whatever will replace those.
Yes, to a large extent, for current real world mass market use cases at least, the only thing keeping the biggest players relevant is that they (and their pals) have monopolised access to GPUs by buying them all and making deals to buy almost everything that Nvidia can produce (even more than they can plug in).
If SMEs were able to buy and host their own GPUs, they would not hesitate to avoid paying a subscription for access to AI tools. They're already starting to do this.
Models already are a commodity, the massive bump in access prices to profitable levels is not going to happen, because businesses will not pay it.
Consider the wisdom of speculating on the incredibly fast adoption of the internet and mobile technologies, and on that basis investing heavily in whatever mobile carrier had the fastest network.
The speculation that the AI investment sector depends on is not just that adoption is incredible and that the value it generates is high, it's ALSO that the companies will be able to capture a significant amount of that value generation for themselves via fees or some other mechanism.
Are the LLM vendors more comparable to Google/Apple in this scenario, or more comparable to AT&T and Bell?
Tne value gap is closing, as corporates wind back token budgets towards relative sanity and reassess what they're getting for their money.
The bubble exists because c-suite people are sheep. They follow the herd, and the frontier CEO wolf-wannabes have been trying to herd them in a profitable direction.
But herds are easily frightened, and there's always a danger they'll stampede in a different direction.
And the risk of a catastrophic AI-enhanced online blackout is increasing.
> The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.
They’re still losing gigantic amounts of money. Whether that is because they’re “flooring it” as it were and growth will eventually catch up, or because the business idea is fundamentally flawed and just temporarily alive because of investor risk appetite, is the fundamental question underpinning whether it’s a bubble or not, and nobody knows the answer to that question for sure. That’s how bubbles work.
The substack linked below only looks into one tiny sliver of the off-sheet circular financing that dwarfs even the most optimistic revenue growth, but it's a good place to start to understand why this is a bubble and that, yes, Anthropic is in fact "losing gigantic amounts of money":
"Boom and bust is not about revenue. To find that kind of risk, start by reading the balance sheet... [The 2008] bust came from off-balance-sheet instruments."
What are you talking about? OpenAI is losing money, projected a $14B loss for 2026 and is themselves not expecting profit until 2029-30. xAI is losing money, about 3B.
Where do you get the idea that Anthropic is profitable from? Is this another one of those “if we remove everything that is a cost, then revenue means profit”-scenarios?
> The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.
How is that compared to the speed at which money is thrown in them? And how much of the profit downstream they might indeed potentially capture with high retention rate?
Bubble can’t be assessed on mere quarter profit expansion over short period (relative to large company group life expectancy).
It is looking entirely plausible to me that Anthropic will make a profit in 2027.
Don't you think that it's less than reasonable to suggest that the technology will take on the order of 20 years to mature into widespread adoption and profitability?
> It is looking entirely plausible to me that Anthropic will make a profit in 2027.
Then you should invest heavily in their IPO, I suppose.
You'll do much better than those of us who believe there is no way their financials can ever work.
I came to my conclusion a different way, I ran different models locally and measured the amout of resources required for basic things. From that I extrapolated some napkin unit economics, and I couldn't find any way to build a profitable business from it.
I don't assume I'm the smartest guy in the room though, so for all of our sake, I hope you're right.
> And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
They may be as expensive to mature (in terms of equivalent duration of total economic output), but by the wall-clock we've gone from "invent Transformers to make translation less bad" to "solves Millennium maths problems" in 9 years.
After 9 years of railroads… well, those were originally an Ancient Greek way to drag ships over an otherwise annoying bit of land, and the more modern attempt had wooden rails and a rope for humans and animals to pull for around 250 years so let's say after 9 years of the steam locomotive: 1784 for the prototype, 1802 for the "first full-scale working railway steam locomotive" (according to Wikipedia), so 9 years past the invention and it was still a prototype.
Semiconductors (well, transistors, i.e. us understanding the process and doing it deliberately, rather than getting lucky with cat's whiskers before we had any clue about quantum mechanics) did at least get commercialised in less than 9 years, but AFAICT in the form of "now our radios are better, sturdier, and smaller" which is a rather smaller change than what's already happened since Transformers.
(Of course, why Transformers, rather than the other attempts at AI? Transformers are when people cared this time around, but there have been previous AI bubbles. I'd argue those are like pointing at 1515 (or the Ancient Greeks) for the railway, or cat's whiskers for semiconductors).
I've been saying for awhile now that the utility is increasingly undeniable. It's amazing technology and for better or worse, is transforming the tech industry (in particular). I just wish we lived in a world without the speculative bubbles, because if this had been developed slowly, deliberately, over time, there would be no necessity to cast aside troublesome details like safety, suitability, and ethics...and we could enjoy the fruits as it was implemented gradually, responsibly, as it matures. As opposed to going "wow, this is really cool, but gosh I feel shitty about using it..."
Nor would it have been thrust blindly into the hands of your manager, the marketing department, the CEO, and countless other folks who have no understanding of how it works, its limitations, or its risks.
> I just wish we lived in a world without the speculative bubbles
Speculative bubbles, I.E. risk is where much of the capital comes from to develop these ideas to start. I’m not sure you can have one without the other.
When you say “increasingly undeniable” do you mean that it’s already plainly undeniable and it’s getting more so? Or do you mean it is still a bit deniable but is threatening to become undeniable soon?
A tech workers union or a stronger regulatory state could help with this but neither party has been that interested in either of those things for the last 50 years.
Well at least from where I'm sitting, it's working fine for us. We aren't in the midst of an oligarchy made up of the Epstein class, for one.
Could thing be better with the tech industry here? Sure, but at the same time, profits and industry leaders aren't the only things that matters. And at the very least companies here can't just exploit the public with total impunity.
I am also a European. Our oligarchy looks different for sure (and look into the history of pedophile scandals in Belgium, BTW).
I am not as optimistic as you. We seem to be stagnating our way into a stark choice between becoming an American tech vassal, a Chinese tech vassal, or losing out the remnants of our ability to compete on the world markets. Half of the continent borrows money at an unsustainable rate in order to pay millions of people to do nothing for decades.
We have made decisions that looked reasonable at the time they were made, but too much has changed in the meantime and now we are stuck in bad equilibria.
The Bismarckian pay-as-you-go pension system was stable as long as people had plenty of kids and died, on average, a few years after retiring. Now they have one kid and spend 20 years in retirement before they die. The maths of that system just does not work out anymore.
The European highly regulated industry was king as long as the rest of the world had no other choice, but now that Asia has industrialized, they are blowing us out of the water. First with lower prices, but now also with quality. German exports are half of what they used to be, with another halving expected. And of course without exports, our current social spending cannot be sustained either - you cannot keep such a massive redistribution mechanism running on an economy based on selling each other cups of coffee in a beach bar.
Our most talented engineers tend to move abroad and stay there. Brain drain is one of the most serious things that can happen to a country or a continent, and we are suffering from it.
We just cannot continue to live like that, something will have to yield, but many voters just don't want to hear that.
We are. In the UK we have the former Prince Andrew and Peter Mandelson selling information to Epstein. There are lots of other people all around the world connected to Epstein and mostly investigations are slow or on-existent or blocked through lack of cooperation by the US.
Epstein's network is one we know about because its been exposed and because of the sexual abuse aspect of it has both drawn attention and made people less likely to cover up for those involved. Its only a fraction of the corruption that goes on though.
> We aren't in the midst of an oligarchy made up of the Epstein class, for one.
Europe even has literally hereditary monarchies; let alone an insular upper class who are famous for their debauchery. They just have much less free speech protection so it doesn't get bought up as much because of the legal risks.
Related note, but while looking up some background info on Wikipedia it appears [0] it was technically illegal to publicly insult Donald Trump in Germany at the start of his presidency which would have added a hilarious angle to a classic joke [1].
[1] In the USA, you can stand in front of the White House in Washington, DC, and yell, "Down with Ronald Reagan", and you will not be punished. Equally, you can also stand in Red Square in Moscow and yell, "Down with Ronald Reagan", and you will not be punished.
The US is great for free speech as its constitutionally protected. However it is also very corrupt. Pros and cons.
There are a lot of advantages to constitutional monarchies in terms of stability. The monarchs do not actually have much power. I on balance prefer living in the UK and would not want to move to the US. There are pros and cons, and its culture and society (e.g. unarmed police) as well as politics and some if it is about personal circumstances (as someone who is obviously ethnic minority or foreign everywhere in the world I prefer a less racist society, for example).
I can yell "down with the King" or "down with the Prime Minister" pretty much anywhere. There are some restrictions on protest in a few places like Parliament square.
Indeed I am, and you're right, there is a reason for it. The tech unions don't need to capture mindshare in order to advertise VC-funded startups ;)
We shouldn't assume that YC is hosting this forum out of benevolence. However it is true that this place is also useful for things other than just advertisement, so that's a neat side effect. We just shouldn't assume that that's the primary purpose of HN.
> I've been saying for awhile now that the utility is increasingly undeniable.
If that was true it is strange that you would find yourself saying as much for a while now, given it sounds most appropriate in response to someone denying the utility of slop generators.
I wonder if physical technological progress is also the same as information treatment progress. Everybody is happy having food and materials faster. But AI bringing you solutions faster is also taking away your job, passion.. the negative side seems larger.
>I wonder if physical technological progress is also the same as information treatment progress. Everybody is happy having food and materials faster. But AI bringing you solutions faster is also taking away your job, passion..
Bubbles aren't caused by betting on the technology. They're caused by betting on the wrong companies. Say you have railway, and there's 5000 railway companies. At this point no one has any idea what good looks like, and it changes every 6 months. Investors bet on all 5000, and once all the drama is over, there are 3 companies still capable of competing in the railway business. The investors in the other 4997 get screwed, but since the investors also bet on the winning companies, theres a good chance the actual successful companies themselves get screwed because their investors are broke. There is then intense consolidation, and one of the 3 to nail the technology is still left standing while somehow there are now 3-4 other companies that won by taking talent and inventions from bankrupt companies or just by peaking after the bubble burst. About 20% of the bubble companies have also survived by finding niches.
That's a bubble. It happens every time because it has to happen every time. If you stay out of the game, the people who stay in the game win and your holdings in redundant technology are worthless.
Bubbles are caused by reckless stock purchases leading to insane, unjustifiable valuations, regardless of the company's prospects.
Hysteria + widespread financially illiteracy means there are lots of people prepared to buy AI stocks at any price. They're adamant to get in (FOMO).
Suppose OpenAI will be a successful company into the long term, and you bought its stock at a $5T valuation. After several reporting seasons and subsiding hype, its true worth turns out to be closer $1T. OpenAI could still be enormously successful, but anyone you bought at much over $1T valuation is going to get hurt.
That's not the difference. The difference is the distinction between core infra, B2B services, and B2C services.
The companies that sold core infra during dot com mostly died, because they invested in fiber etc far too quickly. Likewise with server factories like Sun.
The B2B and B2C survivors were patchy. A few did extremely well but there were a lot of casualties.
The biggest winners offered B2G - business to government - disguised as search and social media. They were actually selling individual user profiling and targeted ads, and that turned out to be the biggest money maker.
The reason Anthropic and OpenAI appear wildly overvalued is because their offering isn't AI in some abstract sense, it's the next generation of surveillance and persuasion technology. But now with corporate surveillance in addition to the traditional profiling and targeting.
You can tell a lot about what someone is doing - and the industry is doing in general - by looking at everyone's code bases. You don't need to train on them or steal keys to extract immense value.
Potentially there's also the option to eventually make claims on corporate IP developed with AI tech.
Anyone who thinks the companies who pulped and torrented whole libraries of books for their training data would never, ever, pinky promise, try to extract value from the code they're generating is being very naive.
>AI models can be deployed on the same datacenter as the old one
This hasn't historically been true. The clusters have been getting bigger along with the models over time. And now we may even want to have these computers have rapid access to extra compute or caches of the internet or other data stores.
> AI models can be deployed on the same datacenter as the old one
Not if the datacenters are in space, you can upgrade the models for a little while but once you need to upgrade the silicon you need to launch new hardware.
Google hasn't retired any of its now ten-year-old V1 TPUs or any of the subsequent generations so it's unclear that AI infrastructure "needs" upgrades.
Of course, Google could be wrong about the economics of running old AI processors. Or maybe TPUs are special.
I do wonder how much of their inference is running on 10-year old TPUs. I’d imagine not much, if only because 10 years ago the justification for massive compute was quite different than today.
What on earth makes people assume there's still huge growth possible in the information sector? We've long been at the very end of the information age and even the latest attempts to squeeze out more benefits using "big data" have failed spectacularly. Especially when it comes to actually producing physical wealth i have seen no indications at all that we can increase anything drastically with a more efficient permutation of information.
For example, Let's assume AI "solves" the traveling salesman and let's assume it indeed does find a better route for the delivery guy, what kind time savings are we looking at? The 8 hour route is done in 7.5? That's not going change any society in any measurable way. The just in time industry might become even juster in time but that's about it.
It might be dated but not too long ago a report showed the effect AI had on the SAP 500 and 490 companies couldn't even find any substantial improvement. Mind you, this is during the craze, this is happening while every manager has all the incentives to use as much of the technology as possible, lest they be the one that missed out on the new steam engine. All of these companies invested heavily in cars because none of them wanted to be stuck using horses while everyone else drives off into the sunset... yet most of them didn't manage to measurably outrun the people on horses. That's not the prediction I would have made if I were switching from horses to cars.
This reads less like a bulwark against sensationalism and more like an attempt to remain ideologically safe in what appears from many angles to be an extremely steep exponential start of what we can probably agree is an S-curve.
> And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
I suspect this opinion may come across as naive for the myriad software engineers whose job has become pressing the enter key. Your point about timing is based on solid principles, but I don't think your comment contains a satisfactory argument that the timing is incorrect.
How do you know we aren't already through the exponential part of the S-curve and about to enter the AGI winter (i.e. AGI will require a fundamentally different technology than LLMs that is still decades off)?
If no one is willing to pay increasingly high costs of training frontier models, we'll still have the models that we have now that can be fine tuned for incremental improvements. Still tons of utility in that, but little investor upside.
> How do you know we aren't already through the exponential part of the S-curve
From a recent METR report, benchmarks point towards an acceleration, LLMs are getting better at a faster rate. There's no sign that we are about to reach the top of the S curve.
> increasingly high costs of training frontier models
Somewhat surprisingly the gains do not seems to come from more compute, we have on the METR report:
> We are confident that the acceleration was not due to an acceleration in the scaling of training compute.
> Somewhat surprisingly the gains do not seems to come from more compute
I think this could form part of an argument that acceleration is slowing. It's only because we are not making as much improvement out of hardware gains that we're having to find the improvements in more complex architecture and more tricky to create training data.
The fact that the gains are not coming from the traditional place gains come from in technology might be a sign that a high percentage of the low hanging fruit has now been picked.
>From a recent METR report, benchmarks point towards an acceleration
On an internal benchmark with internal models, if you look at the public benchmark with public models there is no acceleration. That report also has a big graph labeled "fake data".
This isn't true. Metr reports show consistent increases in acceleration, 1.4x in the November report. Even with the brake of deliberate hobbling by their makers this year.
In 1893, GE nearly went bust as the market got cold feet that electrification was too costly, and the uses too abstract and niche. People were also hugely worried about circular financing, as they were guaranteeing bonds and taking equity stakes in their clients.
As we all know, GE then went bankrupt and we went back to whale oil.
This argument sneaks in the fact that models are commodities which is NOT true. Claude has a different personality and performance than other models for example.
Expect more differentiation in the future and not less.
> Claude has a different personality and performance than other models for example
This kind of thing amounts to a LoRa applied to your model. There's no reason (apart from potential IP infringement) that someone else couldn't produce a "Claude LoRa" for their model (or, more likely, given the dislike for Claude's writing style, Anthropic could offer a "non-Claude LoRa" for their own models)
I completely disagree with this. What you are saying is that performance, all they need a simple trick and they get Claude performance for free. In that case, why do we see open weight models still lagging?
Assuming you meant this for personality, I’d still disagree but not that extensively. I think personality is nuanced and complex - it can’t be simply lifted and shifted. It’s grown using a meticulous process.
> I completely disagree with this. What you are saying is that performance, all they need a simple trick and they get Claude performance for free. In that case, why do we see open weight models still lagging?
Yes, they are lagging.
But "performance of Claude's from 6 months behind" is fine for a lot of tasks. And the performance of today's Claude will be still useful for many (and more and more) tasks in 6 months.
The frontier labs are making models for smaller and smaller share of the pie, as the capabilities of open models grow. That isn't sustainable
I disagree, I always reach for the best model when I’m at work. For normal questions I do use the default ChatGPT model.
Fundamentally, people underestimate the marginal increase in utility for marginal increase in “intelligence” as measured in benchmarks. Like I’d pay 5x for a model that has 2 units higher in AA intelligence.
I don’t think GPT Astra is as good as Opus 5.5, it’s very noticeable how pedantic and poor at intuiting things it is to me. It often calls things done when Claude would keep working on them, and doesn’t seem to have as good of an eye - often leaving clear visual defects and calling things done.
> Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
I wonder why so few people make the comparison to the recent VR hype. We heard the same things there about its unlimited potential, how it's inevitable that we will soon resolve all its shortcomings and that it will become ubiquitous.
I still believe VR will eventually live up to its potential, but no one can deny that bubble has popped after Meta took a huge loss on it. It took too long to resolve the hard problems so the timing didn't work for them.
AI can maintain the multiverse and administrate it. Everyone can create just by talking.
Coding would also be a good usecase, to see the architecure in 3D and talk with Claude, point at stuff, pick him up and put him where the work needs to be done.
I think it's a match made in heaven, AI is just still way too slow an expensive for this.
> With AI VR does have usecases. AI can maintain the multiverse and administrate it. Everyone can create just by talking.
Nobody cared about multiverses before AI, and it wasn't because of the lack of maintenance and administration. It's just a thing that nobody cares about, AI improve it but it's not going to make people care about it so that doesn't really change anything.
Okay, but why would they? And who's going to consume?
If anyone can trivially create, every creation is meaningless. There is no billion-dollar business in abandoned Minecraft servers, so why would this suddenly change when you force people to put on an expensive VR headset to visit them?
> If anyone can trivially create, every creation is meaningless.
Everyone has always been able to trivially create. We live in an attention economy, not a creativity economy. Meaning you can create crap (or ai slop) and people will consume it if you know how to get their attention.
Attention is what gets money. Creation is just fuel for attention. Quality is a part of the equation, but much less important than people generally believe.
Not measurable in ROI according to reports that keep coming out every week from companies capping spending because the wow effect is not translating into growth.
Obviously AI people say it's because they don't know what they are doing, but are you going to tell me a company like Apple couldn't figure out how translate those productivity gains into something more tangible?
It didndump a lot of money into it, improving some things and headset availability. But arguably a most of it went to useless stuff no one asked for, like paying developers to build shovelware for Quest or their useless corporate Metaverse.
> I wonder why so few people make the comparison to the recent VR hype.
AI is already a gigantically transformative technology, with a staggering amount of paid consumption, and virtually everyone I know uses it every day, where I don't know anybody who regularly uses VR in any form.
The ship has absolutely sailed on "is AI a world-changing technology", where "are the leading AI companies worth what's been paid for them" is a completely open question.
AI is transformative _doesn't_ mean there isn't a bubble. Online shopping was transformative too, but there was still a dot com bubble. Amazon survived, Pets.com did not.
Because VR failed to deliver at all. The experience is narrow, isolating, and frankly pretty horrible. People get motion sickness. The set of games is limited never mind the set of non-gaming activities. AR similarly has a lot of issues like narrow FOVs, low resolution, and brightness issues.
VR is a toy with a few very narrow areas of application in industry. That does not describe the current status of AI.
Yes. That's the situation in 2026 in most of the world. But in the 19th (after the railway bubble - https://en.wikipedia.org/wiki/Railway_Mania) and early 20th century, railways were actually very profitable.
> Almost everywhere I the world railways are state owned and heavily subsidised and still aren’t that attractive to customers.
That's a remnant of the past, where only a nation state could afford massive projects such as creating a railway line crisscrossing throughout a massive country.
Since the last couple of decades we see private companies investing in railway lines, including high-speed railway corridors. We also see a major liberalization push in Europe to decouple railway operators from infrastructure maintainers, and now there are plenty of private companies operating railway services across the world. In Spain alone there are 2 or 4 private companies providing high-speed railway connections between major urban centers.
The reasons for this are complicated and nuanced though.
Governments think they push the financial burden of transport maintenance on the people (and its a good tool to register and control people too, but thats not a strong consideration I wouldn’t wager), so they fervently cut spending on public transport and curtail cycling infrastructure in favour of cars.
But roads themselves operate at a crazy financial loss and are one of the reasons governments are in crazy debt (which nobody wants to acknowledge).
This is because road maintenance is significantly more expensive than building new roads. Which means that road use becomes heavily subsidised (much more than trains cost per user) but the cost of vehicle ownership, despite that heavy subsidy, is still enormous and one of the largest expenses an average person can have.
Cars a tool of state control and not public transport? If the government wants to surveil someone it’s much easier to see exactly where they tapped their card on and off, and if they want to stop people getting somewhere all they need to do is tell the drivers to shut it down rather than set up a bunch of roadblocks.
"The fourth power law (also known as the fourth power rule) is a rule of thumb which states that the damage to the road caused by a motor vehicle increases in proportion to the fourth power of its axle load. This law was discovered in the course of a series of scientific experiments in the United States in the late 1950s and was decisive for the development of standard construction methods in road construction."
"Example: Car (total weight 2 tonnes, 2 axles): load per axle: 1 tonnes vs Truck (total weight 30 tonnes, 3 axles): load per axle: 10 tonnes. [...] The resulting damage differential between truck and car is 15,000 to 1"
PS: this is something electric car proponents don't talk much about
The rule you linked to basically shows that whether a car weights 1.5 tonnes or 2 tonnes (an EV version of a car model usually weight about 25-30% more) is completely irrelevant to road damage when you have heavy freight using the road as well. The single 30 tonne truck from the example above gives the equivalent road wear to 15 thousand 2-tonne EV's.
- In Europe the petrol taxes pay for road maintenance. Only electric vehicles operate at a loss.
- Train capacity is much, much higher. You couldn't have the density of France, the Netherlands or even NYC without most people travelling by train. LA struggles with 1/10th of the density.
Europe is a large disparate continent, but I don’t know any country (neither Sweden, Denmark, Norway, Estonia, Finland or the UK) that covers road maintenance costs with fuel levies.
In fact in the UK you would be forced to aggregate all fuel and road tax and offset it against road maintenance because there is no direct correlation to be made, as tax is “one big bucket”.
The majority of the planet are going from one point of interest to another at a relative constant rate- but yes people are spread further and further- ironically this is helped by requiring huge parking space everywhere anyone needs to go, and needing huge houses so you can park multiple cars outside. Its a self-reinforcing problem.
(This is most evident when comparing cities like Malmö in Sweden with Orlando, FL. Malmö has a higher total population but approximately 1/20th the road surface- yet in Orlando I kept bumping into the same people everywhere I went because we’d all drive to the same Restaurants and Shops; and despite that happening we had to drive 20-25 minutes to get there! Weird coincidence or natural concentration points? Someone smarter than me would have to say).
The "road tax" misconception is why I feel the need to point it out as it is inaccurate and is used to weaponise hate against cyclists (e.g. "Don't bloody pay road tax"). In reality, most cyclists also drive (not me, though), but everyone who pays income tax is funding the roads which I agree with as they are shared infrastructure. However, it does mean that drivers are being subsidised by non-drivers.
Of course, with the 4th power weight relationship with road damage, ordinary cars are not really causing much damage compared to buses and trucks, so maybe it's better to state that we're subsidising logistics companies.
I can cycle anywhere in my city in 15 minutes- and anyone else can too.
Yes, it gets cold in the winter in Sweden- and I sometimes have to wear rainware. But the lack of noise, huge highways and being able to walk if necessary is such a crazy feeling of freedom that you can't possibly imagine.
I literally mean, you cannot imagine.
Being able to drink a beer, and then walk with a friend in the evening with my bike in hand, being able to park anywhere at any time, and being able to replace/fix my bike with no money (or, even buy a new bike without crazy paperwork or registration or licensing for my ability to drive it).
Not having anxiety about paying for parking, fuel or ensuring that I have a place to park, not to mention the lack of abusive road-rage folks.
I've lived in other places, and I've visited supposed "car-utopias" where everything is built to "great" standards and honestly the quality of life is just many orders of magnitude inferior. I've never been to the US and stayed in a nice quiet house without it being 30 minutes drive (in stop/start traffic and with aggressive drivers who cut you off) from anything of any value.
So, no, it's not the best option. It's the best option you can think of.
In germoney, IF i remember correctly, it is supposed to be like that. Seeing the state of the overall..well..corruption of autobahn bridges, roads and the "private" company governing it seems like its either not enough or the bucket has holes OR the money goes elsewhere. Pensions maybe, maybe gas power stations, who knows?
Everything is heavily subsidized by socializing the losses and deflation of buying power for the average person.
Corporations demanding payment for everything at whatever margin they want is little different than taxation without representation and local cops shaking people down for money at random; just different legal semantics
The representation is you yourself by voting with where your money goes. At least in a market with competition.
Without competition you are talking about a monopoly and it is your government that is responsible for mitigating its effects. The government that you elected.
The mechanisms are all there, they just need to be used+embraced.
Yeah I know how the American civil religion gospel goes; don't patronize
There is inherently competition with entropy and just getting ourselves out of bed to deal with everything
Corporations and brands and tax codes are just culture war distractions on top of the raw physical competition; middle ages like words of power
The competition is with the physics that binds us to Earth; does it? Or are we just to distracted by rhetorical competition between legal constructs to embrace the levers of science
Rhetorical competition wrt to markets type blather is only seeking internal consistency with an arbitrary rhetorical framework. Not pushing the boundaries of engineering based on science.
Yall can have AI agents simulate Bloomberg terminals so actual physical scientists can focus on the necessities ground truth work you take for granted
Specifically because of massive government investment in road infrastructure in preference of railways. This would be about like the government deciding to funds DCs directly...
Your point is a good one, but it's constructed upon the premise that AI is a single discovery or invention or even a single tech wave.
AI is bigger than the Industrial Revolution, the Internet, than computers, than literally any innovation or invention that has ever come before, combined.
Today you see AI slop and vibe coding.
Tomorrow, all of humanity will be affected, forever.
Everything about the way that we learn, that we think, and that we invent will be changed forever.
Strange comment to make. We shouldn't be honest about AI's capabilities?
How do you think we need honest deliberate action if 95% of people believe AI is a scam and a bubble and mostly just hype marketing? Just take HN's comments for example.
You're defending one extreme by claiming the other extreme is the only option. Which isn't honest behaviour even if it was well intended. Truthfully, AI will cause real problems and we need to handle it, but it's also often exaggerated. It's a rapid technological shift, but like all the other ones anyone who works hard will still find a place to be useful.
"Tomorrow, all of humanity will be affected, forever."
This..is not being honest about AI's capabilities. Its just generic hyperbole. Also, simply not specific enough to be called honest or dishonest. It's just panic-inducing.
We're discussing AI in general here, not individual companies (and none of those companies have a monopoly). Nobody's arguing that AI wouldn't exist if the previous revolutions hadn't occurred.
Each event was so different with different motivations. The popularity of Tulips never back. 2007 was about hiding poor risk mortgages. 2001 every company had money thrown at them and today is about a product that costs more than the market can pay.
You know what was actually a bigger invention? Language. Farming.
Imagine a world without those. Oh, wait: tomorrow people won’t know how to use language, will they? And they won’t have water to farm with, so there you go.
I have not been able to workout if AI means demand for compute for
most people will mean it's infinite since you can have AI to just work
on many different problems and many different possible solutions in
parallel. Or if for most people (developers included) once AI is good
enough for most tasks people will reach a limit of things they could
want to do and find some limit of compute enough and if so where would
this limit be.
I am talking about average users and developers, there will always be
tasks that want as much compute as possible I suppose but these will
be limited and cash constrained as they where in the past, I can
imagine most developers being satisfied with finite compute and not
wanting any more for their AI agents and not needing any more
intelligence for their LLM's so no more training apart from distilling
and making models more efficent.
The thing I don't know is if where AI is basically good enough for
largely all possible things people could want to do, if the lack of
demand for training coupled with the satiated desire for compute will
mean hardware falls massively in price as well as local AI getting
better and cheaper to run. It feels as if some of these data centers
are not going to find have the issue of massive over supply.
The logic here is so flawed it suggests to me a significant lack of financial understanding. For example:
>If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year. For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019. Adjusted for inflation, that's an even more impressive 213 billion dollars.
Alphabet's REVENUE was $162B, not earnings! You can't use all your revenue to pay off your debts because you have to use some of your revenue to pay operating and other expenses. So even if AI generates as much revenue as Alphabet did in 2019 (adjusted for inflation), there is no chance it could pay back $250B/year.
Your point is valid, but there is a flaw in the premise. They don't need to pay off the $5 trillion in 20 years. They just need to be able to service it. If they can use that debt to bootstrap the company, as long as they can pay the interest and issue new debt, they can keep that $5 trillion on the books longer than either of us will be alive
I suspect if you go back to look at say, Microsoft's books from when they went public to today they have never been debt free
The author talks about revenue growth on the basis of Anthropic's S-1 filing, but ignores the third quarter results which were posted later.
The growth slowed and operating profit has swung to a loss. The company can still make money, just not at a pace that has a chance to pay off the investments.
Also, the author conflates AI's technical potential with the market potential, naively assuming the demand will also grow as technology matures.
Most of the "laws" of economy in this form say often as a footnote "all else being equal", which remains to be seen. Otherwise you can justify any attempts at price decrease in any industry the same way.
I absolutely disagree number of users is saturated. Fortune 500 companies, for all of their 'AI accelerating' shareholder speeches, are doing very little to actually integrate or spend on AI for the average employee. I've talked to a lot of people who work white collar jobs that could absolutely be accelerated by AI, and they just... aren't using it at all. Some of the companies make Copilot available. Some people just copy paste their tasks into the free version of chatGPT and spend a lot of time correcting the dumb model. A lot of orgs are concerned about data leakage and just have a blanket AI ban.
I recently joined a software dev team in which nobody is using any AI at all. They weren't conceptually against it, they just didn't feel the need to change their process.
> I don’t think this is naive at all; it’s quite rational that demand increases as price decreases, which does seem inevitable.
But that's not what OC is saying, isn't it? Frontier models becoming better and cheaper at the same time are not a given.
Near-frontier models that have 99% of the capabilities of frontier models, at a tenth of the cost of frontier models, might ironically be what pops the bubble.
Is there seriously a non-frontier model out there that can do what Opus 5.5 does? I can say: go look at a ticket in Jira and figure out how to fix it, write end to end tests for it, make sure it aligns with previous architectural decisions we've recorded in documents, and then test it, summarize all the testing you did, and open up a PR for review?
Anthropic, for one, doesn't seem to be replacing their human engineers and data scientists with Opus 5.5 yet even when they would be better positioned to do so than any other company. https://www.anthropic.com/careers/jobs
A 10x will always be beat by 10x+ai. But ai+10x is better than just ai. The world has an over supply of ai (or soon will) but a finite number of 10x coders.
CPUs (and a million other things) got better and cheaper with time, I see no reason that AI would be any different.
The distinction between frontier/non-frontier models doesn’t matter much here, for the overall market. With CPUs, for example, Pentiums got better and cheaper while Cyrix/AMD did the same.
The real question is not if the technology is going to get better and cheaper, but rather if it's going to meet the market expectations to justify trillions of USD in valuation.
Demand will also grow as technology matures, if that maturing means more efficient and cheaper models. We're already at a stage where only a very small niche cares for better more complex models.
People keep repeating this but is it true? In natural sciences, for example, there are endless applications. Anthropic models have only recently reached the level what I would call useful. Still, they are far from perfect and there's a lot of room for growth. Is science a very small niche? I don't think so.
> But the AI industry doesn't consist of just Anthropic.
That is not the claim. Author uses Anthropic's filing as an example indicating growth of the whole industry. While questionable on its own, it is also not given in full, which I pointed out.
The competition is also a problem tho. Both of them are betting the house pretty much on winner takes all. If competition reduces growth, that's a sign that they both can't fulfill all their promises which ipso facto means one or both is overvalued.
I know this is a bit of an ad hominem attack, but it is hard to take the authors arguments seriously as they also argue that the imperial system is superior to the metric system because people in Italy apparently state their height as 183cm (in real life everyone would use 1.83m) and that calculating things using the imperial system is easier as well...
This is just general comment to all the excitable kids over here.
Anthropic is one company in a gigantic ecosystem.
You want to understand the eco system. Got to flyover the jungle, not just stare at one elephant herd.
Read annual reports of NVIDIA, atleast one from Microsoft/Amazon/Google and atleast one representing activities of Enterprise projects (Accenture/IBM/TCS). Then look at govts budgets/programmes for AI US/EU/UAE/S.Korea.
Isn't that worse because you then get into neocloud companies like Coreweave and Nebius who have taken insane debt to build out data centers and if the demand sinks for those they'll go under
It gets worse. CoreWeave has more than $50 billion in debt. rising costs will make financing economically unfeasible. And a demand shock will push the company into bankruptcy. This will affect NVIDIA and others involved in this entire AI supply chain.
What's weird to me is that you don't see a lot of people making deals that seem way too good to be true but yet the total numbers are absolutely gigantic. Throw on the fact that the stranded assets will have a shelf life, it really seems terrifying. Yet everything is being priced as pretty darn risky. I think maybe there can be a big slowdown without a dramatic pop. A sort of grind down. But there's not a lot of room for error to prevent a pop so I guess it will happen? I don't know its very weird.
Also of course this is all predicated on models not making another dramatic leap. The signals out of the cutting edge labs don't really suggest they have a model that would be as dramatic of a leap as Mythos was when it came out in the hopper. Their code quality seems about as mediocre as what the publicly available models can do. The OpenAI math results are very impressive but that's ripe picking for AI. Also they keep talking about a coordinated training slowdown which is textbook antitrust so they seem to want the government to force their hands off the "spend money on R&D" button.
So I don't really know. I guess I would say purely on vibes that this doesn't feel like a bubble but the size of the build out is just so massive that there really isn't a big window for a success case.
This is what keeps me up at night. That fiber that got overbuilt in 1999 is still there spitting photons. That Vera Rubin gathering dust on a shelf will be a paperweight in 5 years.
So maybe it's not high risk enough for you, but it is definitely high risk enough for the market.
But the best approach is just comparing rates. Go to https://archive.epa.gov/epawaste/nonhaz/municipal/web/pdf/bo... , and you will see a huge list of companies whose bonds are classified by armies of analysts as "less risky" than these 10% Anthropic bonds. Is Anthropic less risky than AT&T, or BP, or Firestone Tire & Rubber Co? The market says "no".
currently US long term bonds have very low yield and are being issued less as usual buyers move to gold/stock -- trust in US gov. has eroded especially in fiscal matters. The budget keeps growing, interest rates are still too low to combat inflation and debt is ever-expanding, without drastic measures taken to stabilize the situation: thus, high bond yield - we pay you premium to have at least someone buy it due to trustworthiness not being the sole attractor anymore
Competing with chinese open-weight model which can run openly or on independent providers. Once the temporary hardware shock is over and becomes more easily available, there will be tons of independent inference providers hosting open-weight models. And I suspect we will see dozens of custom, niche distilations from open-weight models being offered by smaller groups.
Anthropic has no value proposition - just being 3-6months ahead in frontier models vs. open-weight doesn't justify trilion dollar valuation.
People keep saying this about open weight models but I haven’t heard of any real companies any of my friends work at willing to switch over to them.
My company buys Claude because it’s a trusted name with a user-friendly interface and integrations with a lot of business apps.
This “open weight models will take over based on price” is analogous saying that Linux and LibreOfficr will displace Windows and Microsoft Office in the workplace because they’re drastically cheaper.
Another analogy is that people talk a lot about the resurgence of audio CDs, but if you look at sales numbers it’s about 4% of peak audio CD sales from the year 2000. Sometimes in society a trend is more that people keep saying that a trend is happening rather than a trend is really happening.
> This “open weight models will take over based on price” is analogous saying that Linux and LibreOfficr will displace Windows and Microsoft Office in the workplace because they’re drastically cheaper.
You can't just literally drop LibreOffice into an organization and have it automatically and seamlessly work.
LLMs are, definitionally, fungible. Their input is tokens. Their output is tokens. They have minor variations in chat templates and can be trained to perform better in a particular harness (like training GPT in Codex).
Do you or your friends at work use redis? Postgres? Are they hosted at "Redis Inc" and "Postgres Inc"? No, they're commodities hosted at a huge number of places.
I went from using Anthropic via API daily to using Kimi K3 (now 3.1 pro) daily, only going back to Anthropic when a refactor is too much for Kimi to handle (once in a month's worth of work).
That surprised me. I didn't expect K3 to be as good as it was and now with 3.1, it's clear (imo) that the gap is closing. Anthropic still has a minor edge but from the 'feel' of these open models, that gap is negligible, not a moat.
The first. There have been massive waves of price cuts in Q3. OpenAI and Anthropic keep slashing prices to take market share from each other, it seems.
To me these are arguments _against_ the bubble popping. The market reacted to all these signals. The semis took the brunt of it but even there it wasn't really dramatic. These are signs of a relatively healthy environment.
Have the models become significantly more capable in the last 6 months? I'm not convinced we're still seeing massive improvements as we saw at the start. The output certainly isn't getting better, the software these companies produce is actually pretty buggy.
I took a six month break from AI coding after leaving my startup and just restarted this week. I feel like Opus 5.5 is a step function improvement from 4.5 or 5, which I built a large codebase with.
Opus 5.5 is capable of completing very long running tasks and managing many threads of work with subagents without getting confused in a way that simply wasn't possible before.
It does very much vary with the domain though. I haven't personally noticed much improvement, they're still useful but unreliable, in my view the idea of agents, harnesses etc has been a distraction and functions to cover up the lack of progress in the underlying models.
But everyone has a different view of this, it'll be interesting to see how it plays out.
Yes, but what's interesting here is that everything didn't get better. The original "interesting" for transformers was that a translation AI turned out to also be good at generating novel prose. But if we now need to focus to train new capabilities into the models and those don't yield better behavior everywhere then we're no longer in the "free lunch" period of LLMs.
Yes, without a question. Fable 5 and Opus 5.5 are a step-change above what we had before. On the open-weights side, GLM 5.3, DeepSeek 4.1 v1017, Qwen3.8-Flash-Next are all significantly more capable in basically any task you can give them.
You can now have a model work on a task non-stop for a day or more. This was unthinkable 12 months ago, and unstable 6 months ago.
If you are searching for evidence of capability, just look at the amount of projects, blog posts and videos on game remixes, hardware reverse engineering, new compilers and runtimes that has started coming out in the last 6 months. Slop? Sure, but this kind of slop didn't exist at this level 6 or 12 months ago.
It's really very hard to pin down progress in this domain, companies are definitely targeting benchmarks.
One bit of evidence I would look at is the maths results this last year which have been impressive, that was a huge expensive brute force effort, but it is still good progress.
i am wondering in what potential world would average user or even business user require for model to work on a task for more than a couple of hours? Is assumption that models on average are used to build nextt SpaceX or something?
The web of circular financing that is going on leads to higher risk. If some ventures fail, and are unable to supply compute they are obligated to, or are unable to pay for compute they are obligated to buy, this could cause a series of financial problems for other related businesses.
I also think that some businesses may find that it might not (yet) be cost effective to focus AI use on every problem.
The pushback against data centers is getting bigger too. But america at least is usually pretty industry friendly so I'm not too sure how much it'll delay planned projects. I've read the limiting factor is sometimes just finding a source of power that can consistently deliver the electricity needed for the DCs that companies want to build.
Yeah, seems like there is a GPU to engineer sweet spot with models like GLM. This could become a more popular pattern for some organizations. Likely also the reason gov. attempts to intervene
The biggest metric is that anthropic make ~$2 for every $1 of compute.
Two things:
1) it requires a massive money pit to make that compute worth anything
2) it requires charging a premium price
To compare to two previous bubbles; railways in the UK and USA.
running trains was really quite cheap and engines weren't expensive per passenger served. Compared to stage coaches, the number of staff needed was < 50%. the Massive fucking problem was that it needed GDPs worth of money to make the railway in the first place. Moreover as more railways were put in it significantly altered profitably of railways near them. Lines that were there for produce moving suddenly lost profitably because the produce they were moving fell out of fashion
Anthropic's S-1 filing. I don't think the parent is necessarily posing contradicting numbers, but removing the benefit of the doubt I was giving anthropic when it comes to inference costs.
Inflation, house prices, gas prices, interest rates, huge public distrust of government and corporations, votes to restrict data center construction, endless war in Iran, etc.
Some of these red flags aren’t real or backed up by much of anything.
- Is it uncommon or alarming for an IPO to be repriced?
- I’m not sure I see strong evidence to show that open weight models are making significant progress to catching up (and I’d argue that the model quality isn’t even where the business value is)
- What does “SpaceX disappointment” mean exactly?
- Are enterprise buyers buying raw tokens or are they buying productivity tooling for their employees? Not many companies switched to LibreOffice because it was cheaper than Microsoft Office.
- Look up BLS unemployment data, what job loss and entry level hiring collapse are you talking about?
- Does discounts on debt point to anything? Genuinely don’t know.
- What about the Chinese arms race is concerning? Isn’t an arms race evidence of value? There wouldn’t be an arms race if this technology wasn’t worth investing in. Doesn’t China’s desires to compete in this market bolster its merit?
- What is “possible downward spiral” and how do we know that’s real? Isn’t there a “possible downward spiral” with every industry at all times? This seems like pure speculation.
- job loss and entry level hiring collapse --> I will out that as in the group "Not a bubble" (Not that I am thinking it myself that it's not a bubble).
It looks like the end goal of replacing labor by capital is working.
With big technological shifts like this, it's important to keep in mind that the ability to do more with less resources can pan out in two different ways: (1) Keep doing what you're already doing, and lay off the people you no longer need, or (2) Do more than you did with the same amount of resources.
The answer depends on the company, but I think many companies are still figuring out whether and how they can exercise this new capacity. We don't know where it will land yet at a societal level. It's a parameter shift, and it will take some time. Pray that you live in a place with sensible labor laws and safety nets.
No. The biggest layoffs are happening at companies that need to free up additional capex for infrastructure buildout, because they're not getting ROI from their existing capex, and they're having trouble finding more money.
I think that's actually going to moderate the bust, by giving the big companies the moat they need.
It performs the same function as regulation, replacing "the scale to navigate and manipulate the regulatory environment successfully" with "the scale to navigate and manipulate the political environment".
It's essentially grassroots regulation instead of federal regulation.
(I don't think it will stop the bust, but moderate it, allowing more companies to survive.)
I generally agree with you, but have to get this off my chest.
> pareto frontier gap decreasing by the month between open and closed models
We need a temporary moratorium on the words "pareto frontier". Every new model is the pareto frontier by some combination of metrics. Like "Bayesian Priors", people love saying / tweeting it because it sounds Smart.
It's like watching car commercials where every single vehicle has won something from JD Power and Associates. Which, (a), who the hell are they? and (b) you can only pass out so many awards until the gluttony of awards proves the award is worthless.
Agreed. This applies even to simple words like “velocity”. When you keep seeing the same key words/expressions used in support of the same argument, you know the person is parroting an idea as opposed to sharing an original thought.
There is no such thing as a moratorium on words. It's another phrase that people like to repeat (a meme) until the excitement or prestige of using it has waned.
That said, i agree with you that word-salad investobabble is a sign that the bubble has put many people in a raving fever. This current bubble manifests like the cryptocurrency mania.
Also - all the gee-whiz AI companies are headed by people who don't have great business backgrounds, either by people from tech backgrounds (chasing the last of the VC), or by people with the kind of "interesting" background who you wouldn't lend your own money!
> - pareto frontier gap decreasing by the month between open and closed models
This is a problem for Anthropic and OpenAI, sure, but why does it matter for e.g. Nvidia (or any of its suppliers) who will use their GPUs for training?
We need to separate the model companies from the shovel-sellers in these "AI bubble" conversations.
Maybe this is what you meant by "Chinese arms race" but there is, though not quickly, a massive wave of increased hardware supply incoming. Again, let me mention this isn't coming soon in terms of the AI bubble conversation. But China is pushing for a home grown EUV and fabs that should start coming online and having an impact in 2030.
The reason that is important is it will fuel the competitiveness of on premises and local open weight models.
Though maybe it's less useful in the "bubble" conversation since over four more years you could fit in various booms and busts before the supply side significantly starts changing.
Still useful. Especially if Chinese GPUs operating on Nvidia software become performant and accessible, thereby undercutting the price gouging that we're seeing with western chips.
Access here means being allowed to participate in the market. BYD doesn't have a problem shipping cars, it has a problem of being allowed to sell them.
People will need compute. And if the Chinese economy can transition into providing not only factory output but cheap compute, then yes - the future is indeed theirs. Every single country on this planet will need compute.
The EU is also quite skeptical of Chinese exports, though it remains to be seen how that will evolve once the EU-USA relations settle into a new normal. Certainly the EU wouldn't be happy, though, if China managed to break ASML's monopoly on EUV machines, and at seek to retaliate with economic pressure on chips.
problem is if you can just rent cheap compute from elsewhere, you'd need to ban buying compute and all that which I'm not so sure would work. maybe heavy regulations could do it buuut....
unless I am mistaken, you missed that the datacenter infrastructure being a one-time cost. nope. those servers have about a 5 year lifetime and then are likely to be replaced. The buildings aren't the expensive part. The GPUs are.
so I think that's one big miss from this blog, or I don't understand datacenters as well as I think I do (which is always a possibility! hopefully someone will inform me if I am wrong)
> those servers have about a 5 year lifetime and then are likely to be replaced
This made sense in the past when they were not prohibitively expensive, but other than them breaking down why replace them on the 5 year schedule now? Yes, the newer stuff is better but the jumps from this point forward seem more incremental than revolutionary hardware wise. And even then, a company can simply swap out racks as parts fail.
As the friendly article points out, the "AI is a bubble" hypothesis and "entry level hiring is collapsing because AI is replacing their jobs" hypothesis are not really compatible.
As always, reality is likely a muddled mix: there is real utility and there is wasted investment that will never pay itself back.
They actually are compatible, because the idea of replacing workers with AI in large numbers isn't compatible with the kind of economic growth these valuations are predicated on. We are not living in pre-industrial revolution times, where most people are wasting huge human effort on food growth and physical tasks, having room to move into more white collar jobs.
So, if you have massive unemployment because AI really is cheaper than labor, you'll get a massive reduction in revenue for virtually every industry, collapsing all of these valuations. Even for B2B companies, Google isn't going to be making much money selling ads to unemployed people who can't afford to buy anything.
My reality: Before AI Tokens I had to hire someone. External or normal.
After AI Tokens: I don't need to hire someone anymore and if, i need to know why. Because if i have an contractor for 1-2 years and explain a lot to them, i can just explain it to AI.
Also i prefered before already to not hire someone if i had the feeling they don't have potential, now i have even less incentive to do so.
Yeah absolutly. The question though will be if capitalism is able to update its gain function somehow or if the system collapses.
As long as robots are not there yet, there might be enough of buffer in the whole system for non computer jobs to catch but i'm not seeing this either.
It will increase pressure on every industry.
And it might also lead to an increase in speed of adoption because everyone getting fired, will have the tools to fill any niche.
And because of massive overhiring. In that scenario, the sequence of events is: Cheap money (ZIRP), companies hiring developers just to deny them to the competition (there were reports) or for empire building or as a signal to investors (etc), AI and nonzero interest rates come around, need to spend billions on AI with limited money -> lay off overhired developers and say it's because AI replaces them. It's great for companies because it allows to admit no mistake regarding the overhiring and claim to be fully in on the current thing.
No one seems to worry about the demand of ai. That makes me wonder. My neighbors for example are normal working ppl they don’t know how to open a zip file (they call me to help them with something like this) and they are in their early 40s so not that old. Sure companies and workflows are going to become agentic but not everyone needs it or even wants it. Then there are smarter small models like Qwen 27b running on a Mac mini and this will also continue to get better and develop further. If the usage is easy I see ppl downloading a desktop app that’s dead simple instead of paying for ChatGPT. I just don’t see this growth curve going for 20 years for these companies to pay off their dept
I would say your neighbors are the perfect use case for AI. AI can understand human language. Your neighbors probably never were able to use Google for search, compare flight tickets online, Wikipedia etc.
Give your neighbors something like Alexa powered with better AI, and they are sold.
Having dealt with many family members just like this, I can tell you they certainly can use Google for search, they just don't want to. When forced to they use it fine.
If I were to guess, a lot of people, particularly older people, see computer use outside the web browser as an amateur might a chop-saw; a dangerous tool not to be used by inexperienced hands, so an AI that can also directly do the tasks for them might solve that problem (assuming it came preinstalled or someone else did that for them). Though then comes the question of whether they'd be willing to pay for it
Plus hallucinations. Most "regular people" questions aren't inherently easy to verify. And when regular people do have to verify the answers, the value of "AI" drops like a rock.
> Having dealt with many family members just like this, I can tell you they certainly can use Google for search, they just don't want to. When forced to they use it fine.
Exactly! I noticed this when I moved out of my parents' house, when suddenly my mom became an electronics whiz who can hook up VCRs and TVs just fine. I always thought I was the smart person in the house, but now I realize I was just taken advantage of.
Based on my experience, you’re right. I know a woman in her late 60s that is basically tech illiterate (can barely send an email), and she has been making effective use of the ChatGPT free plan. AI has been helping her cut through the noise of navigating the web in 2026 (which in her defense is often a bewildering experience). It can’t help her with everything yet, but it’s a clear boost to her independence online.
I'd say the exact opposite. People who have no idea how technology works shouldn't touch AI at all. "AI can understand human language" is fundamentally false, and it leads people to actually believe AI is "AI" and not just a token predictor.
These people WILL fall victim of their own ignorance and upload all their private information to an AI chatbot.
There is no "honest" AI though. It is always biased towards its prompt and ultimately its training data.
Compare flights? You have to assume it is objective and not tainted by corporate bias or backroom deals(aka ads and sponsored content), unless you selfhost and find an objective model.
Search Wikipedia? You have to constantly check that it is giving you straight facts instead of being constrained by its community and safety guidelines.
The problem with so many of these usecases is treating AI as a super intelligent Vulcan who is always logical and objective aka Spock, when we are better off treating it as if it is a salesman.
This is all true, but which travel booking assistant is really unbiased?
Online comparison sites might not be, the good old 'brick and mortar' travel assistant most likely isn't either.
You don't need an unbiased one. Only someone who cares about the consequences of getting caught while being blatantly unbiased.
Your neighborhood brick and mortar, or an online one may care about that while a faceless trans-national big tech will not even register that you left.
agree. That's why i love models (agents) that just search the web get the answer and use tools to their disposale instead of guessing based on training data which is outdated anyway...
Also, their neighbors are also the exact exemplars of what two generations later will look like. There isn't a lot of hard data on what AI usage might look like, but there is data on tech literacy, and the numbers are clear: We trained precisely -ONE- generation that actually understands computers. The incoming one will happily gobble up all the AI, same way they gobbled up willful ignorance of filesystems.
Much like there is a button that corresponds to "gimme shot videos", there will be a button that corresponds to "Listen to what I'm about to say, and do it".
i dont know about that, they could've adopted using google which you could already search things in pretty much natural language and even got youtube snippets how fix your problem... so i'm not convinced they will adopt
Right. So my worry – inasmuch that I worry for the state of the global economy – is that this doesn't generate valuations that match the current hype.
There's so much ad money to go around. It currently fuels Meta and Google. Let's say Anthropic and OpenAI enter the field and take a slice. That's a fine business, sure. But does it justify the state of the market?
It is a bit comical how the HN crowd spent the last number of decades trying to convince the audience that a programmer's value was actually in their system thinking, not their arcane knowledge of a particular language, but now that the language barrier is removed it turns out it was just the arcane language knowledge after all...
Although I'd posit that the old school HNers were the ones actually right. From what I see in the average Joes is that, while they can communicate with AI, they don't really see how to leverage it. Granted, the younger generations grew up in a school system that tried to promote system thinking more so than in the past, so that gap will likely shrink over time.
This year, a lot of big companies are building the agentic layer.
This has to be build and takes time.
But this is the major thing were ai will be introduced into every single layer of software we have.
The HR System which might have needed x people? The agent has solved 90% of tickets already before they hit a human. Agents will do the repetitive but not that easy to automate tasks.
This is the point. hmontazeri is imagining where AI still has left to grow as a consumer product, but AI is currently more of a business product. Even without a new consumer use case, it's easy to see demand for tokens keep going up.
The agent will be running 24/7 and be triggered by different things. Chat input, scheduler, actions someone executes inside the software etc.
The SEO campaign tool? You might need to fill out a lot of different things, configure stuff then execute. In the future you will start the SEO campaign tool, the agent will ask you missing information, it then fills out everything, gets a budget and runs it. It will adjust the campaign based on feedback from the SEO tools.
Kind of a workflow engine but smart, so smart that it will handle A LOT more than what an traditional workflow engine could ever do.
>My neighbors for example are normal working ppl they don’t know how to open a zip file (they call me to help them with something like this) and they are in their early 40s so not that old. Sure companies and workflows are going to become agentic but not everyone needs it or even wants it.
Maybe not everyone (yet) but your neighbours clearly do need it.
It's definitely an improvement. Most people would not want to depend on their neighbours for opening zip files. I know I wouldn't.
And if one of my neighbours ever asked me to help them open a zip file, I would not leave before I'm absolutely sure that they can do it themselves next time.
Feels like collectively we are under indexing on LLMs as user interfaces. I use a CLI with an agent to things I once did in a GUI. Your neighbors should never have to worry about zip files again.
having to write "hey LLM-du-jour please unpack these files!" os strictly less efficient/more time consuming than clicking "extract" on a menu. I have hard time what usecase you have in your mind. (Unless the user is blind that is)
I'm sorry but if somebody born in the 1980s cannot or will not double click on a zip file to open it they also cannot or will not ask Claude to open it for them.
> companies and workflows [] going to become agentic
What makes you so sure? Some would call it "automated running scissors" and want tools firmly in responsible hands. And could assess "automated (indeterministic) reliability" as still far away in time.
Are there any contexts we may have missed that would surely, safely replace individual deliberate action with "the ANN can do it"?
As in, "machines cost less"? Well, this general principle has always been present, as in "sucking rocks costs much less then cordon bleu", but there are "human costs" (U. Eco said for an equivalent area) involved.
Of course spending less is alluring: but you have to be able to afford spending less. Also considering that giving jobs to people who deserve it prevents other general "boomerang" social costs.
Im not convinced by the claims of infrastructure only being expensive to build initially. Hardware doesn't last forever especially in a data center, maintenance isn't free, electricity is not going to get significantly cheaper. If models get more efficient we won't need super expensive and expansive hardware build outs, if models keep getting more powerful and needy and are hardware limited, each new hardware generation will stomp over all this expensive and soon to be old hardware in electrical efficiency requiring another massive round of funding that players already holding debt will struggle to afford.
> Im not convinced by the claims of infrastructure only being expensive to build initially.
The claims seem to be true.
> A typical one-gigawatt AI data center requires $38 billion in up-front capital expenditure (CapEx) and $0.9 billion in annual operating expenses (OpEx). If CapEx is annualized over each asset’s lifespan, the total cost of ownership equates to $8.5 billion per year. Servers dominate this cost at $5 billion per year, or 60% of the total. Operating costs are small by comparison: even energy, the largest OpEx category, costs only $0.6 billion per year.
If new hardware comes online that doesn't allow you to payback your billions in hardware, what then? The level of rigor in the article is low.
If we steel-they the point in the essay, it still doesn't matter because a bubble is setup in the economics of the system, not whether the underlying tech works and/or is profitable. It can both work, be profitable, but have been overcapitalized before the knock on effects are in-place in the rest of the economy. This is a crazy analogy, but what if 1700s earth figured out, theoretically how to build a fusion reactor that could supply all the energy needs of the whole planet. But it cost in their equivalent dollars, 1 quadrillion dollars? You can't even integrate under their money curve to scrape together that many dollars. The only economy you can make about that is in the research side to push the cost down. Then when its dialed in, you can launch that rocket AND the rest of the system has had time at many different levels to integrate AI at various levels. OpenAI dry barebacked its way into the economy and literally screwed everything up. If they really were an altruistic AI safety lab, they should have kept working on ChatGPT, but demoed it to the world including government, so we could as a planet decide where we are taking this.
It is a huge bubble, AI is over capitalized, we have no idea what the knock on effects are and really the only benefit right now is in programming, which is huge and civilization changing. But it also has massive downsides in that it has given the most powerful even longer levers to move the earth.
I just don't think the author really has any clue.
they seem to think that debt markets are infinite. "There's Still Plenty Of Capital To Raise"
> The ones that have issued public debt, like Alphabet and Meta, have excellent bond ratings. If they need more capital, they won't struggle to find people to give them a loan.
Yes, and that bond rating is based on the idea that they really don't have that much debt, and have a net profit of something like 30%. If they hoover up a load of debt, the rating will plunge.
> I don't think the rhetoric about the AI bubble is really about finances
Its exactly about finances. The market participants that are not bollocks deep in fake Private Equity valuations are worried about circular debt obligations that are off books. The SEC has been rather lenient about reporting of non-traditional obligations (hence why spaceX was allow to do what ever the fuck it did)
For example anthropic has about 450 billion in non-cancellable obligations. the S-1 puts it at 895 billion. all of that is someone's debt.
If openAI are doing the same thing, thats pretty close to 1.6trillion in debt, both for the same customers, at the time when customers are going "hmm this isn't giving us what we hoped for the price"
Agree. CDS on the debt of several players here is getting expensive. Debt markets are getting less promiscuous quite quickly, just when even more money is needed.
It could be that a slowing of data centre build (and hence a slowing in the finance needed) takes the heat out of the situation avoiding a dramatic burst. Even then there will be casualties. If someone knows who and when ... they could do very well. More likely, Microsoft and Google will buy up anything distressed that's worth having.
Similarly, the claim that "Once all the data centres are built, AI spending is going to plummet" feels incongruous with "it's hard for me to see AI consumption not being 10x what it is now" and "Anthropics revenue has to grow to 250 billion a year to pay off debt". Scaling up seems like a constant need, especially in a market that values growth and hype about frontier models so much. (that said, the author raises things like P/E ratios that I had no awareness of, so I'm not claiming to be an expert)
The debt thing is amazing since it ignores the interest on the debt and cherry picks Google's revenue from 2019 for some reason (which is then adjusted for inflation for the comparison).
If you have 20 years to pay off 5 trillion dollars, it doesn't come out to a neat 250 billion per year as that suggests there's 0% interest and no servicing cost. Servicing costs would probably be at least 3 times the stated figure, corresponding to 10% interest paid. Some of the debt may obviously carry less interest, but the financing is opaque (as the author admits to), we have seen interests rates creep across the economy and big tech has had hits to their credit ratings.
I hope the Liberal Party can find a better candidate than the author of this piece ...
That is a weird take and after that I sort of just skimmed the rest. Yes building the datacenter is a huge expense, but you need to replace the hardware to stay competitive, and it doesn't last forever. Previously companies ran three year cycles, now five seems to be the standard. If hardware prices come back down, then.... maybe, if not those refresh cycles are going to put some datacenters out of business.
Another issue is that while e.g. Anthropic and OpenAI seems to be interested in building their own datacenters, they are mostly just renting capacity. If they can shop around, then they can get the price down, but there's a base price, due to power, hardware and staff.
The author also suggest that there's still capital to raise... no there's not. Anthropic is going for an IPO, that means institutional investors, not venture capital or borrowing and Anthropic, CoreWeave and Oracle are rated as junk. That doesn't mean that they can't raise capital exactly, but it will be very expensive and the interest payments needs to be factored into the price of the products.
> electricity is not going to get significantly cheaper
The DC push may be the thing that finally gets renewables over the finish line in the US. For example, TX is around 35% renewables right now. If that continues to rise, we may see power fall in price. It’s all about incentives.
Without a requirement for DC's to add more power to the grid I don't think the price is going down, they're essentially adding to demand and the supply isn't particularly elastic. That being said this is a good opportunity for renewables, they're the cheapest form of power.
And we just saw a pretty strong repudiation of that type of DC buildout. I have a little hope politicians can communicate it’s not DCs that are the issue, but the externalization of all the costs.
It hardly seems to be a one time expense. GPUs and networking equipment may have a useful lifespan of only about five years before needing replacement, and it could be expensive to do so.
The current-gen hardware is expected to have only 2-3 years of useful life. New tech, and especially new architecture is going to make inference much faster and more efficient in the near future, and the big vendors are going to be more limited by datacenter floor space and power/cooling, so that the hardware will be replaced very quickly by better models.
For the time being this thesis hasn’t worked out though. H100’s are only going up in price and becoming more scarce [1]. At what point would you update if now going on 3 years into H100 lifecycle this doesn’t seem close to happen.
But they are becoming more scarce while staying on Nvidia’s shelves buy-and-hold, or being leased back in some cases. They are not being installed by customers at anything like the rate the sales figures suggest. So yeah, these H100s will last a long time in the packaging.
Meanwhile, Meta is selling their spare AI data centre capacity.
They are in Nvidia warehouses. Nvidia is selling H100s and other hardware in advance to firms who do not have data centres to put them in.
Nobody is getting fired for thinking they need to buy scarce H100s in case they aren't able to get them in the future. Nobody is even getting fired for keeping their companies out of Jensen Jail — the fear that if they don't buy the whole server stack from Nvidia they might not get future GPU allocations.
There really are analysts who think Nvidia is channel-stuffing.
If the market deflates there will be a glut of very cheap compute, both in terms of data centres (though many of them will be able to offer managed model training services instead, I suspect) and in terms of unused hardware being made available for sale.
People are buying v100s on ebay to run LLMs and Chinese manufacturers make custom PCIE boards for them. Older cards might not be useful for training but they're perfectly capable of doing inference work.
Look at DeepSeek! They have done full system optimization on their model, architecting it so that it would be cheap to serve. I am proud to call DeepSeek a world class tech engineering firm in a way that 99% of the tech world is not. They are a 1% company.
> The current-gen hardware is expected to have only 2-3 years of useful life.
Expected by whom and why?
Google hasn't retired any of its now ten-year-old TPUs, or any of the subsequent generations. Maybe TPUs are special or Google doesn't understand compute economics but either argument requires data.
If your argument is depreciation, you're confusing a tax concept (depreciation) with useful life. Yes, they're supposedly related, but in practice, the relationship is rather loose. (Residential housing depreciates in 28 years yet typically lasts 2-3x as long and 4x is not rare.)
Google's V1 TPUs
Profits are mentioned but then only revenue is discussed, Revenue != Profit. Revenue can grow yeah but experimentation/token-maxing stage is ending and price wars have started, so I'm not sure about rising profits, AI Labs won't necessary be able to capture the margin(>>China<<), it might be companies/consumers who will capture most of it, if so there is a bubble - it is in AI labs with their trillion $ valuations and they might end up in a race to the bottom, BigTech is not in a bubble for the most part, they will endure a correction and fall back on their existing business.
The entire premise on whether it is a bubble depends on if LLMs can be more than a next token prediction machine or not. Adding reasoning agents, agent swarms, and so on to enhance the results from what is essentially just predicting the next token still means this technology at its core only predicts the next token. It is not AGI and unless something fundamental changes, it never will be. There are certainly valuable applications for this technology as we have seen in coding, math, and other sectors. But it is still not more than taking in a lot data and predicting the next token.
Meanwhile, the major tech companies first financed this build out using the cash flow from their existing revenue. Once this became more difficult to justify, they literally set up companies to keep this off their books directly, and these companies are ultimately designed to take a loss and fail. But before reaching this failure stage and defaulting on debt and obligations, they went to the criminals on Wall Street and designed a financial product structured not all that different from mortgage backed securities. Also, as the "AI" aspects of their existing businesses expand, it eats into the traditional revenue generated by their core business, which if we remember, is the original collateral being used to justify all the debt in the first place. Now these core tech companies have to win "AI", whatever that ultimately means, to not only pay the debt from the build out, but to also replace their existing collateral from their original products. Think in terms of Google Search being replaced by AI. Maybe a clean transition, maybe not. Really just depends on the technology living up to the hype. If not, the hyperscaler companies designed to take a loss a fail, will fail and with it, the circular financing will pop the very real bubble that exists.
Speaking of IPO valuations as well, it's not to say that LLMs are not a valuable technology. Just look at what they have done to computer science already. But there may be increasingly limited returns on investment if every valuable problem to be solved isn't as well defined as logical computer science problems. Without solving the hallucination problem (as a result of the next token prediction limitations), the wide application of LLMs as being sold will never happen.
I never understand the “only a next token predictor” argument. I do share the view that something fundamental about intelligence might be missing in LLMs, but I don’t see how this argument captures anything. If you can predict the exact next word Einstein is about to write, every time, you’re as intelligent as Einstein! It seems very clear that the LLM is also “thinking” about where the output is headed after this next token.
I do hear this position quite a lot though so I’d be interested to hear more about what I might be missing.
The main criticism is, imo, that these models cannot think and have no logic. This is proven time and time again. Every day something new comes out, like counting Rs or the ”drive or walk“ test etc. These are the ”fixed“ or rather band-aided on a case by case basis, but nobody seems to be able to fox the underlying problem.
Having done my master‘s thesis on Transformer LMs, I think this is a fundamental limitation of the architecture.
These models can solve math problems and then fail to count the number of problems they solved. Not being able to count to 68 is not intelligence.
It's a statistical machine, so you only predict a statistical Einstein, ie. something that might appear quite like Einstein and even can be fairly accurate in predicting the next word he types, but which is not him, and will be unable to predict the words that are produced not just statistically but with certain intelligence, direction and intuition.
“It” is not “thinking” for christ’s sake, it’s a massive statistical regression curve fit to a massive, frozen library of text. Unless you’re thinking of “thinking mode” which still is just a template that tells the model to ask questions, which are asked one token at a time, turtles all the way down.
LLMs cannot, by their core design, predict next word correctly every time.
That is the main thing you are missing.
Another thing is that I know that I am not predicting next word based on othr words in my "trainin data", there is much deeper understanding of most subjects that I am using to "predict next word".
This may not be true for everyone, I cannot know for sure, but it might be possible that there are people out there who are just biological LLMs, and in that case I am certain those people will not be able to predict next word Einstein would say either.
> If you can predict the exact next word Einstein is about to write, every time, you’re as intelligent as Einstein!
Einstein is not hold in such esteem because he was really could at spewing out text or formal proofs or theorems, but because he was able to invent something no-one else could. Some people say he advanced physics 100 years wuth his theories in relativity, whether it is true or not.
So even if you can predict what Albert is going to write next, you are still merely copying old Albert. They are amazing at processing large amounts of data, naturally, but creating something new and surprising we have yet to see evidence of (the solutions of the millennium problems seem to be brute-forcing a solution based on fuckton of data to me, but I am not a mathematician so I guess it is up in the air still).
Is there any evidence that you yourself are more than a 'next action predictor with a multi-modal context window'?
I don't think the nature of 'predicting the next token' is the problem. One could argue that any organism works the same way. The problem is that we've had million years of evolution to stabilize the platform such that the important parts of it work without that many hallucinations -- and even then that sometimes fails and you get failure modes like e.g. schizophrenia, where the context window gets confusing and the 'next actions' don't make that much sense anymore, seen from the outside.
An LLM never experienced evolutionary pressure the same way. There's no fundamental difference between how it was incentivized to answer correctly when asked about how many Rs there are in strawberry vs whether it is safe to leave a house through the second story window, even though for any physical organism, those two would be _very_ differently weighted during training.
> Is there any evidence that you yourself are more than a 'next action predictor with a multi-modal context window'?
I love the age of LLMs because people will post stuff like “the fact that I absolutely, unqualifiedly, fundamentally do not understand two separate things indicates that I understand at least one and maybe even both of them more than you do“ and then do that cool move where they pretend they were gonna high five you but don’t
Is your post any better? You attack GP's character without addressing any of his arguments. You could have provided evidence for why you think his speculation is wrong, or tried to point out how his logic falls apart.
It's because of the proliferation of "insight-posting" - people thinking their little insightful "aha" thoughts are unique or have any kind of value at all.
However, insight born of knowledge ignorance is worthless. I really don't care for your great insight about nuclear fusion if you're not a nuclear engineer or physicist working on nuclear fusion, because your insight is just a hypothesis in the void that is your ignorance.
Ironically TFA falls victim to the same thing. It doesn't seem like the author is an economist or an expert of corporate finance. Their assessment of whether something is a bubble or not has absolutely no value, and carries with it the arrogant implication that the hundreds of thousands of people who are actually experts at these things never had the little 2AM epiphany that they had.
It's a smaller version of the same delusion that produces crackpots.
> entire premise on whether it is a bubble depends on if LLMs can be more than a next token prediction machine or not
I don't think current valuations need AGI to work. They just need enterprise and personal spending on AI to keep rising exponentially for a while.
We could absolutely wind up in a case where these companies are fairly close to a trillion dollars each, but not multiple trillions, at a reasonable multiple, all without AGI.
> The entire premise on whether it is a bubble depends on if LLMs can be more than a next token prediction machine or not
A theoretically perfect next token prediction machine has limitless utility. Predict the weather, stock market, lottery number, the next step of groundbreaking cancer research, etc etc. That's not the argument against current LLMs being capable of being AGI.
A theoretically perfect next token prediction machine is just an LLM with a temperature of 0, which performs worse than one with a non-zero temperature (i.e. imperfection)
The "prediction" has nothing to do with predicting the future, it just means choosing the most likely next token for existing human text
If that is all you have then the next chess move will be random (or whatever token is more likely in isolation), or the stock prediction will be the same.
For the chess example to work you need the current board state in the context or the move history. How an LLM performs will depend on what it has learned across the transformer layers and in the residual stream, as well as how much chess data it has been trained on. -- NN models like Leela and Stockfish have been trained on thousands/millions of games and have a game-specific state space representation.
The stock prediction example is similar, but doesn't have a consistent state space in which to make that prediction. It will need information in its context (current news articles, stock history of all stocks, etc.) in order to be able to make a prediction. Its performance will depend on whether it has been able to learn factors that make it possible to make accurate predictions.
> The stock prediction example is similar, but doesn't have a consistent state space in which to make that prediction. It will need information in its context (current news articles, stock history of all stocks, etc.) in order to be able to make a prediction. Its performance will depend on whether it has been able to learn factors that make it possible to make accurate predictions.
This also presumes that accurate predictions are possible.
It's the transformer architecture (token embeddings and cross-token encoding) and residual stream (cross-transformer encoding and token differentiation) that make LLMs powerful tools.
The key question is not whether this architecture is powerful. The key question is how well these models can represent things outside of their design.
With chess for example, are they actually learning the rules and strategies of the game (like AlphaZero/AlphaGo, Leela, etc.) or are they just relying on a large corpus of chess games? For example, a lot of early game moves will have been played out and the best moves will likely be the more frequent moves in the training set. So LLMs could do well early game based on that game corpus.
Predicting the next token well is indistinguishable from intelligence. The bubble is financial, as you point out, just like railroads, radio and dotcom bubbles were. The technology didn’t die, only the companies did. LLMs are in the exploit phase, that doesn’t mean there isn’t more to explore and there may be an architecture-hardware combination that takes us an order of magnitude or two higher in a step-function-like fashion just like LLMs did 2024-2026.
But despite getting better and better at not making stupid mistakes, the models are still just next-token predictors. There’s plenty more room for them to improve without a totally different architecture.
The entire premise on whether computing is a bubble depends on if computers can be more than a bit-flipping machine or not. Adding operating systems, compilers, distributed clusters, and so on to enhance the results from what is essentially just flipping bits still means this technology at its core only flips bits. It is not a general-purpose machine, and unless something fundamental changes, it never will be. There are certainly valuable applications for this technology, as we have seen in ballistic tables, census tabulation, and other sectors. But it is still not more than taking in voltages and toggling transistors between 0 and 1.
> Adding reasoning agents, agent swarms, and so on to enhance the results from what is essentially just predicting the next token still means this technology at its core only predicts the next token. It is not AGI and unless something fundamental changes, it never will be.
A lot of people have dismissed LLMs four years ago for this reason. "It cannot possibly do coding since it doesn't think, it just predicts the next token." On the other hand, Scott Alexander did say in 2019 that next token prediction is a brilliant idea and might cover a lot, including math proofs [1]. So I wouldn't say it is impossible for either 1) LLMs to be all the way to AGI or 2) LLMs to become creative enough to tell us what the AGI architecture is.
They've global procurement, the franchise model means the franchisee shoulders a lot of costs, they own the land the restaurant is on in 50%+ of the locations, and they own the technology system that underpins their app + ordering.
An economy of scale, which McDonalds surely has as GP suggested in various areas, supports durable profits. Seems like a moat to me.
Land is the classic cornered resource. Again a source of durable profits. You can’t open up a franchise restaurant on land owned by McDonalds; the profits from traffic earned in that location are not accessible to you.
CUDA is not a moat because it happens to be unique. CUDA provides a moat for NVIDIA due to the high switching costs and network effects it provides.
McDonalds to my knowledge does not have any network effects dynamics going on however they have entrenched switching cost mechanisms in the form of loyalty rewards programs.
If you think McDonalds has no moat you’re dreaming.
Maybe we just have a slightly different definition of moat, mine focusing more on intangible assets and switching costs.
I don't think the land of McDonalds is a moat. A factor increasing profit, sure. But anybody can by land across the street immediately, only needs money.
Even with infinite money, it is questionable when and if you can cross the moats of ASML and NVIDIA.
Can you describe how that is a moat? A structural advantage that prevents other businesses from taking their customers, or their customers from switching. For example, how does McDonalds real estate holdings make it harder for diners to choose Burger King instead?
It gives McDonalds a lot of intrinsic value. It’s definitely a core part of their business. But it doesn’t prevent people from going to Burger King.
They own restaurants in key touristic areas that have massive amounts of potential customers going around. This is a structural advantage because people often just pick the easiest option (especially when it comes to fast food).
For example, in Paris they have two restaurants on Avenue des Champs-Élysées. While there's a Burger King right around the corner (Rue de la Boétie), a lot of hungry tourists will simply not walk in front of it.
I dont know if the valuations make sense, but the usability is all there and so are the potential profits. Over a decade, that's probably all that matters. If the valuations pop or dip (anytime soon), I think that's mostly irrelevant. I get that the press probably doesn't like to hear this boring conclusion.
P.S: A Matlab/Physics simulation guy running for a Liberal party seat in Canada? Say more!
The decade after the dotcom crash brought about as many if not more long-term successful products as the dotcom boom: Youtube, Reddit, Spotify, Google Maps come to mind
A bubble that funds all kind of crazy ideas and big investments into infrastructure work is great for laying the groundwork. But the end of the bubble won't mean the end of progress, or even necessarily a slowdown in progress
In 2007, It was regular people who lost their homes and their leveraged investment properties. Regular people are not leveraged into AI startups today. If the AI bubble pops, regular feel people will feel the direct financial consequences significantly less than the '07 housing crisis.
> Regular people are not leveraged into AI startups today
this isn't accurate. as an example, anyone with a 401k will be affected by the stock market and the stock market is mostly tech, and tech has bet the house on ai
"Your 401k should never go down at all" is not a realistic expectation.
Hopefully, people who need to pull from their 401k's soon have wound down to less risky positions. 401k exposure is just not the same thing as your house valuation tanking.
For certain values of temporarily. On human time scales, we do see bear markets which last a decade or longer, and that can have a serious impact on someone's earning ability or retirement.
Markets do include significant risk and we would do well to remember that.
Unfortunately, they are, the market (even the global market) is heavily concentrated in just a few US listed companies, and other indexes closely track US indexes. A crisis in AI would impact everyone.
For proof, look at the fall in VWRL last week
correlated with concerns over AI which hit nasdaq.
> In 2007, It was regular people who lost their homes and their leveraged investment properties.
And the debt on which these owners defaulted had been converted into CDOs and mortgage-backed securities, it was financial institutions and investors who lost enormous amounts of money on these.
Today there are GPU-backed securities and compute futures that are helping back all the debt that is financing the datacenter boom. Maybe it’s a bubble, maybe it’s not a bubble, but either way I think its parallel to 2007.
I'm not imaginative enough to see more use cases and who will pay. You can burn a lot of tokens and dollars on math proofs. Same for white hats finding vulnerabilities in all the software. Generating video.
Personal assistants dealing with your email, filling out paperwork, and booking reservations don't use a lot of tokens.
If every software developer paid $200/month, every academic paid what they pay for Matlab or Mathematica licenses, every YouTuber paid for whatever video processing, and every normal person paid a subscription comparable to their TV subscriptions and cell phone plans - does that justify the valuations?
The government pays for surveillance? Companies pay for customer service replacements and target advertisements?
It seems like there are a few small categories that use a lot of tokens, and a large category that doesn't need many at all.
> If every software developer paid $200/month, every academic paid what they pay for Matlab or Mathematica licenses, every YouTuber paid for whatever video processing, and every normal person paid a subscription comparable to their TV subscriptions and cell phone plans - does that justify the valuations?
> The government pays for surveillance? Companies pay for customer service replacements and target advertisements?
Yes definitely. All together that's probably within striking distance of a trillion dollars in annual revenue. Just TV subscriptions + cell phone plans in the U.S. alone is about 400 billion (leaving aside any other country).
Most large, mature tech companies are at 5-10x revenue to valuation ratios (and 10-20x EBIDTA to valuation ratios).
Of course whether you get there is a separate question, but any company that fulfilled what you're asking would pretty easily justify multi-trillion dollar valuations.
That $400B for video and phones though. That's like a $1k average for every adult and child in the country. It's twice what I spend, and I'm not on a family plan.
That's significantly underestimating the potential. Companies typically spend $200k/year for developers (salary, benefits, taxes, equipment, office space...). If AI can make them even 25% more productive, that would justify $50k/year in tokens.
I more than agree on the value, but not on the price. There's currently a competitive market, including free models that are pretty impressive if not "frontier".
I don't think LLM inference is ever going to stop being a competitive market, though. Unless regulation or other factors make the existing players form a cartel.
I think it's pretty safe to say that, unless the government steps in to prevent it, they will form a cartel or some sort of oligopoly. The incentives for them to do so are very high.
Yeah, we need a lot more kinds compilers like LEAN. Give the accountants, physicists, chemists, biologists, and so on the tools where they can verify the results.
The labs probably foresee things looking more like OpenAI solving Navier-Stokes: a company identifies a strategic opportunity locked behind an extremely difficult problem, throws a huge amount of money and compute at it, solves the problem faster than their competition, snatches the opportunity.
It seems like there are a few small categories that use a lot of tokens, and a large category that doesn't need many at all.
It won't be about tokens in the long run, and I think the labs know that. It will be about long-term memory and user relationships, and who gets to own that data.
The disruption and upheaval has barely even started, IMHO.
I don’t think the question is so much of AI is a future market, it is has the capital been allocated appropriately. And no one knows. The challenge Anthropic has is they have a narrow lead on the use case that makes a lot of money today, but new models come out every month and one tiny slip or wrong decision and they are toast. There is also a reliance on the idea that two things continue to happen:
1. Revenue growth increases
2. There are more highly profitable uses cases like code.
Neither are guaranteed.
I also think the narrative that these companies are profitable on compute is erroneous. You can’t ignore the other costs and no they don’t magically go away. Anthropic can’t stop training new models now or in the foreseeable future. Gemini 4 Carbon is about to come out and is apparently competitive. Google has other revenue sources. They are in the ball game. They can wait for Anthropic to run out of money.
The amount of profit they make on compute is also not particularly attractive (at least right now). It ‘s only slightly more than 2x for Anthropic, far off from many big cloud services which sell compute at something closer to 20x.
Is there evidence they are unambiguously making profit off inference when all the supply costs of that inference are considered? It does not seem likely.
Even if Anthropic and OpenAI go to zero, the bubble popping still won't hurt many people like it did in other recessions.
Actually there are two different things going on here
1. model layer companies: OpenAI, Anthropic
2. hyperscalers, data centres, hardware layer
When people say the bubble will pop, they usually mean the second one.
> I also think the narrative that these companies are profitable on compute is erroneous.
I don't think you read the Anthropic filing. It is substantially profitable if you exclude the one type of debt which was actually equity. It so happened that Amazon lent some money to Anthropic but it was tied to the valuation of Anthropic, which became a "problem" because Anthropic's valuation grew 10x. You wouldn't call this a loss in any other circumstance.
> I don't think you read the Anthropic filing. It is substantially profitable if you exclude the one type of debt which was actually equity.
Anthropic filling shows negative EBITDA, how is that profitable? Revenue < (computing + personal + selling expenses). That is not a business, that is a money burning furnace.
That before even taking into account the huge amount of money they promised to datacenters (i.e. "build that for us, and we would rent it for 5 years", non cancellable).
Sorry I meant in Q2 it is profitable excluding loan converted deductibles. There’s close to no chance that q3 and h2 is not profitable as well, unless you disagree and want to make that bet?
It heavily depends on the definition of "profitable". I bet you that "net earnings after everything distributable to shareholders" (i.e. EPS) would be negative for the foreseable future.
(units in Kilo-millions)
Revenue in first 2026 9-months 33.2
Compute and infra expenses 18.8
Personel 4.8
Sales 5.5
Administrative expenses 5.3
Operating loss 1.2 (they are still losing money before ITDA)
Did you read the amended filings with Q3? Because they are not profitable in Q3 either. And their debt due is about to increase substantially. This is all before they inevitably have to issue billions of dollars in bonds with insane yields. 450 billion in commits due over the next 5 years or so.
I think there's just some core arguments here that you can totally reasonably disagree with
>In theory, an AI company has to be worried about their users switching over to a competitor or using an open source model instead. In practice, this isn't a major concern. For starters, the average person...
There's two things to disagree with here. First, this isn't theory, I work at a company and use AI heavily, switching model is as simple as changing a line in a json file. And would I do it? I've done it like 5 times this year. Claude was good, then openai was for a while, then there were times gemini was great for quick trivial tasks, now I'm mainly back on opus-5-5 with astra for some more complex tasks. Secondly, I'm not the average person, no, I'm the person openai and anthropic actually make money from. Sure, a consumer might be stickier, but so far they're not really making money from consumers, they're mainly doing B2B.
The core thing that bugs me is not so much that AI is going to be massively unprofitable, I think there's definitely going to be profitable companies. I just think if you're Google, the best case scenario is you're transitioning from fantastic economies of scale in a monopoly business to a business with costs roughly proportional to revenue and no moat. That is really bad for Google and it's not unreasonable to think it's motivation enough for them to try and undercut Anthropic and OpenAI until they're bankrupt.
It means that the companies have to be highly cost-competitive, they can't win market share and then raise margins. If (as is the case today) astra is bad value for money compared to opus-5-5, then I'm switching, and if (as is the case today) for lots of stuff gemini can do the low hanging fruit, I'm switching. And all of this is happening before my company has really put any boundaries on our spend, in the future I'm going to be on a budget and that's really going to impact my decisions.
But according to Anthropic they are getting the healthy margin right now. If they raise margin it would be “even better”, not just “finally making any money”
But it's not clear that's sustainable. If you look at what happened in the first half of the year, they were very clearly ahead and that translated into both massive growth and great margins. But that looks like it's atleast slowed or maybe even reversed in Q3. And again, that's while there's a shortage of compute and so everyone who has compute can charge a premium for it. Shoot forward 1-2 years as this massive glut of compute hits the market, can Anthropic sustain their margins or do they have to compete with Meta and Alphabet dumping almost free inference to claw back costs on data centre investment.
> Most (or maybe even all) programmers [...] are worse at their respective field than an AI swarm equipped with a good harness
Is this the consensus view now? I'm not any kind of denier and I do recognise that motivated thinking may cause me to overestimate my own value.
But trying to be brutally neutral I'm still not sure this is true. AI doesn't make certain classes of mistakes that humans do, granted. But I'm still surprised daily even in 2026Q4 at things it does seem to make a mess of. Usually, changes that are locally good but subtly globally poor.
I thought the world of measuring programmer productivity by lines of code was dead in the 90s but clearly we see now that people forgot all of that when evaluating AI output.
> These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
What is the most expensive thing in an AI datacenter? The GPU! And what it's its lifespan? 3 - 5 years!
Bubble or not, there is no way AI spending is going to plummet. It will decrease, obviously, but thinking that after 5 trillions of investment everything is profit is so naive.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet
Unless nobody is working on miniaturisation, then this reasoning might hold true.
As seen throughout history, any sufficiently absurd tech starts largescale (refrigerators, computer, radios) and eventually become portable.
Amortising data Center cost will take a long time. Within that time, if someone figures out how to make ai run on cheaper hardware, these data center cost will Become harder to recover.
> AI companies can also cut costs by no longer allowing people to use their products for free and temporarily suspending expensive AI research
> Amortising data Center cost will take a long time. Within that time, if someone figures out how to make ai run on cheaper hardware, these data center cost will Become harder to recover.
Or datacenters are the old mainframes, and when they stop building them and squeezing the market, GPU prices go back to what they were before the frenzy or even lower, so anyone really interested can buy a decent GPU for a fair amount of money, and run models locally for peanuts. Meanwhile, dataframe owners go bankrupt.
Per the article, _somebody_ is overvalued. On the one hand, at some point they will stop needing new hardware, and can drastically reduce costs; in this scenario, NVIDIA will hit a wall because they no longer need to buy hardware. On the other hand, NVIDIA can keep selling; in that scenario, Anthroic &c is doomed to continue purchasing hardware, so their costs can't come down.
>For starters, the average person isn't going to use a locally hosted model, ever.
Let's talk about that in 2-3 years. There is a very clear incentive to have models running locally, from data privacy to cost effectiveness. The software will pop up to easily power this, and if it is cheaper than non-open frontier models it will be used. This can be done today with a bit of hassle, and will only get easier. The consumer hardware vendors are clearly heading this way.
I'm heavily biased here. We are building a ChatGPT/Cowork like solution for data privacy concious enterprise customers.
This is already starting today, not in 2-3 years. Currently customers are using our solution with cloud-hosted LLMs, and for the meantime costs are acceptable, but we are also building capabilities to do organizational rollouts of local LLMs, so that the majority of tasks can be offloaded to the local machines of the employees (which are most of the time underutilized anyways). For a lot of typical office worker tasks, this works great.
This also has the nice side-effect of allowing C-level to use AI locally for highly sensitive documents, without a risk of leakage.
I would expect that the cost-concious customers will push the broader market of harness providers to allow for solutions like this going forward. I am however not sure how much this will take hold on the consumer market.
If you don't talk about competition from DeepSeek you haven't really thought it through all the way. You might argue it's like Windows vs Linux, but operating systems have huge natural motes. Any person can just switch to a provider of an open weight model instantly. Most people don't know about it but we may very well see people learn about this broadly in the next few years. Changing models is not super complicated. Anthropic is unlikely to make $2 to the dollar when competitors are taking only pennies of profit.
The assumption here is that having "the best" model is not worth that much of a premium over a "good enough" model at a lower price. This remains to be seen, though, and people's behaviour paying for Claude Code and Codex suggests at least for now people are willing to pay.
It also assumes Anthropic can maintain the lead. If anything the competition both in the USA and in China seems to be getting closer and closer to parity with Claude.
I use Claude a lot at work and Deepseek for personal stuff, and they feel very comparable to me. I haven't found a task using Deepseek that makes me wish I had Claude - especially given how expensive Claude is.
This is important. I think people would pay a large premium for a marginal increase in intelligence. Primarily because intelligence is correlated with exponential utility which is often ignored.
For example, I'd much rather pay for the highest intelligence model when designing something or developing something because the problems it can preempt will save me more money in the long run.
The whole issue is that intelligence (as measured in AA benchmarks for example) is not linear utility
>Any person can just switch to a provider of an open weight model instantly. Most people don't know about it but we may very well see people learn about this broadly in the next few years.
Sorry what? There are 10x if not more Linux installation vs Windows installation in the world. Linux powers everything, including billions of smartphones, where Windows is nowhere to be seen. Same for routers. And actual proper servers.
If it's like Windows vs Linux it means open-weight models are going to be incredibly more successful and touch just about everything. All the while being cheaper and working not a bit but much better.
The conclusion I've come to is that AI is still revolutionary, and the buildout is probably necessary to some degree, but it is way overestimating the demand and pricing power providers will have.
SW can use a ton of AI. Lots of other applications can't, and they don't need the super smart frontier versions either. We just don't need all that much compute to serve AI for these applications, and people won't need to pay the margins necessary to get access to that AI. SW alone is not going to be enough to pay for all the investment.
I don't buy brand loyalty either. If I can pay 1$/m for the same thing from someone as I pay OpenAI or Anthropic 10$/m for, I'll do that, which almost zero barrier to entry.
The buildout is paying stratospheric prices on the assumption they can charge that much on the other end, but if demand tapers off (or model efficiency doubles again), prices will start falling. Everyone talks about Jevon's, but I don't really see why I'd spend more on AI than I need to do the things I want. I already run out of stuff to ask it to do on a basic subscription, I hardly need more.
The crux is this: I can spend x for y results. If I spend more, I probably don't get 10x for 10y results - I probably get, say, 2y results. So on that equilibrium I just don't spend all that much more because the reward for all that extra spend is not worth it. And if models get better this tradeoff starts to look worse, since I get more of the super-effective benefits for cheaper. Now I'm spending 0.1x for y results, and spending even x for 2y might not make sense, let alone 10x for 3y.
The problem is, that its not just AI for a niche thing, its AI together with compute.
People are already using AI to build them some things with CAD. Electronics too. But now what does AI need to be very good in it without an expert? Simulation and Optimizers. This again needs more compute.
And the agentic layer will be added to all big products out there. Everything you will do will consume AI tokens
To be fair, Jevons suggests that as prices fall, novel use-cases are unlocked. The things you or others currently do with AI might not need frontier models or loads of tokens, but if prices fall or capabilities increase, new applications might arise to exploit that.
Other than the fact that we’re due for one, this one is differently shaped in that there isn’t really widespread “irrational exuberance” and constant reporting about AI fortunes. I think your general investor is quite skeptical about AI and worries about exposure in their portfolio. Sure there is plenty of irrationality happening, but there is also a lot of real growth and changes in assumptions. Like the job market is way stronger than expected for even AI-threatened sectors (not entirely and I’m sorry if this affects your circles) but like we’re really just getting started here. I don’t think AI in today’s form will be the same as this time next year and that real growth is just beginning. All of the destruction of value that PE has wrought will only serve to motivate a new generation of entrepreneurs to build fundamentally better, more nimble and customer-oriented institutions.
McDonald’s is considered to have one of the biggest moats of any business. They have an extremely valuable brand, excellent locations, optimized supply chains, and so on.
Their moat is only for people who feel like eating a McDonald’s. That’s what the article is saying - when analysed they don’t actually have much of a moat because there are so many other options and competitors including cooking at home. Yet they still survive and look as if they have a moat. A moat for McDonald’s would be if their customers ate there every day or even every meal, and few people actually do that.
I can pop into any McDonald across the world and be 99.9% sure that the meal will be clean and taste like what I except McDonald to taste.
I don’t love fast food but I’ve been to a lot of them, across a wide enough range of country , to say that every other fast food chain struggles hard with that.
Again that’s the point - the moat exists if you feel like a McDonald’s, because you know what you’re getting and they are pretty good at delivering on that. But if you’re hungry now and feel like a burger you’re probably not likely to walk 30 minutes across a city to go to a McDonald’s if you’re already standing outside a burger or pizza place that looks alright.
You'd be surprised. Kids want McD's, not some random burger restaurant. They want the real thing. This value proposition exists with any restaurant with franchises though. It isn't specific to McD's. According to my wife (who fancies burgers), Burger King is way better anyway. I wouldn't know. For me, it is a scam. Since what they sell as cheddar isn't cheddar (it has nothing to do with that) but processed cheese.
If it is my son's birthday, he gets to choose. He'd pick either 'pizza from a box' (with which he means New York Pizza for kids) or McDonald's. If I'd pick up a meal from the local snackbar, from a Burger King, or bake my own fries and burger it wouldn't cut it for him. Would it cut it for me? For sure! Cause I like my fries tasting like potato without much salt, not the garbage they dare to serve at a McD's or Burger King. But not for him. It isn't about me; it is about the kid, my son in this case.
Furthermore, if we were on the go, the prolific yellow 'M' can be seen from afar, after which kids yell: 'I see McD's! Daddy can we eat at McD's today?' They want the real deal, and its standardized throughout the country (which has adapted McD's to local cuisine; example: https://news.ycombinator.com/item?id=50423291). So in short, the kids already saw the 'M' sign that day, they know it is around in the vicinity, and you won't get away going to some alternative however good, proper, healthy, or tasty it is: it isn't McD's!
I'm Dutch and lived near the German border. The first time eating at a German McD's was about 25 years ago, and it was an experience. Because they changed the mayonnaise to German customs. In Germany, mayonnaise is sour, not sweet and the specific McD's version (which normally is yellow, and contains, I don't know, pickles?) was rather different.
As such, your 99,9% is overrated since there's even lots of local flavours and customs, but yes it is generally good enough, and if you consider intra culture, you're flat out right.
Napkin math: if the 5 trillion investment is paid back by 1 billion customers, every customer pays $5000. If they want to write off the investment in 5 years, that's $1000 per year.
Even with the current AI capabilities, with the right products that seems doable.
I ain't paying 1000 USD per year (83,33 per month) for this garbage. I pay about 18 EUR per month for a Mistral subscription, and if it cost 3-4 the cost, I wouldn't pay for it.
The beauty is the open models: the margins are going to be thin on proprietary models. The only thing they got going for it, is creating artificial scarcity. But the Chinese are able to buy them from secondary markets anyway, or via Jensen Huang's girlfriend.
Even if the US proprietary models are 6 months ahead, there's going to be a time when the competition is good enough. I believe we are already well past that point.
> When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations, which makes the burst obvious in hindsight. With public AI companies, we don't see that happening. Alphabet, Meta, and NVIDIA are all AI companies that have seen their market capitalization explode over the past few years, but all of them have stable P/E ratios in their 20s. If the increase in their valuations were irrational (or even just in expectation of future growth as opposed to current earnings), you would expect the P/E ratios to go up. Instead, their valuations look exactly what we would expect from a healthy trillion dollar company. When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds. I'm not going to pretend that I'm an expert on financial bubbles, but I've never heard of a bubble where the companies involved had such a modest valuation based on their earnings, even if their earnings (and therefore their evaluations) are enormous.
My worry isn't so mcuh the large players ability to raise funding, it's the long long tail of "AI" in everything else and especially small startup companies that looks more worrying to me.
At the moment you have young grads getting $20-30 million with basically limited evaluation of their skills, all over the place. The area I am interested in is physics and engineering, and the number of Physics AI startups for where there there are not physicists or mathematicians with at least PhDs in numerical computing as founders for example looks to me like a total mess. And I keep seeing things where very clearly solvers haven't converged then getting used as training data and publicised in press releases from companies like this. I assume that most technical areas are seeing similar. This is where the dotcom bubble comparisons are most apt.
The only modification I'd make is to amortize the investments over 5-6 years instead of 20 because most of the cost is GPUs not the buildings. This increases the required revenue by 4x.
Well, if you buy a bunch of GPUs and don't have buildings to put them in (a very real situation right now), then you're just .... holding hardware and not making a return off it and *the building (and related power infrastructure)* becomes much more important than how many GPUs you have.
Compute hardware is ~50% of the cost of a datacenter. If we can assume a 5 year lifespan for the GPUs, that puts the ammortization timeframe closer to ~10 years.
The trouble that that idea is that there comes a point where the generated AI needs machines to test the code that is generated... and it actually puts a limit on the growth of AI.
I enjoyed that article, thanks for sharing. It is nice to hear that people are still having a "rage against the machine" in these heady times. For me, I think a lot of us at high market cap companies think AI is going to do xyz, but most folks are employed way outside of that boundary.
When I look at the current capacity of AI, it really feels like there's so much potential for frontier models to be used that just aren't being applied yet. Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well.
This is a strange conclusion. AFAIK in mathematics it doesn't work yet like article says: there is a problem, point agent swarm on it, and mark it done. The math results from OpenAI and Anthropic were achieved by seeding the surface of known problems and randomly trying to grow solutions; while they got some good solutions indeed, it doesn't mean current LLMs will work for arbitrary math problems.
Correct me if I am wrong on this, I am only a student of mathematics and by no means a researcher.
For the programming part, I think author already says that automated solutions require a good harness, and this is indeed correct, but in my experience building good harness is hard and agents can not do this currently without human closely guiding them. And in pre-LLM programming practice we did not need harness at all, or needed only small part of it: so, solving programming problems with LLM involves more work than without LLM, even if big part of this work can be automated.
The rest of the article seems to be a speculation mostly based on the premise that LLM can really replace human in the loop, which seems to be not the case currently.
Your post repeatedly confuses evidence that the AI industry is commercially viable with evidence that the AI industry's current valuations and investment levels are economically sustainable.
I’m not convinced that something has to be overvalued for it to be a bubble. A lot of the dotcom stocks that went bust probably could’ve taken some piece of the pie, even if small, and made a business out of it if the money hadn’t run dry. Or if they had just come along later. Turns out the internet pie is enormous and it’s still growing. And now people are much more willing to use it for payments. Even some previously bad ideas might be workable now. But investors don’t have infinite patience, and the market can spook them.
Also, great products don’t always make great money. Some companies just never figure out how to monetize and eventually go belly up. Users may not be willing to pay what you expect, and the amount they pay may not be able to pay back your debt. Not every LLM chat app is going to survive the next few years. That’s just how it is when there’s sudden growth. Our economy tends towards boom and bust cycles.
I mostly use Chinese models now. Last 30 days the average token price has been 98% lower than OpenAI. This from my dashboard today. So I now use 10x the tokens on the Chinese models. More people will discover this. I can’t understand how OpenAI can recapture what they have spent.
You mention that training can be suspended, which is true, but don't acknowledge that puts them behind in the arms race against both other frontier companies as well as second line cheaper providers.
Without a moat, training is an ongoing cost that they can't afford to not pay
Training is not infrastructure. It’s not capex, not “investment”. It’s really an ordinary running cost of what they do, for the foreseeable future, and it’s absolutely enormous. They couldn't even dial back training in a winner-takes-it-all scenario, because training of smaller, domain specific models would remain a way to compete.
Yes, and the training data behind a lot of models is already stale and rotting fast.
Maintenance is just as important and the cost is often overlooked. The sober reality is that an LLM is just a natural language search index. For now, Google might be the only one with the capability to maintain "training" at the pace most people expect.
>The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
First mistake. No. Data centers are depreciating. And they're one or two hardware changes away from having most of the hardware in them worthless (because even if you're contend to keep running those, your competitors would move to the next tech).
>For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used. This partially offsets the risk of losing customers to a different product.
Never helped browser vendors maintain their base. Not even social networks, where the graph serves as an incentive to stay. And there's also generational churn. Facebook is a boomer wasteland now, for example.
I like the sentiment, but the author doesn't appear to be accounting for some harsh truths:
- Anthropic is already expected to have junk bond status; this is going to make dealing with their existing debt rough while costs likely continue to rise
- Datacenters are 'cheap' once they come online, but that assumes they will. Power infrastructure hasn't been able to keep up, and several pending datacenters (I've seen estimates in the 50% range) meant for AI growth have been significantly delayed because of it.
- The US has a lot of debt coming up in 2028. The AI investments have been far more lucrative than the traditional 30-year bond. The AI bubble doesn't have to pop at all for the global market to plunge.
> Anthropic is already expected to have junk bond status; this is going to make dealing with their existing debt rough while costs likely continue to rise
Exactly the opposite. This is good for a tech company IPO which opens up all kinds of bond options. Most dont even have a bond rating! BB+ is a get.
> - Anthropic is already expected to have junk bond status; this is going to make dealing with their existing debt rough while costs likely continue to rise
If he is not accounting for this, then he is not saying anything of value whatsoever. So many people are still operating on vibes instead of these companies' actual financial positions, and "because Claude does cool stuff for me!".
> When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations, which makes the burst obvious in hindsight.
I think this highlights a big issue with both sides of the "AI is/isn't a bubble" debate - the idea that there's an established pattern that economic bubbles cohere to. There really isn't, and economists haven't been able to find a consistent pattern that past bubbles have conformed too (that isn't also the case of non-bubbles). "Bubbles for Farma" is a cool paper that tries to do this[0].
Like everyone else, I have my own opinions on whether there is/isn't a current AI bubble. But we should all stay humble and try and avoid stating anything as if it's an absolute fact.
Just because a technology is going to be successful in the future doesn't mean that the current beliefs are accurate of which company is going to capture the most profits from that technology.
The internet is successful and cisco is having a blast but its valuation has still been enormously wrong during the dotcom bubble. Same goes for pets.com, etc.
So in other words, maybe the valuation of some companies is totally wrong. Maybe Anthropic is not worth two trillion, or even a hundred billion, but Nvidia is worth 20 trillion. Maybe it's the other way around. Maybe the legal framework changes and the value will be captured by the major rights holders, reddit, twitter, bertelsmann, etc.
I think the biggest problem is people are conflating current AI being useful with current AI companies being financially viable. Dot com bubble had its winners and losers but the biggest thing is that many of the current AI companies are valued at prices the winners were decades after the dot com bubble. My prediction the big 3 infrastructure AMZ, MSFT, GOOG will lead openai and anthropic to a point where they owe them trillions in payments and use it to take them over.
Some indications in the text suggest that their insights are somewhat limited.
> In other words, once you ignore the cost of building infrastructure, AI is very profitable.
Once a provider stops such investments, they stop being "frontier" which makes them unattractive immediately. These two things are tightly connected.
> If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year. For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019.
Revenue vs profit
> Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness
Anybody doing serious software engineering will wholeheartedly disagree.
> People can theoretically switch from model to model, but people tend to settle in with a product and develop brand loyalty. For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used
LLMs are just one piece of the tool chain, a mostly interchangeable one. Which one a concrete product uses is not of concern to the end user. This quote very much misses the role of LLMs and their providers in the big tech picture.
> In 2023, a memo from Google called "We Have No Moat, And Neither Does OpenAI" was leaked, which outlined how the open source community was keeping pace with AI labs
Labs releasing the weights to their models is not what is typically called the "open source community".
I could go on but it's just not a very well-argued opinion piece. Unclear how it made it to the HN front page.
I'm actually convinced that an AI Bubble doesn't matter at all.
The quality of models like Opus 5.5 are already here, smaller labs can mimick these with a little bit of a delay.
Alone the current status has massive impact.
The status quo will get just faster and cheaper.
Then we have labs like Deepmind, companies like Google and then countries fighting over it (US America vs. China). At this point its not just private VC investing in it, its also countries aka military budgets or Research Budgets.
In parallel universities are doing their research anyway and if an AI bubble would burst tomorrow, Open Source and community driven would finally have a little bit of time to coordinate.
There was just last week an article about open post-training data. I think we want this but as long as frontier is as fast as it is, its not fun to play catch up with a moving target.
Yes, people dream of going back to the way things used to be, and irreversible steps have been taken regarding artificial intelligence. But while a certain segment of society is getting richer economically, humanity is heading toward irreversible sociological destruction.
To use an analogy, just like the scenario in post-apocalyptic movies where people eat each other, we are heading toward a grim end as a species.
I am increasingly open to the possibility that AI is undervalued. There have been recent applications of GPT 6.1 to the client business that genuinely threaten to fully eliminate their development team and put the business in the captain's chair again. I'm inches away from convincing them they have what it takes to prompt the robot to manage their codebase 100%. AI adoption is entirely a people problem now. It's a much better developer than me in most ways the business cares about.
Everyone keeps pricing this as a technology thing for some reason. This is more like a new kind of thermodynamic cycle than a new kind of software. The downstream applications are inevitable. We can't continue milling grain by using live animals walking in circles if we wish to remain competitive.
I think the bubble narrative is a necessity for some technology people. It now comes off as a fantasy that this will all blow over soon and we might go back to some semblance of normal. I'd argue that we've burned the proverbial ships at this point. There is no going back.
Think about how far we've come in such a short time. It's barely been one year of competent coding models. Some parts of tech/media seem only impressed by instant and globally successful results for some reason. Something this disruptive takes time. It took two fucking decades before I had a broadband internet connection. There are massive crow eating buffets that are still attended by those who missed the eventual consequences of the internet.
I agree that AI as a technology is probably undervalued right now. But that doesn’t mean the (current) AI companies are undervalued. I think it may well turn out that AI is a bit like electricity; transformative for the economy, but not spectacularly profitable.
> As one final note, I don't think the rhetoric about the AI bubble is really about finances. Conversation about the AI bubble is too widespread for it to be a contrarian position, and too gleeful to be driven by investors' fear. More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be.
I call this “SchadenFraud”: reveling in and playing up every little negative thing that happens in the AI world to make it look like it’s all just hype or completely irrational and will come crashing down any moment now.
There are enough people aggrieved by GenAI -- traditional media, creators, artists, coders, Mathematicians, and the list is fast growing fast -- that this kind of narrative will have an audience for quite a while.
They're not looking for solutions, they're just looking for rageviews. Worse, it gives folks the hope that things have not irrevocably changed and will go back to normal after the bubble pops.
Me, I’d rather get a stark but realistic picture of what’s happening now so I can prepare better for the future.
so many variables here. Do anthropic and openai as "AI" companies last longer term? If LLMs curve off in terms of what they can do, and local hardware becomes capable of running local models, seems like they won't last that long. But if they keep improving such that massive amounts of compute will always be needed beyond what can be done locally? Then sure. But then does it just become a commodity market where all the margins get stripped right down? Do they invent AI tech well beyond what LLMs are capable of at the moment? I would imagine, no matter which way it goes, hardware companies will win.
Well, first of all, these are not one off purchases. GPUs do break and newer generations come out rendering the old GPUs outdated. I don't know if the 2.08 figure is very valid. This is based on costs of leasing GPUs from companies that may not be making money on the lease deal. What I do know is that Anthropic weren't profitable in Q3 and I'm hearing rumblings they are losing even more money in Q4.
Also Popeyes never had over $800 billion in uncancellable compute agreements. And I'm glad you're having fun with GPT, but I've moved all my projects to use DeepSeek 4.1 Flash. There are plenty of different providers. And as I use OpenRouter the change is one line.
The reason I think the PE ratios are lower than previous bubbles are because everyone already knows its a bubble. Also all these tech companies already have massive businesses and AI is not going to change that.
>There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
[would -> should]
There is no logical contradiction even, just another race to the bottom, drying out of markets. The good ol' disruption playbook, like Amazon shrank or destroyed the markets to capture them, only now they are targeting all white collar jobs.
The conclusion starts:
>I didn't write this article as some big takedown of the AI bubble theory. Even with this evidence, there are serious risks to the financial sustainability of the AI industry.
But the presented "evidence" is largely only: I don't know, nobody can tell the future. I mean there are also good reasons to be skeptic of skeptics like Ed Zitron, but Idk "modest valuation" seems like a big stretch. Also, I'd challenge the notion, that you'd only ever have to build that data centers _once_. Especially with the AI pace the chips are said to be replaced every three years, while the building itself could last 25 yrs. So even, if the run at the frontier cools down, I don't think running your data centers and models would come for free.
> Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well.
We can START discussing this when someone defines "worse" and we all agree on it. For SW engineering, we've been trying to define "good" for decades and we still don't know shit.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
No, that's a total straw man. For example, I am afraid of how AI will change the world, but for that it doesn't have to stop producing slop. I'm afraid of how the slop is CURRENTLY changing the world, people and societies. Frankly, I don't even care whether someone ever pays for the slop. This isn't a "I'm going to lose my job" kind of fear, it's a "we're running education into the ground and rapidly transforming into a low-trust society" kind of fear.
“AI can't replace lawyers and doctors, but they can automate their paperwork and free up time for them to do other things”
So my question is - in software engineering (where this has bitten hardest): does anyone know _anyone_ who is less busy and has time to “do other things”? Nope, nor me.
I do many more "other" things that I previously wouldn't even dream about doing in the same time. Work that would have taken weeks now takes minutes or hours.
So yes I am doing many more "other" things. But those other things are still building features for my product.
Yeh - same. But I read “other” as “leisure things”. It’s not like we’re all having a nice time elsewhere, we’ve just compounded the amount of stuff we can deliver so we deliver even more. It’s the polar opposite of the dream where the tech does the nasty shit so we can all go chill.
The bubble is not about whether AI is profitable or whether profits can grow.
The question is how much?
If they grow 5x, then these companies are very profitable, changed entire industries,... etc.
Yet, the bubble still pops because investors assumed 10x at least.
I liked his point about bubble theories being more about antipathy towards the tech firms rather than rational thought. I definitely experienced this myself, where acquaintances are like "the bubble will burst any time now" and when talking to them, it's very clear that it's an anti-AI stance that motivates them to say it rather than any sober economic analysis.
>Anthropic's S-1 filing says that they make $2.08 for every $1 of compute. In other words, once you ignore the cost of building infrastructure, AI is very profitable.
If you ignore the cost of building cars, they are also very profitable.
>Profits Can Grow, By A Lot
They can also decline by a lot, and Anthropic's S-1 suggest that is what is happening, as their losses are widening, not narrowing.
>Google Has No Moat, But Neither Does McDonald's
That's why no one is sinking several trillion dollars into a new fast food chain.
>Alphabet, Meta, and NVIDIA are all AI companies that have seen their market capitalization explode over the past few years, but all of them have stable P/E ratios in their 20s.
The "E" in their PE ratios is circular revenue that is ultimately derived from debt and equity, not real end-user demand from profitable companies. For some of these bubble companies, if you subtract the debt-fueled private stimulus money, their actual PE ratios skyrocket.
sounds pretty good. if I made 2.08 for every dollar of my spend I would make more money. Plenty of businesses operate on much much thinner margins, and thats not utility-like at all.
I could only find these numbers in articles on something called crypto briefing dot com, not sure how the author got this number from the S-1.
When using traditional business metrics, such as operational net income, we'll see that Anthropic has been in the red for the entirety of its history, at times to the tune of $3B per quarter, except in Q2 2026.
When looking at amended S-1 (with Q3 numbers), page 23, "Consolidated Statements of Operational Data", we can see that for 2026 YTD (Q1 - Q3) net loss was $103B; operating loss was $1,260B.
Anthropic also disclosed up to $895B (multiyear) commitments for infrastructure build-up.
Basically, the company's financial results do not allow for any reasonable valuation (which would be more than 0), leave alone $2T or even $1T. If this is not an indication of a bubble I don't know what is.
If your hyper growth startup is not in the red for its entire history, it fucked up. One can argue their "profitable" Q2 was a mistake, startups at this stage should be far from profitable
While this is true, most of Anthropic's expenses involve locking up compute for the future. But they have no idea how much compute they will actually need in the future, or how much they're going to be able to charge their customers for it. Per token prices are dropping like a rock, and performance optimizations are making model serving more efficient. Combine that with the comparatively slower growth in organic AI demand (compared to debt-driven/investor subsidized demand), and you have the very real possibility of a 2000-style "dark fiber" problem.
That’s 52% gross margin. Not much higher than the overall S&P 500. If you break it down by sector, Real Estate, Health are, Financials (and of course Tech overall) have higher gross margins at the sector level.
>That $1 of compute is not taking into account R&D spending, infrastructure, etc
People assume this, but it is wrong. Check the S-1 for Q2, the $2.08 in revenue is for every $1.00 spent on "compute and infrastructure expenses" which also include R&D compute. (Chart is on page 137)
"Compute and infrastructure expenses primarily consist of costs attributable to operating our technology infrastructure, which supports the full range of workloads for our models, including research, model development, serving customers, and internal usage of Claude." page 146 in the S-1/A-1
It is not really clear. In the graph on page 137 does the "dollar of compute" includes only compute required for inference to provide the answer? Or does this include training as well?
It would be logical to assume that they show expense of compute required to run inference. Leaving training compute (which is huge expense) out.
It's damning that out of the only two companies in the world that absolutely must dogfood Anthropic's products, one of them, the investment bank behind Anthropic's IPO (Goldman Sachs), didn't find enough of a use case for to justify its continued use by most of the bank. Now LLM use is heavily restricted to specific departments and even in those departments spend must be approved. They've also repeatedly dropped the IPO target by almost half in just 4 weeks. That's WeWork levels of adjustment right there.
It's almost as if they know something about the finances that the rest of us don't...
You assume these companies telegraph their motives, they could just as easily be feeding this narrative so they can buy more of the company at a lower price.
What a weird thing to say: a business is profitable if we ignore a key part of its cost base. That’s true of any business, you know. Even if they make a 100% coverage of every request served, it’s still a far cry from the usual internet business model where responding to a request is almost free and most of the expense is the initial investment in software and its upkeep. This is why Alphabet used to be flush with cash and we get free services.
The trouble with Anthropic and OpenAI is, that not only do they need to get operationally profitable, they also need to be sufficiently profitable to provide a return on capital on all the money they have burned to date. Much of which has gone (and keeps going) not to assets but to expenses like the electricity bill. Thus, once the models are efficient enough to be served profitably, a relatively small amount of assets needs to be generating income for a relatively large amount of capital. It is my understanding, that currently the efficiency of the models does not grow even fast enough to outpace the need for new capital.
One more thing, that gets conveniently ignored when talking about AI cost is the cost of training the models. From one hand, there’s but a few institutions capable of building them and, from the other, it’s like the models grow on trees. Yes, you could run your open-weight model on your own hardware and pretend it’s profitable because you don’t have a comprehensive costing model and the infrastructure costs from security to electricity to risk are mixed in with your general cost base. But someone has spent astronomical amounts of money training that model you run and would probably like to get it back with some interest. Let’s hope that return is just money and not, say, the ability of a certain country to control the global software development process.
"If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year", 20? more like 5 years befroe they degrade and be useless
So is Google! They buy real estate on your phone to make themselves the default browser. They pay other browsers and phone manufacturers for integration.
The only reason why Google search is doing well is because it's a real estate business like McDonald's. And lately their search quality has been similarly crappy.
I think one of the things that people analysing the finances of the AI sector often miss is that the capital expenses associated with data centre buildout involve spending a lot of money on compute hardware which depreciates and becomes obsolete very fast.
Large companies have often changed the depreciation schedule for server hardware to 6 years rather than 3. I think it remains to be seen if this is justified as it’s not as though the pace of change in hardware has slowed down by 50% arbitrarily.
I think it’s pretty obvious to me that a data centre full of ampere GPUs is worth less than a data centre full of Blackwell GPUs. I suspect this is also why AI companies prefer to talk in terms of gigawatts of compute, as if efficiency per watt does not exist.
Data centre costs are not one-off. Some are, for buying land and building buildings and infrastructure. The actual compute however is a constant investment. I’m not familiar enough with the market to know what proportion of expenses are ongoing but I bet it’s higher than people think.
> Large companies have often changed the depreciation schedule for server hardware to 6 years rather than 3. I think it remains to be seen if this is justified as it’s not as though the pace of change in hardware has slowed down by 50% arbitrarily.
Anecdotally, I bought a new desktop over 7 years ago and it still feels blazingly fast with plenty of room to grow. And the cost of a new machine is prohibitive, so a replacement at this time doesn’t feel very compelling.
His analysis fails to consider trend lines in inference costs. A couple of years ago inference required a supercomputing cluster. Today, you can run quantized frontier tier models on video cards with a $1600 MSRP. At current trends we'll be running frontier models on a typical gaming computer setup in a couple of years. That is going to destroy centralization possibilities.
It also fails to consider the impact of 'good enough.' A lot of the build-out right now is in anticipation of trying to push the frontier level. But there comes a time, a time we're already arguably past, where models being used for a specific task will already be able to complete that task efficiently. Going way beyond the present should enable LLMs to solve ever more esoteric problems, but in economic terms if a company's goal is to harvest corn, and your corn harvester works well enough, they're not going to care whether or not that corn harvester can also solve Fugazi's 7th Great Conjecture in exomathematics.
Hey, I run local inference on a few RTX 6000s and will say it's pure delusion to think you get "quantized frontier tier inference" on a modern $1600 video card.
I get that you want this to be true. It is not true.
It is not frontier tier, it is not anywhere close, and if you think this you are out of touch with the frontier
Here [1] are some various benchmarks for a RTX 4090 on models like Qwen 3.8, where it usually hits around 43 tokens/s. You can find similar vids on YouTube showing it live for comparison, 'proof' if you doubt, etc. And yes, it is frontier tier.
I don't think these companies would be calling for regulation if there wasn't a bubble. I think they know that these models cannot keep the same progress forever, and potentially not even a lot longer. Instead of admitting this to the shareholders, they instead want to go the government and say "REGULATE ME DADDY", and blame them for not being able to keep all the promises that they're implying.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses
Companies are investing in scaling models. Doesn't that mean that they'll need to keep investing to stay on the edge? After all, when they lose the edge they lose customers.
> For starters, the average person isn't going to use a locally hosted model, ever
If open weights weren't a problem for these companies, ram altman wouldn't have accused the people of stealing his compute.
When are we getting an AI_OS? I just tell the prompt what to do and the entire OS is run by an agent (likely Linux with sudo permissions). Everything including browsing and copying files and what not is an agent driven task.
This sounds like a nightmare but so is using windows and office everyday and everyone is already using it.
This is a remarkably weak article to be getting as much attention as it is.
It would be easy and boring to shred the whole thing systematically, so here's one example:
> When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations [...] Alphabet, Meta, and NVIDIA are all AI companies that [aren't that]
Alphabet, Meta, and NVIDIA are not AI companies, and [I don't really need to explain this, right?] ignoring the inflated elephant here is a hell of a thing.
Honestly, criticising this is beneath me (sorry) and I don't really want to be here, doing that. Oh, but the irony. This article ("I'm not convinced the AI bubble exists") -- this copium -- being a busy front page submission today is such a neat way to demonstrate the bubble's existence. Truly, a thing of beauty.
Just replace AI by block chain or Internet and you'll see that none of these arguments hold water to claim there's no bubble.
> I don't think the rhetoric about the AI bubble is really about finances
But that's purely what bubbles are about though. People see something become popular, want to be in, over invest, and the thing becomes overvalued.
Internet was a bubble. I don't think anyone would claim the value people saw in it doesn't exist, and there's way more investment in internet today than there was when the dotcom bubble burst. It was a bubble nonetheless.
I am convinced for a long time. Just the way the companies CEOs talk and move suggest that they are looking for infinite investment without delivering AGI or singularity and we are always "almost there", though it does not arrive.
AGI and singularity are the Nuclear Fusion of energy.
Like others have said, it is not that AI/LLMs//ML is not perfectly valid technology, it is a auestion of whether the current investment into it is justifiable relative to the demand and profit. We still buy and sell tulips y’know.
> Outside of generative AI, self-driving cars are becoming a reality, with some evidence to suggest that they're even better than human drivers.
Haven’t they have been just around the corner for 10 years now? And the problem is not technology as such, but the fact you are inventing trains that use roadnetwork hardly large enough for cars?
> AI adoption takes time, but there's no reason to think that AI usage isn't going to massively skyrocket, even if AI doesn't get any better than it is today (and it certainly will get better). In other words, there's massive unmet demand for AI companies to profit from.
[Emphasis mine]. This might a good example of a person writing and thinking with an LLM and letting the model enchant you with its sycophantic bullshit. We might speculate endlessly about AI or any technology, but none of that is a basis on which you should throw money at, especially not as a retail investor. We can see value production of AI in software development, but the productivity increase promised is either not in the data, or it is too early to say yet. So the real valuation of AI is really difficult say still, a lot of things needs to happen, before it can be seen to produce concrete value increase. People need to stop really really fucking hating it for one.
It kinda pointless to talk whether there is a bubble or when it will pop if there is. It might happen today at market open or never. Timing it is a fools errand. Also it is kinda pointless to shout into the void how much you love/hate AI. Better use of time would be to develop the different systems and applications that will produce value for somebody. It all sounds too similar to crypto to me still…
"The AI Industry Doesn't Look Like Any Bubble We've Seen Before" - It looks exactly like the two railroad bubbles. Potential profit is there, in the future. But you have to invest in the infrastructure first. If you can't make a profit and pay back the loans fast enough, the infrastructure goes bye-bye, along with all the funds used to finance it.
"There's Still Plenty Of Capital To Raise" - IPO is not how you raise a trillion dollars. A million, sure. A billion, maybe? But a trillion puts retail investors - mom and pop - at risk. IPO is not how you raise funds for a railroad.
"Once AI companies become profitable" - there's the big red flag. This isn't Uber or Lyft that can float along on tens of millions in corporate hand-outs every quarter. They need 100x that much money, and nobody has it to hand out. So you can't wait around for profits. There's no magic cushion to float on.
"For starters, the average person isn't going to use a locally hosted model, ever" - They will in 4 years. "normal users don't want to have to deal with the hassle or expense of setting up something" - What hassle? You download a program with a fancy GUI and run it. There are a half dozen of these projects already, that provide ChatGPT on your computer from one binary. All you need is a GPU.
"Profits Can Grow, By A Lot" - Only if they remain cheaper than local hardware or open-weight inference providers. Which won't happen because they owe more money than they can make.
The sign of the bubble is all the investment made not based on stable returns, but on hype. The weird geniuses from San Francisco said the future will depend on this one technology, all work will be changed forever, and only their company can deliver it. If you don't get in now, you'll miss out on a fortune! That hype just doesn't work out. And the proof is here: the superhuman machine sucks at writing, the documents it creates are lies, and we seem to be spending more on it than we get out of it. Because it's not magic. It's just a machine.
> It looks exactly like the two railroad bubbles. Potential profit is there, in the future. But you have to invest in the infrastructure first. If you can't make a profit and pay back the loans fast enough, the infrastructure goes bye-bye, along with all the funds used to finance it.
That's just not true. Profit is there now. If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!
Totally unlike the railroad bubbles which built out years and years ahead with no customers and then collapsed.
> "There's Still Plenty Of Capital To Raise" - IPO is not how you raise a trillion dollars. A million, sure. A billion, maybe? But a trillion puts retail investors - mom and pop - at risk. IPO is not how you raise funds for a railroad.
You're totally confused about what an IPO is. You're confusing the valuation of Anthropic with what they will raise. They intend to offer about $100B in stock, their valuation is at 1T.
> "Once AI companies become profitable" - there's the big red flag. This isn't Uber or Lyft that can float along on tens of millions in corporate hand-outs every quarter. They need 100x that much money, and nobody has it to hand out. So you can't wait around for profits. There's no magic cushion to float on.
Most AI companies are profitable aside from expenditures to build new infrastructure to reap more profits. That's a very healthy sign.
> "For starters, the average person isn't going to use a locally hosted model, ever" - They will in 4 years. "normal users don't want to have to deal with the hassle or expense of setting up something" - What hassle? You download a program with a fancy GUI and run it. There are a half dozen of these projects already, that provide ChatGPT on your computer from one binary. All you need is a GPU.
Uhuh. A GPU. In your phone? In your laptop? The average laptop in the US is sold for about $700, desktop for about $1k. A good GPU to run a reasonable model is $10k+. It makes zero financial sense for most people to run their own inference. The issue isn't, and never was, the software, it's the hardware. And one thing we've seen from the past years of growth is that people don't look at the models from 3 years ago and say "Oh well, that's good enough, at least I can run it slowly on my $5k+ desktop". Until that happens this is a fantasy.
> If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!
They owe hundreds of billions in loans on datacenters that aren't built yet and don't currently have capacity for the new users and additional tokens they promise are coming. If they don't build capacity the increase of demand would cause more outages and higher prices which would crater both their use and valuation. They have to grow or die.
> They intend to offer about $100B in stock, their valuation is at 1T
That's one company, the aggregate debt of the US AI industry approaches 1T.
> Most AI companies are profitable aside from expenditures to build new infrastructure
Which they have to complete to make the profit they promised which they need to pay off the loans they already have.
> A good GPU to run a reasonable model is $10k+
Today. Due to supply and demand. Funny thing, supply and demand. It's very reliable. High demand makes supply short for a while, but then causes supply to go up. When supply goes up, eventually demand goes down, which brings prices down. The chip factories and economies of scale that will be complete in a few years will send prices crashing, as they have before. Which might bankrupt the remaining chip companies.
"Reasonable model" isn't a thing. Many different models of different size have capacity to do many tasks. I have a laptop from 3 years ago that runs local AI models. My phone from 5 years ago runs local AI models. And any strong CPU can run models (no GPU needed). Speech to text, image/video generation, image classification, research, analysis, computer use, troubleshooting, etc, all works on tiny models. I think you don't understand the tech here. Remember that many people spend a grand every year just for a newer iPhone/Samsung. They'll spend the same for one with better AI and more RAM.
> They owe hundreds of billions in loans on datacenters that aren't built yet
Covered by their current revenue plus 10-20B.
> to make the profit they promised which they need to pay off the loans they already have.
No, they don't need to make profit for this. Their loans and infrastructure expenses are mostly covered by a small increase in revenue
> The chip factories and economies of scale that will be complete in a few years will send prices crashing, as they have before.
All signs point to demand increasing to match. Economies of scale let you better capitalize on existing demand, so the companies stand to profit even as prices decrease.
> They have to grow or die.
This is the story of every startup. Even large companies are subject to this in the stock market.
There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for. You can't really have it both ways here.
This argument only works because the author uncharitably and incorrectly added “nobody would ever pay money for” to the end of it. If you remove that sabotaging clause, you can “have it both ways”: Ai is gonna change the world, it’s gonna be slop, and you’ll have to pay for slop while grumbling that things used to better.
It's more accurate to say that there's an OpenAI-Anthropic-X bubble. Financially sustainable investment in LLMs is useful and not going anywhere, as China shows. Dumping trillions of dollars into companies that make models that are 1.2x better than Chinese models at 100~1000x the cost, ehhh... good luck with that.
Im in China, and scared. Everyone is broke, losing their jobs while on the other hand there are these giant datacenter campuses being build in the warmest and most remote places. if the bubble bursts (there is 1000% a bubble, its the same like mass maker and innovation, sharing bikes and now ai policies)… it will be bigger than any financial crisis before. The only shit I see coming out of ai hackathons and competitions here is ai tools for marketing (ai). The rest is ai slop microdramas.
I think this is correct except that it ignores the hidden bubble in the SPVs: a lot of the firms who look only partially at risk have simply moved debt off their balance sheets- but to whom? Pension funds, sovereign wealth funds etc.
And “OpenAI bubble” alone has one massive implication: the likely demise of Oracle, a blue chip business that is deeply embedded into the western world.
On the one hand, if the air is let out of the bubble before those SPVs get themselves too deep into building things that won’t be needed, that is good. But nevertheless the failures of the SPVs will harm people who are not in the exclusive billionaire club.
Also there is a considerable money-go-round; the money for everything will dry up fast when the cloud LLM businesses can’t convince people to give them money to buy Nvidia GPUs anymore l
This analysis marginally better than Rob Ford smoking crack, I'm not convinced you should run for office, but if you succeed, please get us more Shorsey.
The first reason I became suspicious of the AI bubble is it is all the same people hyping it up. The grift just switched to the new thing.
A part of the bubble that people don’t talk about are all the small AI products and integrations that don’t do much of anything. But they took VC money to build an AI project management program and now they have to cram it in there even though it is useless and actively costs them more money. There is going to be a lot of pull back at some point with these pointless integrations and I think that is going to be felt on total AI usage.
Sure, the largest IPO in history was slashed by fifty percent before it came onto the market because no investor wants to underwrite its original offering, but it’s not a bubble
How much do you think these same people will be complaining about AI companies in 5 years when they realize they missed it? AI is the greatest wealth building opportunity of my generation and people are just dooming themselves away from it.
"If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now."
-------------
If everyone, or even the majority, believed that there is a bubble that's going to pop with near certainty, that would tank stock prices immediately. Why would anyone hold a stock they know is going to go down precipitously? If the majority of people know it's a bubble, the bubble immediately pops. AI stocks still have worth, so most people believe there is no bubble or that prices will go up before going down.
People talk about bubbles. They worry about bubbles. For the most part, though, they go on investing as though there isn't going to be a bubble. Every bubble looked like a solid investment to most people right up until it burst, or it wouldn't have been a bubble.
So, no, most people do not think AI is a bubble. That doesn't necessarily it's not though.
> If everyone, or even the majority, believed that there is a bubble that's going to pop with near certainty, that would tank stock prices immediately.
That seems to assume:
1. Those people are the invest-class and have control over what gets kept/pulled.
2. Those people don't think it'll pop later and that there's Great Fool Money to be had for just one more day at the tables.
dump n pump. day trading. long term strategies. build up for the next big thing or delusion or hysteria or lie.
if you read mainstream narratives like "AI bubble", translate it to several millions of people making lots of bucks via a lot of fraud mixed into some good, genuine, valuable work of productive people.
investors and wall street people doing anything financial is not tied to bottom up signals but to insider knowledge and top down supply chain LEGAL manipulation by people not directly involved in the actual work done.
or it's a wrap up ( made money, time to shift or prep something that will hypergrow again, maybe already re-pre-hyped ).
or it's reactions to having maxed out some part of the supply chain which had negative consequences to some other market which reduced overall gains enough to realize "oops", if we don't re-act now, future gains in other markets will be reduced almost catastrophically.
bubble talks always feels like a mix of hyped people really getting fucked over by fraud runners
the details are not too relevant for the moment as I can't have play money to invest but the narratives are interesting because they directly translate into what I read in the suggestions and advice, subliminal or direct, from people to other people.
This person lives in a bubble. IPO of Antropic is literally a sign there is no more capital in private markets and it’s up to the public to take the shit bucket.
Let’s not even talk about the second headline that costs are constrained, while one company spends few dozens off billions per quarter.
The thing is, the bubble exists because even though there is no guarantee of any big profit coming from the big techs, there are billions of dollars funneling into these companies. If you look at their current numbers, they are drowning in losses. Yet they are forcing to keep this bubble alive
The discussion of the “AI bubble” often conflates two different things: (1) is AI going to justify the valuations we are seeing? and (2) are the specific AI companies we have going to justify their own valuations.
The rapid progress of multiple closed frontier models, along with the rapid progress of open models, suggests that LLMs will be like electric motors, not like social media. Are electric motors worth a huge amount of money economically? Sure. But none of the companies who invented electric motors are especially valuable today, because the technology became a commodity.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
Because hardware doesn’t need updates.
Newer hardware is more capable which gives competitors with newer hardware an advantage.
> Outside of these expenses, AI doesn't cost that much to run
> If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now.
You complain that people take the bubble as a given. Yet you continue by doing what a lot of people on the other side do:
> Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well.
This isn't established. Like, at all. There's serious doubts as to whether AI-driven codebases remain maintainable, understandable and trustworthy over time. You look at a snapshot and conclude, while riding the hype wave.
The technology itself is here to stay i believe, but until they figure out a way to make it more efficient and not just burning money, a part of it does feel like a bubble
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
One thing to keep in mind when you're "debating" things online is you're actually debating with a tag team of people, with a whole spectrum of views. And at any point people with stronger arguments (starting from different precepts) will reply while people with weaker arguments will fall silent.
That's why "the average hacker news poster" is simultaneously rich and poor, powerful and powerless.
> More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be. People are afraid of what's going to happen with AI, and rightfully so.
> So, people wishcast. They tell themselves that this is all a giant bubble, and one day soon, everything will crash down and immediately go back to how it used to be.
+1 to this. I think this is why Ed Zitron gets so many readers - people want someone to tell them things will go back to “normal” someday.
I see it as the opposite. I think the this idea, the "seeing the LLMs as a bubble representing the hope that things can go back to normal" is itself projection from people who know their future is going to suck thanks to "AI" but really wish differently. Whether the LLM companies crash and burn, or succeed wildly, the outcome is going to be negative for the average American (and probably for the average human being).
It will eventually blow up of course but it wont be anything really expectacular, a bunch of larpers will be talking about giant loses for CEOs and surely, hopefully anthropic, all while it doesnt have a big effect on the people that know little about all of this.
But still, nothing will be back to how it was, ever
I am pretty sure the biggest thing in question is if OpenAI and Anthropic are as valuable as they say they are vs any other AI labs. I guess we will see, but I still feel it doesn’t take many ex-employees from them to build something that can wrestle them, and also it’s weird seeing other players that also have comparable AI models (Alphabet, Meta, whatever company owns Grok these days, …) do not get as much in value. There is also the “Can local models disrupt this?” question (and the Chinese labs), and there is also the question of hardware availability (will some player change this?).
I feel the projected value of both OpenAI and Anthropic requires that nothing changes in a space of technology where things seems to change a lot and reasonably suddenly.
Does Open AI or Anthropic fall behind? I think if either falls behind people are going to panic and the eventual stock price will tank. Both companies have a ton of commits so much in fact that one of them failing might take one or two other companies with them. That is what I think the big risk is.
No, no one believes that. Everyone can see that AI is going to change how many industries work completely. The parent is constructing a strawman argument "the reasoning is motivated and that motivation is nostalgia (or similar)".
Yes, quite a few people believe AI is a big scam and it’s not actually good for anything.
I know this might seem crazy to software developers who handed all their dev work over to Claude 6 months ago. But most people are not software developers.
I think a lot of software developers are convinced it must be transforming other jobs and it really isn’t and may never to the same extent. LLMs are just particularly good for technical writing and code.
But that is the potential issue. It is valued as it will and that just isn’t proven.
I have also seen this from many SWE and hobbyist programmers as well, actually, as not all workplaces have AI mandates.
Those people usually bury their head in the sand and have such strong opinions about the capability of these models as they haven't ever used anything but the free tier chatbots, if any at all.
I mean if one of the big labs collapse, then yes it mostly will? Like sure, the tech will stick around in some way or another as long as it makes money or provides a good user experience, but the forcing a garbage chatbot into every product surface, the constant flavor of the week hype, the endless doom from every CEO will all cease to exist.
No, but it is incredible for somebody to use the word “wishcasting” and then describe a group of people that would need to materialize in order for their point to make sense
I think a lot of people who post on HN don't want to admit that, apart from chatbots, AI is completely invisible to the user. A lot of users already hate chatbots to begin with and avoid using them.
So, it has already mostly disappeared, and the lack of practical applications means it will continue to bleed money. Software devs aren't using AI either for the bulk of smaller tasks that require teamwork with non-technical stakeholders.
Yes, it totally can disappear from society. The remaining use cases go with local LLMs and specialized software built around it. From the user's perspective, that's the way things always were and we'll be back to "the way it was" when the bubble is done bursting.
-1 to this. I think this is why Ed Zitron gets under the skin of AI evangelicals. They want everyone on this surrealist funny bus on a journey to whatever-the-fuck. Fingers crossed! Ignore the weird AF obfuscated financials. Ignore the crippled hardware market for everyone except data centre overlords. Nothing to see here!
I think a data center bubble exists and may have just been forcibly popped. But that's never really been an AI bubble. It's a real estate bubble wearing a shirt that says AI.
I stopped reading after the 10th sentence or so demonstrating AI-pilled psychosis and hand-waving. That it is obviously this or obviously that and obviously some other thing.
I look forward to these types of posts fading in frequency much more so than AI use itself.
Meta’s margins are apparently getting crushed as a result of AI investments.
The lack of a clearly profitable operating model and huge (trillions) in commitments for buildout is what concerns people.
I am not convinced by this person who is not convinced there is a bubble. The bubble, by the way, specifically refers to the valuations. Is Anthropic worth 1.2 trillion when it’s not clear they will have positive cash flow in the foreseeable future?
> Is Anthropic worth 1.2 trillion when it’s not clear they will have positive cash flow in the foreseeable future?
Maybe think of it as a bimodal thing. I can easily imagine a future in which Anthropic builds a sustainable profit model, grows into new sectors, and is credibly worth 7T in ten years.
I can also easily imagine a future in which AI capabilities continue to increase but the amount that enterprises are willing to spend on Anthropic’s services to do, and in which that, plus ever-increasing competition, cause Anthropic’s revenue to grow slowly, stall, or even contract. And that Anthropic would be unable to raise sufficient capital to make up for it and would therefore forced to default on their obligations. (Or that they would be able to raise enough capital but on poor enough terms that existing investors take substantial losses as a result.)
> Maybe think of it as a bimodal thing.
And how does that justify the valuation that they are planning to IPO at?
It's just a total gamble, is that the claim?
> I can easily imagine a future in which Anthropic builds a sustainable profit model, grows into new sectors, and is credibly worth 7T in ten years.
I can easily imagine a future in which Anthropic fails to find a market fit that results in the kinds of profit that will cover their current spending commitments, and where the company is another casualty of progress.
“A risk” is not the same as a gamble. Of course there is risk. But companies don’t have trillion dollar valuations when there isn’t a clear path to profitability.
Every one of the few companies with that kind of valuation generates tens or hundreds of billions per year in free cash flow.
A fair critique. Even I, someone who started as a (righteous) angry doomer over it, have come around for the most part.
My viewpoint also shifted around the shape of the bubble. The technology isn't going away, obviously, but the valuations of these companies like OpenAI and Anthropic feels divorced from reality. Inference continues to go down, most folks don't need frontier models for common tasks (including coding), and local models provide a competitive edge in several domains the frontier labs cannot (namely value and privacy). Combined with demand still largely subsidized and the companies being deeply unprofitable, and I just don't see the investments in Anthropic or OpenAI or the sort actually paying off. That is the bubble I'm concerned about, not the tech itself.
...now when the debts get written off and all this compute is just sitting around idle, available for bottom-dollar prices to try and eke out any sort of ROI for whichever creditor(s) inherit the assets? That's going to be pretty cool times.
>In other words, once you ignore the cost of building infrastructure, AI is very profitable.
Its not clear that the infrastructure will ever be paid back. Its also not clear that once models are trained and distilled that the infrastructure will be useful. Ignoring 1.9 trillion dollars of debt is probably not the greatest financial analysis.
>However, there's no reason for why this level of spending has to be maintained in the first place.
Exactly. So Anthropic and OpenAI might be holding the bags for latecomers who distill their models and get ahead. Sort of the Commodore64 and Atari of the AI space.
>Once all the data centres are built, AI spending is going to plummet.
Well data centers need maintenance, and GPUs and similar do have a very short lifespan. But even if spending plummets repayments on existing debt needs to be accounted for.
>In a pinch, AI companies can also cut costs by no longer allowing people to use their products for free and temporarily suspending expensive AI research.
Research keeps them in the headlines, see above about the latecomers.
>For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019. Adjusted for inflation, that's an even more impressive 213 billion dollars.
Yep, so all Open AI or Anthropic need to do is get acquired by a much more profitable business to pay back their debt.
>at some point, the rate of increase has to level off, but that doesn't have to happen anytime soon.
The rate of increase of revenue in the AI sector also doesnt need to be apportioned to OpenAI or Anthropic. If it wasnt for their generous loss leading subscriptions or their rnd, theres not much there to hold on to people.
>it's hard for me to see AI consumption not being 10x what it is now.
Right AI consumption will increase very likely. But will it increase at the cost of an already subsidised subscription, or increase the use of cheap Deepseek 4.1 tokens.
>For starters, the average person isn't going to use a locally hosted model, ever.
Isnt this expressly where Apple Intelligence is heading? Building small low cost models that run on Apple hardware? If I was nVidia I would have both plans being run down at the same time, looking towards selling GPUs to hyperscalers on one side, and working with OEMs to deploy AI coprocessors in laptops on the other side.
>For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used.
N = 1. OpenAI is already what halfway to their trust thermocline anyway?
>A potential way for a bubble to pop is for the companies involved to run out of capital to spend while they grow, causing them to fail.
Isnt Anthropic already down to BB+?
> NVIDIA are all AI companies that have seen their market capitalization explode over the past few years
NVIDIA is a hardware company that can has so much demand on their products they will likely survive 3 AI market collapses.
>Alphabet, Meta, and NVIDIA
Alphabet, Meta and NVidia all have other sources of revenue. They are companies building AI, not AI companies.
>When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds.
Anthropic and OpenAI dont have "E" (as in earnings)
Anthropics Debt to Equity is 363%.
>A potential way that this could be a bubble is if AI proves to be a fad.
A potential way this could be a bubble is if the makers of consumer devices point customers at cheaper or local models as part of their onboarding experience. OpenAI and Anthropic get thoroughly zuckerberged by the people holding the primary relationship with the customer base.
Or perhaps, the funding to the top 1% of token purchasers also turns out to be loans, and that dries up too.
> The AI bubble is a speculative concept that the AI industry is extremely over-valued. If true, the value of most companies involved in AI (several of whom are worth well over a trillion dollars) will collapse, devastating the entire economy. If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now. I don't think that this is a reasonable assumption, though.
Matteo's analysis is very flawed though. For instance, he projects that the
outcome is "devastating the entire economy". Why would that have to be the case though?
The AI bubble can collapse partially. There are tons of examples of how this has happened in the past. Not everything is dot-com 2.0. What we see right now is a "consolidation" bordering on partial collapse of some of the companies' projected revenues. It makes sense to assume that there is not a complete collapse, and some companies will end up dominating more than others.
> If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now. I don't think that this is a reasonable assumption, though.
To say that it is (essentially) taken for granted isn’t a great start to a rebuke. You have to engage with the arguments that they have, and they have presented arguments. Just see born-again clanker Dan Luu respond to Zitron’s claims.[1]
I don’t see any engagement with skeptic’s arguments in this piece.
> In other words, there's massive unmet demand for AI companies to profit from.
Of course. There are hundreds of millions of white collar workers that are jealuously hoarding their unrightful wages.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for. You can't really have it both ways here.
But you can. A bubble of this magnitude will negatively impact the world, not just AI investor cretins. You can take the extreme position that AI is useless and still be worried about the fallout.
> It's completely valid to be afraid of AI taking your job, but money is going to change hands when that happens.
The conjunction here implies a contrast which I fail to see.
> As one final note, I don't think the rhetoric about the AI bubble is really about finances. Conversation about the AI bubble is too widespread for it to be a contrarian position, and too gleeful to be driven by investors' fear. More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be. People are afraid of what's going to happen with AI, [...]
The always-amazing observation that your intellectual opponents have motivated reasoning. Likewise, most AI positives have motivated reasoning, like that they can enrich themselves or that they are getting so much code out of their current subscription prices.
When I write about AI it’s clear when I argue against AI itself and when I am whining about my boo-boo thoughts about losing all of my economic prospects.
> I’ve applied to seek the Liberal Party of Canada nomination
>>> As one final note, I don't think the rhetoric about the AI bubble is really about finances. Conversation about the AI bubble is too widespread for it to be a contrarian position, and too gleeful to be driven by investors' fear. More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be. People are afraid of what's going to happen with AI, and rightfully so. Beyond fear for the future, AI in the present has become uncomfortably common, with AI slop everywhere online and increasingly even IRL. This has all happened so fast, it's easy to forget that ChatGPT came out a little less than 4 years ago. Our world has changed in such a scary and odd way, and while I have a pretty optimistic view of AI overall, I get why people would want it all to go away. So, people wishcast. They tell themselves that this is all a giant bubble, and one day soon, everything will crash down and immediately go back to how it used to be.
And on a personal note, this probably goes against all economic theory, but I think the larger the scale gets of money, the less "real" it becomes. The US national debt is $40.5 trillion. I personally don't think a fraction of that will ever really be paid. There are rich and poor, but once you become rich, that added richness is essentially useless, and the debt we see is a mixture of money missing from the poor person's wallet, or the money missing from a billionaire's portfolio. One of these is more sensually palpable than the other.
If your product is good enough, shows potential, and has meaning, it will get funding. I have never done a startup or gotten venture capital so it's pretentious for me to say, but I think the money just comes, through effort (and I think 100's of pitches for the new startup), and we can't always know from where. It's just there. My brain is too small to consider where that money comes from. We're not valuing money at a certain point, but just sheer utility, and so investors will be there to back you up. In decades-long terms, sometimes centuries-long.
The debt will keep building up, but so will be our infrastructural changes. Early adoptees will get rich to degrees, and investors who held a conservative view on major progress like AI (will still be rich but) will be slightly less rich. Maybe there will be unpaid debt, like the 500 billion incurring a $400 billion loss during the dot-com crash. But the technology will continue growing.
After all, money doesn't mean anything anyway (in the theoretical sense).
It buys us soup, clothes, and a place to pay rent to, but it is just an agreement. And with the changes in the world economy that some of us are expecting to see after the AI revolution grows even more, I really don't know what the meaning of money will become.
Large data centers never made sense. If an additional pc with a graphics card was put into every home, we wouldn't need power grid upgrades. Our power grid is designed around human distribution i.e. homes. Large data centers cherry pick where it may be possible to quickly connect power connections up. Data centers need to secure their power connections before the utility rate hikes on everyone else generate public opposition.
If large data centers cannot secure free power connections paid for by everyone else, they cannot make any of this profitable. The whole thing was a scam from the start. This is why cloud AI companies bought up so much pc and server hardware. Artificial scarcity was the only way to slow down local llm adoption and sell the false narrative that cloud AI has value.
Plucking a graphics card out of every home and centralizing the cards to a handful of large data centers that require a trillion dollars in power grid upgrades is an idea so dumb, we should all be laughing at it. They even have to use active cooling or evaporate drinking water instead of free air cooling. The consumer doesn't benefit in any way from this. Consumers would essentially be paying more just to be datamined.
LLMs are not a bubble, large data centers are a bubble. Look at how large data centers are now the first target in any war. A graphics card in every home, doesn't give someone a concentrated target. National security makes large data centers extremely bad.
Large data centers are dramatically more efficient at utilizing both the power and the hardware.
At it's core, running a top-tier LLM means linear load of trillion+ weights from RAM for every token. The cost of this operation absolutely dwarfs the cost of doing the actual math. If you are only serving a single user, you can only amortize this over a handful of token streams, times like 2-4 for speculative decoding. If you are running in a datacenter, you can batch thousands of streams, and get correspondingly much more tokens out of the same expenditure of energy and hardware.
Be slightly careful with this. Large datacenters use power to operate the facility (this is measured as PUE > 1). Datacenter-oriented hardware often contains low-diameter, high static pressure fans that consume quite impressive amounts of power that are not counted as overhead in the PUE calculation.
A home- or office-oriented workstation will have much less power-hungry fans (go compare a Noctua datasheet to a server fan datasheet - the Noctua wipes the floor with a server fan if measured by CFM/W, and the internal temperature rise is mostly controlled by airflow and not by static pressure). And in some climates and times of year the marginal facility HVAC cost associated with running a computer may be zero or even negative.
And lower utilization hardware may run in on a more efficient point on the power-vs-performance curve. [0]
That being said, your overall point is correct.
[0] For a dramatic demonstration of how much this curve matters, you can run a non-CPU-saturating, non-latency-sensitive workload on various Intel CPUs with and without turbo on. Seeing a 2x or greater difference in reported power is quite possible. (Intel is terribly at writing platform firmware and the default behavior may be wildly suboptimal.)
In addition to the batching issue, "a graphics card" is not going to cut it for the multi-trillion parameter frontier models. You're looking at a 512GB Mac Studio to support anything close to those model sizes, and that's with heavy quantization and much worse performance.
agreed, gets some mid range hardware, some click installing agents with descent customisability, and im talking boring customisability like 'help me plan what to buy this month for my home' kind of tasks. Thats what we are going towards
There's also the fact that the US and China are in an AI arms race. So there's an inherent existential benefit to being ahead in AI regardless of where the financial feasibility/profitability discussion lands.
And at least currently, keeping up in that arms race means a maximal push on growth at the expense of profitability. Realistically, this will go until we hit some bottleneck.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
Fair enough. But the real problem, it seems, is that people love slop.
My family is trying to sell a house in the South of France; the real estate people are sending us 500-words emails every so often, that are so obviously AI it hurts. Well written, and content free. Nobody cares. I'm the only one having issues with it.
At work, a new team let Claude do all their work and write all their status reports. Everybody seems to love it. It's so fast! And mostly correct! And so fast to correct when it's wrong! And here too, I'm the only one having issues with it.
A Canadian author wrote a book in French that turned out to be 100% AI ("C'était ça ou mourir"). Before that was revealed, every critic loved it, in Canada, in France and everywhere else (25 translations were started). The public loved it too: it's one of the top best sellers, and success continued even after the "scandal".
The quote is not a contradiction at all, it can both change the world and be low quality.
If that's true for every AI output is up for discussion, but it's reasonable that if someone is skeptical about the technology, they will also be skeptical about the quality.
That the tech is taking over many areas is not a stretch at all, that's just something that can be observed. Big companies (not just big tech) have been trying to get away with lowering (cost cutting) the quality of their products for a long time, AI lends itself to the same purpose, and will (and already is) obviously be used for it. Starts with customer communication as you mentioned, and customer service, and believe it or not I don't think every company should generate their own tech stack and software for their little products.
I think the problem is calling all AI output “slop”. If people genuinely like the content then is it still “slop”?
And, playing devils advocate for a moment, does it matter if a robot or a human authored it if people like it? People
loved Magic Eye pictures in the 90s and they were computer generated too.
The problem is also with the creator. You learn nothing when AI produces things for you. When it's art, to each their own, I guess.
But when I ask my realtor questions, he doesn't know anything. I get the same answers by typing into ChatGPT.
When I ask the Claude team at work about why this or that, they don't have a clue. They "will get back to me", which means they'll forward the question to Claude and Claude will output another 500-words email with a lot of things that are obvious and not part of the question, another couple of assumptions that are plainly false, and in the middle of it, sometimes, yes, an interesting insight. But the insight doesn't come from the people sending the email, they didn't produce it, they don't see it, and they never act on it.
> But when I ask my realtor questions, he doesn't know anything. I get the same answers by typing into ChatGPT.
Not going to comment on your workplace specifically, but I’ve found this to be an issue with other industries like real estate long before ChatGPT were a thing.
The problem is the people who pickup the phone, greet you at the shop door, or take the photos, they’re the lowest paid juniors and thus know nothing.
So in most industries, you were always lucky to get through to someone with knowledge on the first attempt.
That's highly unlikely because then the company wouldn't exist. At some point in a (to use our example: a real estate business) you are going to need skilled professionals.
I think the bigger risk is what happens when they retire if we are not training the next generation of professionals.
> This is exactly what we were saying with programming during early day of LLM.
Yes I know. I was there too. ;)
> But that does not stop people from doing 100% vibe-coded projects.
And why should it? Let people "code" however they want to "code". Just like some people will find creative ways to play the guitar while others will just play covers.
By the way, you haven't addressed the core point posed: if people like AI output then is it still "slop" (in response to comments where people describe all AI output as "slop")
"slop" is less so about if something is a popular product or not, it's more about the quality. You can think that AI can produce high quality outputs, but it's unreasonable to think that this applies to the majority of it's use. There has always been corner cutting, and using AI to cut corners/safe money in product design/production inevitably results already, and will more so in the future, in a ton of bad products. That does not mean that there can't be good ones too.
> "slop" is less so about if something is a popular product or not, it's more about the quality.
But we are talking about popular culture here. That was never an industry where quality was measured by artistic excellence. It's an industry that is entirely motivated by mass appeal. Thus something being popular IS the measure of success.
So if AI is slop, then so is 99% of the human-made content it's replacing.
Slop exists because the cost of creating content has dropped to zero, which made the signal to noise ratio crater as well. Concision and density used to be a virtue. Now it's the opposite: content is measured by weight.
People absolutely cared about "C'était ça ou mourir", his publisher ended up dropping him. The reason the success continued is that people denied that he used AI and had all sorts of shaky political reasons for this (e.g. claiming that Haitian writing trips up AI detectors). Even the reason it got popular in the first place is to do with the themes of the novel rather than its writing.
That's not really the point. Critics the world over LOVED the book; they only distanced themselves from it when it became clear it was AI, but when they didn't know, they couldn't stop raving about it.
most of the infrastructure will be garbage in 4 years. As before, it will become a monopoly of 1-2 big players and everyone else's datacenters will become ... crypto miners or something.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
Fatal misunderstanding.
Datacenters are not warehouses. Amazon.com being borderline-profitable owing to massive ongoing capex on warehouse expansion is not at all like AWS becoming borderline-profitable owing to massive ongoing capex on LLM datacenter build-out and SOTA model training (it's CAPEX, but carries OPEX affect because of GPU depreciation schedules and competitive hardware upgrade cycles), and opex of delivering LLM inference services, and being able to price it all competitively.
(edit: expand explanation of why it's a misunderstanding)
There's also a failure to misunderstand the arms race nature of it all and the need to be in the lead since everyone is assuming it's a winner take all market. That's probably close to correct.
A data center is a warehouse for sure but the building is only a few percent of the cost. The GPUs are probably 80% of the total cost or more. And they probably only have a 3-5 year primary lifecycle.
Pay the building off over 20 years? Sure but 80% of your capex problems remain. 2 trillion already spent? Just those GPUs need to generate 500 billion a year over the next 5 years.
> everyone is assuming it's a winner take all market.
I guess this is the big gamble, the idea that if you just spend enough money as fast as you can you can become the gods of AI forever and can just charge whatever you want for your magic beans.
But what is the moat? If in reality these things become commodities and just compete with each other to spit out cheap tokens then there's no way the margins are there.
> everyone is assuming it's a winner take all market
I don't think anyone serious believes this any more. The tech is thoroughly commoditized. There are a lot of plays to be made, like OpenAI cornering memory, or spending heavily on unique data for RL, if you can stop distilling, but there is no proprietary core IP at all.
There is no way to monopolize a technology this big. It would be like monopolizing literacy.
It's not the AI technology itself that is being monopolized, it's the data necessary to train it.
More specifically: when everything you could previously have earned money with is subsumed by AI, the "frontier" of profitable knowledge exists solely with the highest concentration of capital. Because that's where the highest amount of compute and consequently valuable data is.
Already now the idea of the tech being "commoditized" is true only if you exclusively look at the bubble of rich countries. AI-worthy compute is prohibitively expensive otherwise for individuals.
Power, natural gas, gasoline, water, mobile contracts, railways, internet access are probably better analogies. They have tendencies to (local) monopolies but also a lot of precedent and knowledge on how to regulate and break them, how to introduce competition.
they broke up Ma Bell, but they've never done anything about the internet access. Mobile is an oligopoly.
> and can just charge whatever you want for your magic beans.
'Winner take all' doesn't mean convincing others to buy your product. If others have something to buy your product with, you've not taken all, fundamentally.
'Winner take all' means having the most powerful army. The assumption is that once AI crosses a certain threshold then there won't be anything the rest of us can do to stop the machine of whomever gets there first, so they will truly be able to take all. Hence why the AI labs sometimes hint about the risk of human destruction.
The army interpretation exists, and might sadly be relevant here, but it's not the only one. Google, for instance, won a nonviolent winner-take-all competition for search.
You're right that Google has been greatly helped by having what has been considered the most powerful army (the US army) backing them. They wouldn't have had a hope in hell of surpassing the "mom and pop" stage without that.
But I am not sure you are right that it is powerful enough to secure all. I search daily, but cannot remember the last time I used Google. The other AI companies do it better. And I'm not alone in that sentiment. Google's stock price recently took a nosedive because investors have grown quite worried about the competition. Without a sufficiently powerful army stopping them, there will always be someone else ready to try to outdo you.
> since everyone is assuming it's a winner take all market. That's probably close to correct.
why? I keep seeing this over and over again but I've not seen an explanation. I would really appreciate some insight into this line of thinking, since it's not obvious to me. What is the benefit of holding hundreds of billions of computer hardware when it's not proven that it's even needed.
Lets say the total market for an arbitrary winner-take-all industry is $100 Billion/year in profits. Just so that we have a round number. And for another round number, lets assume the fair market PE ratio is 20x. This over-simplification would then have a fair market valuation of $2 Trillion total.
Just simple napkin math.
Now, lets say there are 5 companies competing to be the winner. Only one can be the winner, but all 5 are trying to be the winner. How much should each company individually spend?
Well... starting-bid is $2 Trillion, or ~$10 Trillion in the aggregate. Its basically the waste of competition. Because everyone is vying for monopoly power, the only fair for them to all invest over $2 Trillion in the hopes of being the dominant one at the end. Even if $8 Trillion is wasted, that's "other companies", your competitors, who lose out.
But wait: if one company spends $2.1 Trillion, that one company likely becomes the winner and the other 4 companies lose out. So then another company spends $2.2 Trillion, to get ahead of that war. Etc. etc.
The idea is, innovation will concentrate where the necessary ingredients are.
When AI replaces human workers and researchers, the "frontier", with which you can earn money and gain an edge over your competition, resides in those data centers with the highest concentration of computational power.
If you own those data centers, you control the frontier of human knowledge, the edge where money and power is being carved out.
A "failure to misunderstand" is a fitting turn of phrase.
People not only ignore second order effects, they also don't think about what would actually happen when a "winner" really "took all".
While AI is a tremendously useful technology, business types wildly misunderstand it and its consequences. Greed eats brain.
You got me on that one, geez! I'm not going to edit it though because I think you're right. Nobody really knows if there's going to be a winner-take-all or not. During the arms-race portion though people are definitely acting like it will be that way. But watching how fast models catch up to each other it's not obvious it will play out that way.
It's probably closer to 45-60%.
https://epoch.ai/data-insights/ai-datacenter-cost-breakdown
https://www.csis.org/analysis/impact-tariffs-ai-data-center-...
Thanks for the links!
The first one says that networking infrastructure costs 13% and the other says 1.5%. Not a great start.
Let's imagine that a traditional full rack can host 80ish H100s and that each one costs $25k. That's 2 million right there. There's probably only 10x cases in that rack so that's 10x dual core motherboards, RAM, network, etc. Call them 20k each. Maybe you need a 32 port 100gig fiber switch, that's 8k. Let's say it's 2.3 million per rack and 2 million of it is just the GPUs. We're at 85% of the rack cost right there in GPU, but the other stuff in the rack isn't durable either.
I suspect that it would take more than just another 100k per rack to finish out a data center so my "80% of the cost is in the GPUs" is definitely wrong. But I didn't see anything in these articles that indicated to me that 50% of the cost of a datacenter is anything really durable like buildings or electrical infrastructure or cooling. One rack might only take up 20ish sqft itself another 100sqft of isle. These buildings don't cost $10k/sqft to build the envelope. There's still an enormous amount of cost in stuff whose performance follows more of an exponential performance curve, not linear. It's just an absolute nightmare to try and amortize exponential performance curves with linear repayment.
These days there's usually two networks. Multi-trillion parameter models are too big to fit on one box of 8x H100's, so many of these are running RDMA (800Gb/sec on GB300 series) for rack-scale/many row-scale inference.
For GB300 series, pricing along the lines of $500k/server, 9 servers/rack ($4.5M/rack) before looking at water cooling, 800Gb RDMA network, and your 100Gb IP network
but is it a winner takes all market? currently we have models that are close-but-not-quite SOTA which are still heavily used and very profitable because they are _good enough_ (i'm talking about deepseek, of course).
or is the "winner takes all" mentality based on the assumption of being the first to achieve the technological singularity?
> or is the "winner takes all" mentality based on the assumption of being the first to achieve the technological singularity?
Alternatively, it may be based on becoming the favored provider of the defense or intelligence industries.
I read somewhere that the expected 3-5 year lifespan of inference GPUs was not holding up to scaring: in practice, the GPUs have the same lifespan as any datacenter CPU or RAM. I think that estimate was based on Bitcoin mining ASICs, which wasn't an apples-to-apples comparison.
Older GPUs still have an active secondary market but they are probably not profitable at market rates for inference because newer chips are so much more efficient.
If Semi-Analyses is correct [1], even H100s would lose money at current prices - I doubt that is actually true of H100s in practice because from what I can tell it is based on a 3-year depreciation from the new price, but the point is that there is a point at which old hardware's operating inefficiencies dominate.
https://inferencex.semianalysis.com/infx/profit-estimator/de...
I'm not certain it's the longevity of the actual hardware so much as there is a huge incentive to replace your hardware with faster, more capable equipment. Who wants to invest in the AI company that's using 10 year old GPUs when the next company is using brand new?
Old GPUs today are selling and renting for more than they were new which suggests the typical depreciation pattern isn’t holding.
There’s a risk that likely will happen that once demand is met capacity will overshoot significantly. However, when will that occur is very up in the air. Everyone is training every larger models and there’s more AI usage than ever
There’s still a cost to underestimating asset depreciation from an accounting perspective. Imagine being left holding the balance sheet when you budgeted for X amount of depreciation but the demand drops significantly more than you expect during the next fiscal year. I guess my point is that artificially constrained supply is not a free lunch.
Power generation is also extremely expensive ongoing - either through gas turbines that eat fuel or on-grid power that datacenters will need to pay to capitalize (depending on ratepayer agreements of course - but clearly people aren’t stoked about subsidizing these)
In aggregate it’s a lot of electricity, enough to impact local grid markets.
But as a fraction of inference revenue, there is good reason to estimate that the direct electricity costs are perhaps 1-5% of inference revenue.
That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
Lawrence Berkeley National Lab, November 2026: https://eta-publications.lbl.gov/sites/default/files/2026-11...
There are substantial economies of scale on both power generation (grid-scale) and inference power consumption (batching and utilization).
> That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
But that’s a feedback loop that could lead to more supply being built. Or it could lead to regulations that put enormous surcharges and low limits on data center power consumers, which effectively shield consumers directly from the lack of supply. Which in turn removes the incentive for consumers to expand power generation capacity to keep up (or to allow data centers to expand it themselves).
There is enormous opposition to a new data center project 30 miles outside my city because they want to build a 1000MW natural gas power plant and a 500MW biomass plant (that will take millions of acres of forest to sustain).
If that opposition continues, I see little chance of anything but electricity cost rising as a percent of inference.
Fun part ... all those GPUs are going to shredder after use.
The problem is likely the GPU price, NVidia has a very large margin - but it doesn't seem like the the use case supports it. If NVidia cut the GPU price by 2-4x then these problems likely go away.
I wouldn't be surprised if most hyperscalers are betting on cheaper GPUs on their next upgrade through either pricing pressure on NVidia, competition from AMD/Broadcom or in-house chips.
Normally you'd replace the device because there is more power efficient one on the market so running one at current energy prices costs more than buying new, more power efficient one and saving power.
IF prices of GPU/RAM come back to pre bubble costs, we might see the 5 year cycle of replacement that's relatively common in normal datacenters pre-ai bubble.
At current prices ? those GPUs that are there now will be running there in 10 years still.
I'm of two minds on this issue. Off the cuff I agree, but then again where are all the used "enterprise" GPUs coming from right now? There's plenty of secondary market for used Telas K, M and V series GPUs right now and the V100 is only 9ish years old, and that only if you bought it on the release day.
I do agree that 5 years isn't the hard cutoff I was suggesting but I'm not convinced that 10 years is a reasonable timeframe to consider paying the debt down when Moore's Law is involved. I know it's no longer the 12-18mo doubling it used to be but even if it's 2ish years then over 6 years you get 3 doublings or 8x performance gains. At 10 years it's another 2 doublings or 32x performance gains. It's hard for linear payback curves to deal with exponential performance curves past 2-3 doublings.
> At current prices ? those GPUs that are there now will be running there in 10 years still.
If they don't burn out after two or three. Annualized failure rates are about 9% according to Meta.
A quarter will have failed within three years. Less than half of the original hardware will be running after ten years. In both cases if you assume that rate does not increase over time. Do H100s come with replacement guarantees?
https://arxiv.org/pdf/2503.11901
This paper puts the failure rate of A100s closer to 0.3%/yr, H100s under 1%.
I guess we will see. But I would not expect GPU hardware to last that long until there is some underutilisation in data centres.
Mind you, most of the H100s that have ever been sold are still in their packaging, I suspect. Or at least not plugged in.
The gamble: gunning for market supremacy before investor cash dries up.
The real gamble: betting there is a large enough pool of c-suite that doesn't fundamentally grasp the concept of vertical integration that they're willing to pay to train their competition. Because if you think AI companies aren't training models how to do everything your company does with your data you are deeply delusional.
The real real gamble: praying sovereign governments stay anesthitized to the national security implications of a handful of extra-national organizations operating the conputer security equivalent of WMD with no oversight.
> And they probably only have a 3-5 year primary lifecycle.
I think in 3-5 years close to 100% of current GPUs/RAM will still be in service. New hardware matters only if gains in efficiency are large enough to put you out of business because the competition gets them. Extreme case of that are bitcoin miners. With AI inference there are way too many factors that influence the profitability. It's not like you can't run new model without the new hardware. It can just run on more of the old one. 3-5 years for GPU are estimates when the main application was gaming with new games with new requirements for running on a single GPU being constantly released.
Thanks for pointing that out. That has been my concern as well. Sure the post says "they make $2.08 for every $1 of compute" but that seems to only be accounting for OPEX. If we start accounting for CAPEX and variations with interest rates things look a lot less rosy. I mean a lot of folks are suspecting Anthropic will get a junk bond rating when it rolls out to public markets.
I definitely get "Community-Adjusted EBIDTA" vibes from that number.
(for those who don't remember WeWork, this was their made-up variant of EBIDTA that they tried to use in their S1 filing before crashing out, here's a random article on it https://privateequitymodels.com/blog/community-adjusted-ebit...)
You seem to be wrong about this
> A typical one-gigawatt AI data center requires $38 billion in up-front capital expenditure (CapEx) and $0.9 billion in annual operating expenses (OpEx). If CapEx is annualized over each asset’s lifespan, the total cost of ownership equates to $8.5 billion per year. Servers dominate this cost at $5 billion per year, or 60% of the total. Operating costs are small by comparison: even energy, the largest OpEx category, costs only $0.6 billion per year.
https://epoch.ai/data-insights/ai-datacenter-cost-breakdown
This is my heuristic for the capex split and effective lifetime of the asset it creates (I'm no expert of course, this is based on various reading):
- 20% Capex has ~20-year lifetime: The shell (land, building, cooling gear, rack mounts, power lines etc.).
- 80% Capex has ~3-year lifetime: The computers, including bulk commitments to power and maintain them.
And interest rates are why this split is a ticking time bomb. The sheer quantum of investment to build one is almost entirely debt funded (~90% of total outlay). I would guess the endgame would be bad even if the split for these entirely different asset classes is 50/50, which would be very generous, based on what I've understood.
Datacenter funding is extremely leveraged. For comparison, only some banks are allowed that much leverage for that size of capital, that too only if they have federal reserve backing. Even in such a case, the retail account holder (pseudo-investor, because our deposits are loans to someone else) has practically no downside protection. If the bank collapses you will lose everything.
1 Gigawatt * 1 year = 8760 GWh * 0.8 load factor * $0.2 per kWh = $1.35B of only electricity.
They're paying way under 20 cents a kWh. They're paying less per kWh than residential households. If the residential rate in an area was 20 cents, I'd guess the datacenter is paying 40-60% of that rate
They get bulk discounts because their load is 24x7, so the fixed costs covered by a per kWh charge is spread over many more kWh than a residential customer
Yes, and that is small compared to the building cost.
The main problem is that once training is complete, you only need inference which needs far less compute.
Even if we allow for the strange “it’s only a one-time cost” hypothesis, the author’s next paragraph outlines another actual bubble, just 5-10 years from now:
> Once all the data centres are built, AI spending is going to plummet.
Yeah and what happens if none of those data centers translate into more clients? Its not a one off, its a gamble too.
Please, Giancarlos. Let's all behave with the times... I think it's culture-appropriate to say "You're absolutely right. Datacenters aren't more clients. We build. What if they don't come? It's not just a one-off. It's a massive gamble."
/s
(edit: fix name)
To be clear: it's a tongue-in-cheek riff, macroexpanding Giancarlos's (I think humorous) comment. I don't mind the downvotes---acerbic / black comedy is risky business :sweat-smile:
To be fair, I could be wrong, we could all be wrong, we don't know what we don't know. On the other hand, I do think they should come up with a better plan than just trying to retrofit them everywhere they can find. I do think some states make more sense than others.
We will know once it all unfolds, on the other hand, at least there's people who are getting paid to build these.
Worse case, we have acccess to (hopefully) really well-built data centers that can be used for anything hosting related in the future.
I'm hoping the "worst case" you state becomes real, because that would be a great case! Blogged it here:
After "AI": Anticipating a post-LLM science & technology revolution
https://www.evalapply.org/posts/after-ai/index.html
I, for one, welcome the coming age of the post-LLM-datacenter-overinvestment-bust-fueled backyard GPU supercomputer revolution.
(edit: in which case, I'd say more people would get paid to build more of these, except not as obnoxious township-destroying behemoths ... the community library, as opposed to the giant book seller in the cloud)
(edit 2: And I'd go as far as to say that it is anti-employment-generation to allocate economy-scale capital to a very small number of very large scale industrial installations... You wanna incentivize a large number of small-to-medium ones instead, that employ locally in many many more places. That is how you win back industrial capacity.)
this also misunderstands the data center rollout itself. openai and anthropic etc aren't building any data centers. they're promising to buy billions of compute and 3rd party companies are taking those promises and using them to fund the construction of data centers which will then lease out that compute back to them. so it's not actually a capital construction project at all. CoreWeave or whoever is making the capital investment. OpenAI is just promising to make that a good investment.
I mean that's not entirely true
https://openai.com/index/building-ai-infrastructure-with-the...
Yeah, Open AI and Anthropic may be buying the bulk of their compute from other providers but they are now moving into building their own infrastructure directly. There is a scale beyond which it makes more sense to build vs. rent
Zero cash flow is not the same as zero earnings. They may end up with negative earnings sometime in the future depending on how much additional capex they spend, how long they take to depreciate the capex, how fast revenue grows, etc
GPU seem to last lot longer than expected.. And have good resale value for other use cases. The $2 for every $1 inference cost quote from Anthropic I presume includes all opex.
I reckon it depends on how hot they run GPUs for how long. Running any chip at clock or overclocked (likely) will burn it out faster than "business as usual" workloads.
Defraying of capex also depends on how the hyperscalers scrap used hardware.
AFAIK, all hyperscaler hardware is effectively a consumable. Disks, memory, boards, chips; all the high-value pieces do not get refurbished into new computers for sale into secondary markets. The commercial and legal liability is too high---leaking their own proprietary IP as well as leaking customer IP, not to mention secrets and hardening configs and so forth.
I can't recall where I read it, but if a datacenter is modeled as a bunch of black-box hot-swappable shipping containers, then for pre-LLM workloads, per-container unit economics works only as long as a container delivers over 30% capacity. After that you got to swap it out, and send the black box into a furnace to recover them rare earth metals for the next generation of chips.
To a first approximation, high-throughput, gpu-maxxing math workloads (LLM inference) are surely harder on chips and memory, than are bread-and-butter SaaS / data lake / ETL / RDBMS workloads. Speculating further, I'd guess that the ~40% operating margin of datacenters is due to the relative compute-friendliness of non-math workloads for lack of better words.
> Running any chip at clock or overclocked (likely) will burn it out faster than "business as usual" workloads.
Could it be that they _have_ to run them like that because of the shortage of compute volume and once there are sufficient data centers, they can be run relatively cooler and thereby extending their lifespans? That doesn't bode well for NVIDIA/AMD.
> AFAIK, all hyperscaler hardware is effectively a consumable. Disks, memory, boards, chips; all the high-value pieces do not get refurbished into new computers for sale into secondary markets. The commercial and legal liability is too high---leaking their own proprietary IP as well as leaking customer IP, not to mention secrets and hardening configs and so forth.
Maybe that could be a catalyst for going for more "commodity" approach to hardware instead of highly spcialized (atleast for some sections of the hardware) so that the IP is not the hardware but the model/software and they can re-coup the costs by selling the hardware to "tier-2" providers that might emerge based on using cheaper second-hand hardware?
> Running any chip at clock or overclocked (likely) will burn it out faster
server parts are tuned for power efficiency and longevity
B200 has an operating range of 10-35C. https://resources.nvidia.com/en-us-dgx-systems/dgx-b200-data...
RTX 5090 (and most consumer CPU/GPUs) have thermal limits at 90-100C.
I suspect GPU and CPU failure rates have more to do with stress from thermal cycling, if they can keep them loaded and running without heating/cooling cycles, they should last pretty long. Their lifetime is probably dictated more from when do they stop being competitive on a flop/watt basis than by their failure rate.
Interesting... Upon reflection, I feel like it's a "yes, and" proposition coming down to failure mode.
[Type A] Constant use a constant power level: Quantum world rules would dominate.
When run hot, electron tunneling will kill off individual transistors, and AFAIK that is a cascading failure mode once it takes root. A big enough failed spot, even if localised, could effectively kill a core (if not the whole chip---too many errors). A local hotspot will directly affect neighbours, and indirectly the chip as a whole because the same work is being pushed on fewer cores.
[Type B] Thermal cycled: Mechanical and electrical factors would dominate...
Microfractures / mechanical damage in weak areas of whole chips, and motherboards. In a hyperscaler server rack, if a board dies that part of the rack is dead---micro-maintenance is probably too expensive.
And speculating further, it stands to reason that Type B failures would likely cause Type A failures, and vice-versa.
> The $2 for every $1 inference cost quote from Anthropic I presume includes all opex
No, it does not, its just revenue minus compute. Per their S1, Anthropic has never turned an operational profit.
That's simply not true. See page 160 of the S-1 - clearly shows positive operating profit in Q2 2026.
That's because there's a temporary supply side crunch.
That only explains why the price of those GPUs at resale is high, but it doesn't invalidate the fact that those GPUs are still very useful way past a 2 to 3 years depreciation schedule advocated by Michael Burry et al.
Older GPUs can still be used to run model inference, albeit more slowly than newer GPUs. We already see frontier labs turning this insight into an opportunity to charge a premium on token delivery speed, when models run on newer hardware. As AI becomes more widely spread throughout the economy, not all workloads will require fast inference, and there will be many use cases for previous generation hardware. The only exception is if the latest batch of specialised chips become ubiquitous very quickly, but then price of compute AND intelligence per watt will both go down dramatically, which would reduce the need for sustained high capex and opex
Yes it is a supply-side crunch (chip manufacturing), but IMHO, it is not temporary. Only two things will remove the supply constraint.
- Demand drops like a stone: raw TFLOPS / per model. This will happen if someone figures out insanely fast and power-efficient math (specifically, multiplication) for existing chip architectures, that makes it dirt-cheap to keep re-computing ever-better SOTA LLMs, as well as to on-line compute inference with ever-bigger models (scaling laws rule). These are different workloads, AFAIK.
- Moore's law re-starts, because someone figures out entirely new transistor designs that can deliver at least up to 100x per-core frequency gains over Si transistors (which would be nothing considering that Si transistor tech delivered ~1Mx gains since invention) and which can be manufactured at similar bin sizes, at similar scale (i.e. equivalent per-die yield), price-competitively (no more 10x the equivalent in SOTA Silicon compute).
I'd say that a math breakthrough might be more accessible than the transistor-manufacturing breakthrough.
IDK if Moore's law restarting is possible. But I wouldn't rule out some sort of non-silicon compute breakthrough that functionally jumps us a couple orders of magnitude. The entire world economy being aggressively focused on compute seems like a good environment to drive investment into figuring out photonic computers or whatever. We do know at least one other non silicon highly energy efficient compute technology (biological brain) exists and works very well
> Moore's law re-starts, because someone figures out entirely new transistor designs that can deliver at least up to 100x per-core frequency gains over Si transistors (which would be nothing considering that Si transistor tech delivered ~1Mx gains since invention)
That assumes that there is no hard, physical limit on Moore's Law, which is obviously false. There is a limit, and as long as we depend on mass flow of charge through conductors, it's nowhere near quantum sizes.
Where that limit is has not yet been determined, but at this point the support structure to write transistors is the size of a house trailer, using a staggering amount of toxic chemicals (just to run the exotic lasers). So, advancements today may scale very nonlinearly in cost.
The very phrase "Moore's law restarts" acknowledges that we are no longer on-track.
Yep, I mean to imply practically impossible, in context of bubble-timelines. Theoretically anything is possible given enough decades. Practically, AGI has thus far been 20 years out, as has "clean" fusion power, 1 gee EM drives etc.
> And have good resale value for other use cases.
Do they? Is H200 $-competitive with an ASIC for crypto? Bearing in mind they're 4 figures.
because of GPU depreciation schedules and competitive hardware upgrade cycles), and opex of delivering LLM inference services, and being able to price it all competitively.
Hacker News increasingly feels like being a horse in 1916, reading in the Horse Bulletin that Ford is a scam and a bubble because the mass-market automobile will soon be commoditized.
There's an object-level claim to consider, but behind it the more baffling implication that this would be good for horses.
No horsing around here, just money math that says interest rates on debt sunk into datacenters will far outstrip the most optimistic revenue projections as of 2028-29-ish, when the bulk of the debt will come a-calling.
Also, inverse Pascal's Wager, because horse v/s motorcar is a false dichotomy... they are bound by the same laws of physics (faster horse, less shit, more noise and more smoke), merely different economic realities.
Hyperscaler bet is "Superintelligence" and/or "AGI", or we will make a God. That is a technological singularity. There will be a "before God" and "after God" world, and we will very likely never be able to understand what the AGI/Superintelligence's world looks like. (Because to understand what it is like to be a bat, you have to become the bat, but at that point your meaning of "understanding" itself changes entirely).
So I believe ten trillion dollars and a bunch of pirated books and source repos, and well pirated everything, mostly of very dubious quality and provenance, will not result in a God. I've been acting accordingly.
I mean Ford didn't get a trillion dollars because he is to be only manufacturer of any and all transportation. There are still horses, and Ford and Dodge(Anthropic) are not even the biggest manufacturers.
Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
Everybody knew it.
But some investors don't have the patience for these "sure things" to mature into their commercial destiny and others know they can siphon off a killing by taking advantage of them as they speculate.
These bubbles aren't about potential, they're about timing.
And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
Most of the top players involved in AI right now are going to be rich. And in 20 years, the world is going to involve a lot of AI in soooo many things. But a lot of shirts are going to be lost below those top players and a lot of companies are going to fold because infrastructure, refinement, and integration take time while the FOMO speculation starts as soon as the big discovery is recognized.
> Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
Canals... where a repeating pattern was the company underestimating difficulty/costs, and going unrecoverably into debt. Another company, which then bought the largely finished canal, cheap, they had potential profit... until tech change (railroads).
> Another company, which then bought the largely finished canal, cheap, they had potential profit... until tech change (railroads).
Absolutely right. Somebody may buy the DCs. The tech changes very very fast, the DCs may be entirely worthless. US has high energy costs, it can never be cost competitive.
Why Data Centres Will Change the World and Bankrupt Their Builders — John Hempton https://www.youtube.com/watch?v=GY5YyPLfqp0
Useful (like solar, airlines, fiber), but it won't be profitable. High capital cost, low marginal cost, subject to competition and with technical improvement --> these four factors guarantee businesses will never be profitable, and they go bust.
> US has high energy costs, it can never be cost competitive.
A non-trivial percentage of that is artificial.
When push comes to shove and enormous profits (and anything remotely resembling national security) are at stake, the US is decent at removing artificial problems.
You've already seen a ton of that related to chips and data center build outs.
Why is solar more expensive in the US than China? Mainly tariffs. Same for batteries.
Why is nuclear more expensive in the US than South Korea? The government.
Why is SMR never going to beat conventional nuclear if you remove artificial problems? Physics. Though specialized applications will be viable with SMR.
The UK has even higher energy costs and people seem to want to build data centers here? (Or at least they say they do - looks more like land speculation to me):
https://www.globalpetrolprices.com/map/electricity_industria...
The UK has got a huge amount of wind power that can be purchased ahead of time at better rates than listed in that table, it's also got some bottlenecks with power distribution (especially between Scotland and the rest of the UK) which will take years to fix so there area areas where decent power contracts can be had for customers who can build out fairly quickly (datacentre operators with large power requirements).
The UK has way less cheap wind than the US has cheap hydro. Let alone cheap solar.
"The UK has got a huge amount of wind power"
Very little wind today here in Scotland and for most of this week - which is a pleasant change I guess...
Weather is not climate. A datum is not data.
Oh look! A bird!
The reality here in the UK does not always match the ambition though.
https://www.reuters.com/business/media-telecom/microsoft-pul...
Both our high energy prices and the current capabilities of our energy distribution network relative to the renewable generation and storage capacity we are trying to move towards are significant barriers.
People want to do all kinds of idiotic things which make no sense. We've had excess compute/gpu since 2010. People didn't know what do with all that excess and started bitcoin and tens of thousands of crypto, followed by NFT. The claim was crypto will solve all problems, which hasn't happened. Now, everyone moved on to AI, AI will do this, that and everything else and also take away all your white collar and creative jobs.
Even third world countries like India which can't provide basic services to their citizens, have no water, no power, no sanitation or cleanliness are giving hundreds of billions of dollars in incentives (land, power, water, tax subsidies) to hyperscalers, to create at most 2.6 jobs. Google clearing land to build India's largest data centre: https://www.youtube.com/watch?v=dBJPf6Pxfvs
> The UK has even higher energy costs and people seem to want to build data centers here? (Or at least they say they do - looks more like land speculation to me):
There is no meaning to what politicians and crony capitalists do. It is all lies, theft and grift.
The island where residents fight data centres for water: https://www.youtube.com/watch?v=WuToaVyWmL4
Because European countries are politically stable and are under the rule of law. Things that are rather priceless and can't be bought with money.
Those drones hitting high-rise hotels in Dubai were rather eye opening.
I agree with your point about rule of law, but let’s not gloat too quickly about nasty SA getting its comeuppance. It’s going to look pretty ugly in Europe when Russia follows through on its potential to disrupt or destroy critical infrastructure. We are pretty vulnerable too.
the UK is stable? economically it's falling pretty hard and the governments seem to be a revolving door
“When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds. I'm not going to pretend that I'm an expert on financial bubbles, but I've never heard of a bubble where the companies involved had such a modest valuation based on their earnings, even if their earnings (and therefore their evaluations) are enormous.”
Do you know the situation with rail, canals, etc?
Railroads had some of the same issue encountered by the dotcom bubble, and in particular dark fiber: Overcapacity.
Once railroads where demonstrated to be (extremely) profitable, people kept building them even if the existing demand was already met. The exact same thing happened with dark fiber (looking at you Enron).
AI is a bit different in the sense that the capacity can be used for research, and more importantly is not bound to a specific connection (between locations), and therefore a lot more fungible.
Not sure about canals, but while I think there are issue with current AI investments, I would not conflate them with the railroad bubble on this basis.
AI is a bit different in that unlike the rail tracks and the dark fiber the data centers arent actually going to be finished. It's much more of a stranded asset - there will be piles of chips and 1/4 finished construction projects, not usable assets.
No doubt somebody will use the assets for something but it's still wildly more wasteful than either railroads or dark fiber.
You don’t think there were half-finished rail lines abandoned when the companies ran out of money?
Not at all. They were just a lot more salvageable.
I'm reminded of vertical hydroponic farming in old silicone fabs in Japan...
https://www.japantimes.co.jp/business/2026/07/09/companies/p...
Also, (setting aside Standard Oil’s business practices), this is a common meme in oil exploration historically, and even today. Early prospectors were honestly pretty stupid, and their principal differentiation was risk tolerance.
There a lot of parallels with AI…it doesn’t take a genius to see the long-term potential, but the risk is in implementation, so it’s a mad rush to the finish line by a wide variety of less scrupulous early players. Rockefeller got his big start by sweeping up failing producers who were all capital weak, and I’m sure something similar will happen here during the inevitable bust.
I can do the hard, risky work of drilling wells / deploying AI, or I can sell you a vision of that future, today. The most dangerous technologies in terms of grifting are the ones that everyone can understand.
Restaurants and bars!
This pattern repeats across many industries with large up-front costs.
Capital cycle theory: https://www.youtube.com/watch?v=2J2Fb1bBufA&t=577s
You should check out the fundamentals of all those businesses; quite different. AI might be closest to a railroad based on these Capex outlays, and like railroads and other utility businesses the margins are thin and paybacks long. We wouldn't have Microsofts and Googles at 3% a year, so OpenAI and Anthropic are going to need something else to support this level of investment and speculation.
> Most of the top players involved in AI right now are going to be rich.
Almost all the "top players" in the dot-com crash died or lived on in relative poverty; only Google and Amazon turned out "very rich".
A large number of companies failed in the dotcom crash, but the founders did pretty well if they'd made it to an exit event. Don't forget that it was a stock market bubble imploding. The founders and early employees had already sold shares in the IPOs that put the companies on exchanges. It was the investors who bought stock that lost money.
The after effects were that it was virtually impossible to raise capital for a while afterwards, so the next round of founders struggled. But then interest rates went to practically zero and money got cheap so it was all good.
The founders doing well is not mutually exclusive to the overall crash as experienced by the public. Also, a lot of founders are doing well even before they take on a venture. 8/10 founders from successful high growth startups self report as middle to upper-middle class background and 50%+ have elite education from top 10 schools.
You're right, the broad market drags down all retirement savings and those who are about to hit the age of 60+ pay most dearly.
Wait. Are you talking companies, or people? Companies come and go. They are replaceable virtual containers for people. I don’t think it’s appropriate to apply the term “poverty” to companies.
The Yahoo founders left with never-fly-commercial money before re-joining the billionaire club by investing the money they had. Others that survived post-IPO include eBay and Dell.
There were high profile failures but we have had those recently too eg. pet.com (dot com) and color.com (post-dot com, domain has since been repurposed)
> The Yahoo founders left with never-fly-commercial money before re-joining the billionaire club by investing the money they had.
I think you're confusing people cashing out with actually creating a successful company. Yahoo is not what people would describe as flagship company, or pets.com or Ask Jeeves or myspace or Voat or whatever. The dotcom bubble made a lot of people rich, but most examples were at expense of bag holders.
You are the one confusing what success is.
If you go by current standings in the market, yes, Yahoo is no longer at the top. Yet Yahoo was one of the most valuable and influential tech companies in the late 90s, early 2000s.
Yup, if you google what founders in those 'failures' did was that they pretty much jumped vc side or founded more companies and have had very successful career. Not mind boggling like Bezos or Brin but still very cool career.
taking the money and running is the dream of the VCs
if you believe the milton friedman approach to economics, a company has no other value in that what it can provide to the owner (shareholder)
take your millions while it burns and, as the parent said, never fly commercial again
For a VC "running" would mean re-invest in the next thing to get more money. What alternative would you prefer? Keep doing the same thing without investing in new stuff?
They don't "take the money", there are other people that believe some things (profits/growth in a field) will last forever and will bring 10x return, so they are willing to give them the money.
No successful company is on top forever. Just take a look at the top companies 20 or 30 years ago: where are they now? Nowhere near the top.
eBay, Priceline (now Booking.com) did very well too
It would be hard to make a list of things we forgot about because they failed 30 years ago
I know numerous multimillionaire dot-com techies. I guess they're not "top players".
Jim Cramer lost big but has continued spouting nonsense ever since pretty successfully
Mark Cuban?
> AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
That's an interesting claim but it needs support. How are you comparing their maturity rate? AI is arguably maturing very fast.
Another argument in favor of the AI build out, usage of the data centers is not geographically bound like the other technologies considered. A datacenter in Cheyenne WY can be used by the entire continent. They are not even bound to AI usage.
Its profitability ain’t.
>But a lot of shirts are going to be lost below those top players and a lot of companies are going to fold because infrastructure, refinement, and integration take time while the FOMO speculation starts as soon as the big discovery is recognized.
Another reason is that the way we find out what works and what not is through experimentation. Especially in a transformation like this it’s impossible to predict which configurations will work and which won’t.
> they're about timing.
Yes. And valuation. I think the leading frontier labs will be fine in the long-run and NVidia will be fine. Some of their investors who bought at too high of a valuation, expecting to flip it in too short a time frame will lose a chunk of their money when they have to sell for less.
The hyper-scalers in the middle are the most vulnerable. They borrowed billions at increasingly high rates to lock up GPU and power capacity. They are 'ground zero' of a lot of the 'circular money' deals. Big banks who financed hyper-scalers initially, expecting they could package that debt up and sell it on to diversify, are already struggling to find buyers for the debt, causing its value to drop, which then forces the banks to cover with more collateral, and that's how the snowball starts down the mountain.
Your logic is backwards.
NVIDIA will take a huge hit and we will probably end up with anti circular financing laws.
Frontier labs are losing money and are most likely to go bankrupt and be absorded by bigger companies.
Hyperscalers will take a huge beating but they are ultimately profitable and they already control the frontier labs. So after a hard crash they will be fine. A lot of their debt is actually offloaded onto pension funds, insurance companies, etc. Those risk bankruptcy.
They will be just fine. They might not see the astronomical price margins of the current bubble, but there is a lot of untapped demand that is curbed by current mem/GPU prices and if they go down the demand will rapidy grow.
Wait... with hyper-scalers, we are talking about Google, Meta, Amazon, Microsoft, are we not?
And do these not all have established business models that are ridiculously profitable, and huge cash reserves?
How the heck are THEY vulnerable, and the "bare" frontier labs like Anthropic and OpenAI with no preexisting business models or profits, but equally ridiculous borrowing are supposedly not?
I don't think they are talking about the ones you listed, but companies like CoreWeave.
Google, Meta, Amazon, MS have huge cash-producing businesses they can just fall back on. Of course the stocks will get hammered in the short term, but they are still sitting on cash volcanos.
Pure AI plays are what are most at risk. In the bad case, all the CoreWeave type companies have their assets acquired on the cheap by the existing big players. In the armageddon case, companies like Anthropic and OpenAI also end up in the discount bin.
Coreweave & co are neoclouds.
Hyperscalers are FAANG.
I think it was implicitly assumed that one or more of those would fail. The main question is, who else?
> And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
This is so demonstrably incorrect it's hard to even begin a counterargument to this point.
Agreed,
I believe AI is here to stay as general concept turned into practically working implementation.
What that practicality means is we all who consume AI are learning but we're still in the early adopters phase of all that.
Of course AI companies invest a lot and try to stay on top of the traction but I also believe not all of them will survive test of time.
I personally believe that a lot of over investment in data centers and supply of energy will be burnt and will have to be reinvested somewhere else when technology advances enough to "miniaturize" AI enough so it won't require couple of football fields to provide what it does today.
I do believe cloud based AI will be always attractive to small business if price is right and to big enterprises so they don't need to invest inhouse for the scale they need, but that will be sort of alliance based enterprise B2B market.
anything between (including me for the scale I represent business wise), I can't wait until I will be able to buy appliance from 12 to 48U maximum that can do all what me and my company needs AI wise as local AI.
moment subscriptions become equally expensive as pay as you go models, I switch to mac studios with 96GB and reasonable weighted models that will be specialized in something instead of one LLM/AI to prompt everything to it and hoping it can address all sort of problems as one solution.
AI one day must fly to space, be called Mother and discover aliens. it can't be 6 football fields as physics won't let it take off. It must be one day small enough so we all can have ChatGPT 6.1 medium model at home without subscription, just as fixed price with TCO over lifecycle of the whole appliance (software + hardware together)
that may become a drama to many companies that apparently may not be worth trillions unless they diversify enough its AI offering.
> AI one day must fly to space, be called Mother and discover aliens.
We're discussing economics, not science fiction.
50 years old sci fi is today’s economics. That was just metaphoric highlighting of the direction which always includes miniaturisation at some point.
If that line about space is the only line you picked up as meaningful, then you missed whole point made and your economics is science fiction, buddy
> Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
I don't see all these things as being comparable to AI... At the bottom, they were "100%" technologies. Trains were just able to transport commodities, and the internet was able to reliably connect computers. So their path to success was about gaining market.
Since they were 100% technologies, they disappeared. One could build solutions using them and forget they are building blocks, because they just worked.
LLM AI is different at the fundamental level, because it's non-deterministic. It needs human-in-the-loop. It won't ever disappear. Gaining market share doesn't fix anything.
Their point is (and I tend to agree), that LLM AI is not different at the fundamental level. The idea is that e.g. coding capability will be a standard feature of future computing hardware, not a very expensive service you rent. Anthropic is the current-gen mainframe operator selling time-shared access. In the future, that capability, too, will just be a commodity, i.e. a part of the silicon in your notebook or smartphone or whatever will replace those.
Yes, to a large extent, for current real world mass market use cases at least, the only thing keeping the biggest players relevant is that they (and their pals) have monopolised access to GPUs by buying them all and making deals to buy almost everything that Nvidia can produce (even more than they can plug in).
If SMEs were able to buy and host their own GPUs, they would not hesitate to avoid paying a subscription for access to AI tools. They're already starting to do this.
Models already are a commodity, the massive bump in access prices to profitable levels is not going to happen, because businesses will not pay it.
>It needs human-in-the-loop.
Not AGI Pilled. NGMI
AI is taking just as long to mature as anything in the past, and it will take 20 years?
I don't think so.
The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.
Further, there is the expectation that breakthroughs in applying AI to science, robotics, or medicine may unlock trillions in value.
That this will take 20 years, rather than happen before 2030, is just speculation. None of us can see the future.
Consider the wisdom of speculating on the incredibly fast adoption of the internet and mobile technologies, and on that basis investing heavily in whatever mobile carrier had the fastest network.
The speculation that the AI investment sector depends on is not just that adoption is incredible and that the value it generates is high, it's ALSO that the companies will be able to capture a significant amount of that value generation for themselves via fees or some other mechanism.
Are the LLM vendors more comparable to Google/Apple in this scenario, or more comparable to AT&T and Bell?
If you are comparing AI to other technologies you fundamentally just don't get it. Yes it is actually different this time...
While it's an interesting question whether Anthropic and OpenAI can capture that value, I find the telecom comparisons not so apt.
In this comparison they are more like Apple and Google: they are ahead in research and are able to offer better technology than their competitors.
The six month gap between the open-weight models and the frontier has not closed yet.
Tne value gap is closing, as corporates wind back token budgets towards relative sanity and reassess what they're getting for their money.
The bubble exists because c-suite people are sheep. They follow the herd, and the frontier CEO wolf-wannabes have been trying to herd them in a profitable direction.
But herds are easily frightened, and there's always a danger they'll stampede in a different direction.
And the risk of a catastrophic AI-enhanced online blackout is increasing.
I disagree. For my own enterprise job, Opus 5.5 with a $500/month API budget is the best option.
The cheaper open-weight ones, even if they were available with our enterprise contracts, would be a huge downgrade.
> The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.
They’re still losing gigantic amounts of money. Whether that is because they’re “flooring it” as it were and growth will eventually catch up, or because the business idea is fundamentally flawed and just temporarily alive because of investor risk appetite, is the fundamental question underpinning whether it’s a bubble or not, and nobody knows the answer to that question for sure. That’s how bubbles work.
No, they are not "losing gigantic amounts of money".
They are already profitable and are investing enormous amounts into datacenter buildout to meet demand over the next years.
The substack linked below only looks into one tiny sliver of the off-sheet circular financing that dwarfs even the most optimistic revenue growth, but it's a good place to start to understand why this is a bubble and that, yes, Anthropic is in fact "losing gigantic amounts of money":
"Boom and bust is not about revenue. To find that kind of risk, start by reading the balance sheet... [The 2008] bust came from off-balance-sheet instruments."
https://kennyfraser.substack.com/p/anthropic-ipo-part-one-wh...
Patrick Boyle goes into a deep dive that's well worth the watch:
https://youtu.be/Ic1Jf7VlXRM?si=DjiP6LVdEykBlWwU
What are you talking about? OpenAI is losing money, projected a $14B loss for 2026 and is themselves not expecting profit until 2029-30. xAI is losing money, about 3B.
Granted, maybe I extrapolated Anthropic's profitable operating result to other companies where it's not yet looking so good.
I believe your $14B OpenAI number is from 2025 though.
Where do you get the idea that Anthropic is profitable from? Is this another one of those “if we remove everything that is a cost, then revenue means profit”-scenarios?
Anyone who counts cost as part of their P&L shouldn't be allowed to make decisions! /s
> The reality is that the adoption is incredibly fast and that revenues of the AI labs are like doubling by the quarter.
How is that compared to the speed at which money is thrown in them? And how much of the profit downstream they might indeed potentially capture with high retention rate?
Bubble can’t be assessed on mere quarter profit expansion over short period (relative to large company group life expectancy).
It is looking entirely plausible to me that Anthropic will make a profit in 2027.
Don't you think that it's less than reasonable to suggest that the technology will take on the order of 20 years to mature into widespread adoption and profitability?
> It is looking entirely plausible to me that Anthropic will make a profit in 2027.
Then you should invest heavily in their IPO, I suppose.
You'll do much better than those of us who believe there is no way their financials can ever work.
I came to my conclusion a different way, I ran different models locally and measured the amout of resources required for basic things. From that I extrapolated some napkin unit economics, and I couldn't find any way to build a profitable business from it.
I don't assume I'm the smartest guy in the room though, so for all of our sake, I hope you're right.
I agree about timing, but not the speed:
> And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
They may be as expensive to mature (in terms of equivalent duration of total economic output), but by the wall-clock we've gone from "invent Transformers to make translation less bad" to "solves Millennium maths problems" in 9 years.
After 9 years of railroads… well, those were originally an Ancient Greek way to drag ships over an otherwise annoying bit of land, and the more modern attempt had wooden rails and a rope for humans and animals to pull for around 250 years so let's say after 9 years of the steam locomotive: 1784 for the prototype, 1802 for the "first full-scale working railway steam locomotive" (according to Wikipedia), so 9 years past the invention and it was still a prototype.
Semiconductors (well, transistors, i.e. us understanding the process and doing it deliberately, rather than getting lucky with cat's whiskers before we had any clue about quantum mechanics) did at least get commercialised in less than 9 years, but AFAICT in the form of "now our radios are better, sturdier, and smaller" which is a rather smaller change than what's already happened since Transformers.
(Of course, why Transformers, rather than the other attempts at AI? Transformers are when people cared this time around, but there have been previous AI bubbles. I'd argue those are like pointing at 1515 (or the Ancient Greeks) for the railway, or cat's whiskers for semiconductors).
I've been saying for awhile now that the utility is increasingly undeniable. It's amazing technology and for better or worse, is transforming the tech industry (in particular). I just wish we lived in a world without the speculative bubbles, because if this had been developed slowly, deliberately, over time, there would be no necessity to cast aside troublesome details like safety, suitability, and ethics...and we could enjoy the fruits as it was implemented gradually, responsibly, as it matures. As opposed to going "wow, this is really cool, but gosh I feel shitty about using it..."
Nor would it have been thrust blindly into the hands of your manager, the marketing department, the CEO, and countless other folks who have no understanding of how it works, its limitations, or its risks.
> I just wish we lived in a world without the speculative bubbles
Speculative bubbles, I.E. risk is where much of the capital comes from to develop these ideas to start. I’m not sure you can have one without the other.
When you say “increasingly undeniable” do you mean that it’s already plainly undeniable and it’s getting more so? Or do you mean it is still a bit deniable but is threatening to become undeniable soon?
Is the utility increasingly undeniable? What does data actually show?
A tech workers union or a stronger regulatory state could help with this but neither party has been that interested in either of those things for the last 50 years.
> A tech workers union or a stronger regulatory state could help with this
Look to Europe to see how that has worked out.
Well at least from where I'm sitting, it's working fine for us. We aren't in the midst of an oligarchy made up of the Epstein class, for one.
Could thing be better with the tech industry here? Sure, but at the same time, profits and industry leaders aren't the only things that matters. And at the very least companies here can't just exploit the public with total impunity.
I am also a European. Our oligarchy looks different for sure (and look into the history of pedophile scandals in Belgium, BTW).
I am not as optimistic as you. We seem to be stagnating our way into a stark choice between becoming an American tech vassal, a Chinese tech vassal, or losing out the remnants of our ability to compete on the world markets. Half of the continent borrows money at an unsustainable rate in order to pay millions of people to do nothing for decades.
We have made decisions that looked reasonable at the time they were made, but too much has changed in the meantime and now we are stuck in bad equilibria.
The Bismarckian pay-as-you-go pension system was stable as long as people had plenty of kids and died, on average, a few years after retiring. Now they have one kid and spend 20 years in retirement before they die. The maths of that system just does not work out anymore.
The European highly regulated industry was king as long as the rest of the world had no other choice, but now that Asia has industrialized, they are blowing us out of the water. First with lower prices, but now also with quality. German exports are half of what they used to be, with another halving expected. And of course without exports, our current social spending cannot be sustained either - you cannot keep such a massive redistribution mechanism running on an economy based on selling each other cups of coffee in a beach bar.
Our most talented engineers tend to move abroad and stay there. Brain drain is one of the most serious things that can happen to a country or a continent, and we are suffering from it.
We just cannot continue to live like that, something will have to yield, but many voters just don't want to hear that.
> We aren't in the midst of an oligarchy made up of the Epstein class, for one.
Aren't you? Or, haven't you noticed yet?
We are. In the UK we have the former Prince Andrew and Peter Mandelson selling information to Epstein. There are lots of other people all around the world connected to Epstein and mostly investigations are slow or on-existent or blocked through lack of cooperation by the US.
Epstein's network is one we know about because its been exposed and because of the sexual abuse aspect of it has both drawn attention and made people less likely to cover up for those involved. Its only a fraction of the corruption that goes on though.
> We aren't in the midst of an oligarchy made up of the Epstein class, for one.
Europe even has literally hereditary monarchies; let alone an insular upper class who are famous for their debauchery. They just have much less free speech protection so it doesn't get bought up as much because of the legal risks.
Related note, but while looking up some background info on Wikipedia it appears [0] it was technically illegal to publicly insult Donald Trump in Germany at the start of his presidency which would have added a hilarious angle to a classic joke [1].
[0] https://en.wikipedia.org/wiki/L%C3%A8se-majest%C3%A9#Germany & https://en.wikipedia.org/wiki/Freedom_of_speech_by_country#G...
[1] In the USA, you can stand in front of the White House in Washington, DC, and yell, "Down with Ronald Reagan", and you will not be punished. Equally, you can also stand in Red Square in Moscow and yell, "Down with Ronald Reagan", and you will not be punished.
The US is great for free speech as its constitutionally protected. However it is also very corrupt. Pros and cons.
There are a lot of advantages to constitutional monarchies in terms of stability. The monarchs do not actually have much power. I on balance prefer living in the UK and would not want to move to the US. There are pros and cons, and its culture and society (e.g. unarmed police) as well as politics and some if it is about personal circumstances (as someone who is obviously ethnic minority or foreign everywhere in the world I prefer a less racist society, for example).
I can yell "down with the King" or "down with the Prime Minister" pretty much anywhere. There are some restrictions on protest in a few places like Parliament square.
Hereditary super class is by miles less dangerous then tech bro CEO. The latter are so closed to actually finishing American democracy
Yes, America is somewhat freer then Russia. Despite aligning itself woth Russia, trying to help it as much as possible last years.
But no, America does not have more freedom or human rights defense then Europe. It has also much worst powerty.
Is that a joke?
You are posting this on a forum made by American venture capitalists, and not a forum made by European tech unions, for a reason.
Curious to know what you think the reason is exactly. Nuclear weapons? Economic imperialism? Ruthlessness?
Indeed I am, and you're right, there is a reason for it. The tech unions don't need to capture mindshare in order to advertise VC-funded startups ;)
We shouldn't assume that YC is hosting this forum out of benevolence. However it is true that this place is also useful for things other than just advertisement, so that's a neat side effect. We just shouldn't assume that that's the primary purpose of HN.
Can you read the non English speaking forums? Because they exist and people post on it but if you're a monolingual American you won't know it.
> I've been saying for awhile now that the utility is increasingly undeniable.
If that was true it is strange that you would find yourself saying as much for a while now, given it sounds most appropriate in response to someone denying the utility of slop generators.
I wonder if physical technological progress is also the same as information treatment progress. Everybody is happy having food and materials faster. But AI bringing you solutions faster is also taking away your job, passion.. the negative side seems larger.
>I wonder if physical technological progress is also the same as information treatment progress. Everybody is happy having food and materials faster. But AI bringing you solutions faster is also taking away your job, passion..
and, apparently, our brains https://www.media.mit.edu/publications/your-brain-on-chatgpt...
> But AI bringing you solutions faster
Is it?
Every single vibe-coded software I've ever used looks pretty on the outside but has massive, glaring holes in functionality that make it unusable.
It's like having food faster but then you find out it's actually full of silicone and microplastics.
You don’t have to “vibe code” anything to get enormous benefits from AI.
Yes, but "vibe coding" is what's driving the investment bubble.
I cannot comment but I would like to know if there's data to support that assertion.
Bubbles aren't caused by betting on the technology. They're caused by betting on the wrong companies. Say you have railway, and there's 5000 railway companies. At this point no one has any idea what good looks like, and it changes every 6 months. Investors bet on all 5000, and once all the drama is over, there are 3 companies still capable of competing in the railway business. The investors in the other 4997 get screwed, but since the investors also bet on the winning companies, theres a good chance the actual successful companies themselves get screwed because their investors are broke. There is then intense consolidation, and one of the 3 to nail the technology is still left standing while somehow there are now 3-4 other companies that won by taking talent and inventions from bankrupt companies or just by peaking after the bubble burst. About 20% of the bubble companies have also survived by finding niches.
That's a bubble. It happens every time because it has to happen every time. If you stay out of the game, the people who stay in the game win and your holdings in redundant technology are worthless.
Bubbles are caused by reckless stock purchases leading to insane, unjustifiable valuations, regardless of the company's prospects.
Hysteria + widespread financially illiteracy means there are lots of people prepared to buy AI stocks at any price. They're adamant to get in (FOMO).
Suppose OpenAI will be a successful company into the long term, and you bought its stock at a $5T valuation. After several reporting seasons and subsiding hype, its true worth turns out to be closer $1T. OpenAI could still be enormously successful, but anyone you bought at much over $1T valuation is going to get hurt.
Nah.
Every single railroad company failed. That repeated several times.
The ones that existed on the end were created by assembling the capital brought on the bankruptcy of the original ones. Most of those failed too.
That's how a bubble is. No original investor got out ahead.
The difference is that all those things needed physical infrastructure to be linearly build out.
AI models can be deployed on the same datacenter as the old one and be either orders of magnitude cheaper to run, or orders of magnitude more capable.
The only thing that can pop this bubble is if we're not smart enough.
That's not the difference. The difference is the distinction between core infra, B2B services, and B2C services.
The companies that sold core infra during dot com mostly died, because they invested in fiber etc far too quickly. Likewise with server factories like Sun.
The B2B and B2C survivors were patchy. A few did extremely well but there were a lot of casualties.
The biggest winners offered B2G - business to government - disguised as search and social media. They were actually selling individual user profiling and targeted ads, and that turned out to be the biggest money maker.
The reason Anthropic and OpenAI appear wildly overvalued is because their offering isn't AI in some abstract sense, it's the next generation of surveillance and persuasion technology. But now with corporate surveillance in addition to the traditional profiling and targeting.
You can tell a lot about what someone is doing - and the industry is doing in general - by looking at everyone's code bases. You don't need to train on them or steal keys to extract immense value.
Potentially there's also the option to eventually make claims on corporate IP developed with AI tech.
Anyone who thinks the companies who pulped and torrented whole libraries of books for their training data would never, ever, pinky promise, try to extract value from the code they're generating is being very naive.
>AI models can be deployed on the same datacenter as the old one
This hasn't historically been true. The clusters have been getting bigger along with the models over time. And now we may even want to have these computers have rapid access to extra compute or caches of the internet or other data stores.
> AI models can be deployed on the same datacenter as the old one
Not if the datacenters are in space, you can upgrade the models for a little while but once you need to upgrade the silicon you need to launch new hardware.
And data centers are practically free to build, are welcomed everywhere, cost little to run, and never need to have their infrastructure upgraded!
Google hasn't retired any of its now ten-year-old V1 TPUs or any of the subsequent generations so it's unclear that AI infrastructure "needs" upgrades.
Of course, Google could be wrong about the economics of running old AI processors. Or maybe TPUs are special.
I do wonder how much of their inference is running on 10-year old TPUs. I’d imagine not much, if only because 10 years ago the justification for massive compute was quite different than today.
Citation needed.
What on earth makes people assume there's still huge growth possible in the information sector? We've long been at the very end of the information age and even the latest attempts to squeeze out more benefits using "big data" have failed spectacularly. Especially when it comes to actually producing physical wealth i have seen no indications at all that we can increase anything drastically with a more efficient permutation of information. For example, Let's assume AI "solves" the traveling salesman and let's assume it indeed does find a better route for the delivery guy, what kind time savings are we looking at? The 8 hour route is done in 7.5? That's not going change any society in any measurable way. The just in time industry might become even juster in time but that's about it.
It might be dated but not too long ago a report showed the effect AI had on the SAP 500 and 490 companies couldn't even find any substantial improvement. Mind you, this is during the craze, this is happening while every manager has all the incentives to use as much of the technology as possible, lest they be the one that missed out on the new steam engine. All of these companies invested heavily in cars because none of them wanted to be stuck using horses while everyone else drives off into the sunset... yet most of them didn't manage to measurably outrun the people on horses. That's not the prediction I would have made if I were switching from horses to cars.
This reads less like a bulwark against sensationalism and more like an attempt to remain ideologically safe in what appears from many angles to be an extremely steep exponential start of what we can probably agree is an S-curve.
> And today's AI, for as revolutionary as it may be, is taking just as long to mature as anything in the past.
I suspect this opinion may come across as naive for the myriad software engineers whose job has become pressing the enter key. Your point about timing is based on solid principles, but I don't think your comment contains a satisfactory argument that the timing is incorrect.
How do you know we aren't already through the exponential part of the S-curve and about to enter the AGI winter (i.e. AGI will require a fundamentally different technology than LLMs that is still decades off)?
If no one is willing to pay increasingly high costs of training frontier models, we'll still have the models that we have now that can be fine tuned for incremental improvements. Still tons of utility in that, but little investor upside.
> How do you know we aren't already through the exponential part of the S-curve
From a recent METR report, benchmarks point towards an acceleration, LLMs are getting better at a faster rate. There's no sign that we are about to reach the top of the S curve.
> increasingly high costs of training frontier models
Somewhat surprisingly the gains do not seems to come from more compute, we have on the METR report:
> We are confident that the acceleration was not due to an acceleration in the scaling of training compute.
Link to said report: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleratio...
> Somewhat surprisingly the gains do not seems to come from more compute
I think this could form part of an argument that acceleration is slowing. It's only because we are not making as much improvement out of hardware gains that we're having to find the improvements in more complex architecture and more tricky to create training data.
The fact that the gains are not coming from the traditional place gains come from in technology might be a sign that a high percentage of the low hanging fruit has now been picked.
>From a recent METR report, benchmarks point towards an acceleration
On an internal benchmark with internal models, if you look at the public benchmark with public models there is no acceleration. That report also has a big graph labeled "fake data".
This isn't true. Metr reports show consistent increases in acceleration, 1.4x in the November report. Even with the brake of deliberate hobbling by their makers this year.
> I suspect this opinion may come across as naive for the myriad software engineers whose job has become pressing the enter key
FYI Claude Code’s default is Auto Mode these days
> appears from many angles to be an extremely steep exponential start of what we can probably agree is an S-curve.
In every bubble some say this. And it could be the exact same herd behavior (mentality) that creates a bubble.
> These bubbles aren't about potential, they're about timing.
Oh, I thought they were about unchecked greed and insufficient regulation, but I suppose that works, too
You aren't AGI pilled. NGMI
Seriously this is just missing the entire point
In 1893, GE nearly went bust as the market got cold feet that electrification was too costly, and the uses too abstract and niche. People were also hugely worried about circular financing, as they were guaranteeing bonds and taking equity stakes in their clients.
As we all know, GE then went bankrupt and we went back to whale oil.
They did go bankrupt, just a bit later :)
you are assuming overbuilding. we are probably underbuilding. it's exponential, railroads weren't exponential.
Maybe
fiber too
Why?
This argument sneaks in the fact that models are commodities which is NOT true. Claude has a different personality and performance than other models for example.
Expect more differentiation in the future and not less.
> Claude has a different personality and performance than other models for example
This kind of thing amounts to a LoRa applied to your model. There's no reason (apart from potential IP infringement) that someone else couldn't produce a "Claude LoRa" for their model (or, more likely, given the dislike for Claude's writing style, Anthropic could offer a "non-Claude LoRa" for their own models)
I completely disagree with this. What you are saying is that performance, all they need a simple trick and they get Claude performance for free. In that case, why do we see open weight models still lagging?
Assuming you meant this for personality, I’d still disagree but not that extensively. I think personality is nuanced and complex - it can’t be simply lifted and shifted. It’s grown using a meticulous process.
> I completely disagree with this. What you are saying is that performance, all they need a simple trick and they get Claude performance for free. In that case, why do we see open weight models still lagging?
Yes, they are lagging.
But "performance of Claude's from 6 months behind" is fine for a lot of tasks. And the performance of today's Claude will be still useful for many (and more and more) tasks in 6 months.
The frontier labs are making models for smaller and smaller share of the pie, as the capabilities of open models grow. That isn't sustainable
For lots of use cases you don't want the BEST model, you want the model that gets the job done for as little money as possible.
I disagree, I always reach for the best model when I’m at work. For normal questions I do use the default ChatGPT model.
Fundamentally, people underestimate the marginal increase in utility for marginal increase in “intelligence” as measured in benchmarks. Like I’d pay 5x for a model that has 2 units higher in AA intelligence.
You could tell me AA int scale is logarithmic and I would believe you
Claude’s personality is the worst aspect of the product!
Disagree but I don’t think this negates my point
"Cursor" Grok has the exact same (non-)personality as Claude.
GPT has the same performance. And performance is definitely a commodity. Why wouldn't it be?
I don’t think GPT Astra is as good as Opus 5.5, it’s very noticeable how pedantic and poor at intuiting things it is to me. It often calls things done when Claude would keep working on them, and doesn’t seem to have as good of an eye - often leaving clear visual defects and calling things done.
Something something than you can remain solvent
The market can remain 3.14159...
> Railroads, semiconductors, plastics, networking... the utility was there for all of them, and so was the potential profit.
I wonder why so few people make the comparison to the recent VR hype. We heard the same things there about its unlimited potential, how it's inevitable that we will soon resolve all its shortcomings and that it will become ubiquitous.
I still believe VR will eventually live up to its potential, but no one can deny that bubble has popped after Meta took a huge loss on it. It took too long to resolve the hard problems so the timing didn't work for them.
The major problem with VR is that it simply doesn't have a good use case beyond the initial wow effect. AI certainly does.
With AI VR does have usecases.
AI can maintain the multiverse and administrate it. Everyone can create just by talking.
Coding would also be a good usecase, to see the architecure in 3D and talk with Claude, point at stuff, pick him up and put him where the work needs to be done.
I think it's a match made in heaven, AI is just still way too slow an expensive for this.
> With AI VR does have usecases. AI can maintain the multiverse and administrate it. Everyone can create just by talking.
Nobody cared about multiverses before AI, and it wasn't because of the lack of maintenance and administration. It's just a thing that nobody cares about, AI improve it but it's not going to make people care about it so that doesn't really change anything.
> Everyone can create just by talking.
Okay, but why would they? And who's going to consume?
If anyone can trivially create, every creation is meaningless. There is no billion-dollar business in abandoned Minecraft servers, so why would this suddenly change when you force people to put on an expensive VR headset to visit them?
> If anyone can trivially create, every creation is meaningless.
Everyone has always been able to trivially create. We live in an attention economy, not a creativity economy. Meaning you can create crap (or ai slop) and people will consume it if you know how to get their attention.
Attention is what gets money. Creation is just fuel for attention. Quality is a part of the equation, but much less important than people generally believe.
Here's a legacy document explaining all this since long before ai or the internet was around: https://archive.org/details/TheManualHowToHaveANumber1TheEas...
Not measurable in ROI according to reports that keep coming out every week from companies capping spending because the wow effect is not translating into growth.
Obviously AI people say it's because they don't know what they are doing, but are you going to tell me a company like Apple couldn't figure out how translate those productivity gains into something more tangible?
It didndump a lot of money into it, improving some things and headset availability. But arguably a most of it went to useless stuff no one asked for, like paying developers to build shovelware for Quest or their useless corporate Metaverse.
Dot com crash is a mush better comparison. Over promise and under delivery in a immature market.
> I wonder why so few people make the comparison to the recent VR hype.
AI is already a gigantically transformative technology, with a staggering amount of paid consumption, and virtually everyone I know uses it every day, where I don't know anybody who regularly uses VR in any form.
The ship has absolutely sailed on "is AI a world-changing technology", where "are the leading AI companies worth what's been paid for them" is a completely open question.
AI is transformative _doesn't_ mean there isn't a bubble. Online shopping was transformative too, but there was still a dot com bubble. Amazon survived, Pets.com did not.
Because VR failed to deliver at all. The experience is narrow, isolating, and frankly pretty horrible. People get motion sickness. The set of games is limited never mind the set of non-gaming activities. AR similarly has a lot of issues like narrow FOVs, low resolution, and brightness issues.
VR is a toy with a few very narrow areas of application in industry. That does not describe the current status of AI.
VRChat is not really isolating.
Almost everywhere I the world railways are state owned and heavily subsidised and still aren’t that attractive to customers.
Yes. That's the situation in 2026 in most of the world. But in the 19th (after the railway bubble - https://en.wikipedia.org/wiki/Railway_Mania) and early 20th century, railways were actually very profitable.
I think this is untrue for Japan.
One of the few countries shaped perfectly for railways (Italy also comes to mind)
Florida East Coast railway is also privately owned, and Florida is also shaped right for railroads, being mostly a flat "straight shot".
https://en.wikipedia.org/wiki/Florida_East_Coast_Railway
Probably important to note that Japan privatised its state-owned rail network, so the private companies did not have to build out the network.
And China, and India
And commuter trains
> Almost everywhere I the world railways are state owned and heavily subsidised and still aren’t that attractive to customers.
That's a remnant of the past, where only a nation state could afford massive projects such as creating a railway line crisscrossing throughout a massive country.
Since the last couple of decades we see private companies investing in railway lines, including high-speed railway corridors. We also see a major liberalization push in Europe to decouple railway operators from infrastructure maintainers, and now there are plenty of private companies operating railway services across the world. In Spain alone there are 2 or 4 private companies providing high-speed railway connections between major urban centers.
Based on what data?
There isn't much money to be made in agriculture but food is very attractive to customers.
Yep, trains had a good run, but highways and automobiles took over. On-prem inference is the automobile.
> but highways and automobiles took over
The reasons for this are complicated and nuanced though.
Governments think they push the financial burden of transport maintenance on the people (and its a good tool to register and control people too, but thats not a strong consideration I wouldn’t wager), so they fervently cut spending on public transport and curtail cycling infrastructure in favour of cars.
But roads themselves operate at a crazy financial loss and are one of the reasons governments are in crazy debt (which nobody wants to acknowledge).
This is because road maintenance is significantly more expensive than building new roads. Which means that road use becomes heavily subsidised (much more than trains cost per user) but the cost of vehicle ownership, despite that heavy subsidy, is still enormous and one of the largest expenses an average person can have.
Cars a tool of state control and not public transport? If the government wants to surveil someone it’s much easier to see exactly where they tapped their card on and off, and if they want to stop people getting somewhere all they need to do is tell the drivers to shut it down rather than set up a bunch of roadblocks.
ANPR is much more effective than trying to track public transport usage, this isn’t even contestable.
Blocking a bus route has a large knock on effect, same for trains, and nobody is taking a bicycle away from you either.
> The reasons for this are complicated and nuanced though.
Is it? People love to have personal sovereignty over their transportation, given the choice.
I don't know why it's treated as an exclusive choice. IMO they complement each other well. Take public transport where convenient, else take the car.
You have to understand that road wear is governed by the https://en.wikipedia.org/wiki/Fourth_power_law
"The fourth power law (also known as the fourth power rule) is a rule of thumb which states that the damage to the road caused by a motor vehicle increases in proportion to the fourth power of its axle load. This law was discovered in the course of a series of scientific experiments in the United States in the late 1950s and was decisive for the development of standard construction methods in road construction."
"Example: Car (total weight 2 tonnes, 2 axles): load per axle: 1 tonnes vs Truck (total weight 30 tonnes, 3 axles): load per axle: 10 tonnes. [...] The resulting damage differential between truck and car is 15,000 to 1"
PS: this is something electric car proponents don't talk much about
The rule you linked to basically shows that whether a car weights 1.5 tonnes or 2 tonnes (an EV version of a car model usually weight about 25-30% more) is completely irrelevant to road damage when you have heavy freight using the road as well. The single 30 tonne truck from the example above gives the equivalent road wear to 15 thousand 2-tonne EV's.
worth noting that many electric cars weigh 2 tons or thereabout, like the car in your example. internal combustion cars tend to be lighter than that
> PS: this is something electric car proponents don't talk much about. In France, we're getting a "car weight tax" too (https://www.service-public.gouv.fr/particuliers/vosdroits/F3...).
I hope trucks transporting fuel for ICEs get equal treatment.
- In Europe the petrol taxes pay for road maintenance. Only electric vehicles operate at a loss.
- Train capacity is much, much higher. You couldn't have the density of France, the Netherlands or even NYC without most people travelling by train. LA struggles with 1/10th of the density.
Europe is a large disparate continent, but I don’t know any country (neither Sweden, Denmark, Norway, Estonia, Finland or the UK) that covers road maintenance costs with fuel levies.
In fact in the UK you would be forced to aggregate all fuel and road tax and offset it against road maintenance because there is no direct correlation to be made, as tax is “one big bucket”.
The majority of the planet are going from one point of interest to another at a relative constant rate- but yes people are spread further and further- ironically this is helped by requiring huge parking space everywhere anyone needs to go, and needing huge houses so you can park multiple cars outside. Its a self-reinforcing problem.
(This is most evident when comparing cities like Malmö in Sweden with Orlando, FL. Malmö has a higher total population but approximately 1/20th the road surface- yet in Orlando I kept bumping into the same people everywhere I went because we’d all drive to the same Restaurants and Shops; and despite that happening we had to drive 20-25 minutes to get there! Weird coincidence or natural concentration points? Someone smarter than me would have to say).
> in the UK you would be forced to aggregate all fuel and road tax
Just to be clear, there's no specific "road tax" in the UK - it was abolished in 1937. Instead general taxation (e.g. income tax) pays for roads.
Your point about places becoming further apart has been made by Cory Doctorow who refers to it as "geometry hates cars".
Edit: found the Cory link - you might find it interesting, but I suspect it's detailing what you already know
https://thewaroncars.org/transcript-episode-79-the-end-of-ub...
> Just to be clear, there's no specific "road tax" in the UK - it was abolished in 1937.
While this is _technically_ correct, most powered vehicles on UK roads have to pay Vehicle Excise Duty, which is popularly known as "road tax".
And is a totally different thing, and doesn't fund the roads.
The "road tax" misconception is why I feel the need to point it out as it is inaccurate and is used to weaponise hate against cyclists (e.g. "Don't bloody pay road tax"). In reality, most cyclists also drive (not me, though), but everyone who pays income tax is funding the roads which I agree with as they are shared infrastructure. However, it does mean that drivers are being subsidised by non-drivers.
Of course, with the 4th power weight relationship with road damage, ordinary cars are not really causing much damage compared to buses and trucks, so maybe it's better to state that we're subsidising logistics companies.
> Its a self-reinforcing problem.
Problem? It is THE BEST solution we have. Any suggested alternatives are worse and less convenient.
I can cycle anywhere in my city in 15 minutes- and anyone else can too.
Yes, it gets cold in the winter in Sweden- and I sometimes have to wear rainware. But the lack of noise, huge highways and being able to walk if necessary is such a crazy feeling of freedom that you can't possibly imagine.
I literally mean, you cannot imagine.
Being able to drink a beer, and then walk with a friend in the evening with my bike in hand, being able to park anywhere at any time, and being able to replace/fix my bike with no money (or, even buy a new bike without crazy paperwork or registration or licensing for my ability to drive it).
Not having anxiety about paying for parking, fuel or ensuring that I have a place to park, not to mention the lack of abusive road-rage folks.
I've lived in other places, and I've visited supposed "car-utopias" where everything is built to "great" standards and honestly the quality of life is just many orders of magnitude inferior. I've never been to the US and stayed in a nice quiet house without it being 30 minutes drive (in stop/start traffic and with aggressive drivers who cut you off) from anything of any value.
So, no, it's not the best option. It's the best option you can think of.
In germoney, IF i remember correctly, it is supposed to be like that. Seeing the state of the overall..well..corruption of autobahn bridges, roads and the "private" company governing it seems like its either not enough or the bucket has holes OR the money goes elsewhere. Pensions maybe, maybe gas power stations, who knows?
Highways are also heavily state subsidized.
Yep, that's part of why I think the analogy works.
Everything is heavily subsidized by socializing the losses and deflation of buying power for the average person.
Corporations demanding payment for everything at whatever margin they want is little different than taxation without representation and local cops shaking people down for money at random; just different legal semantics
The representation is you yourself by voting with where your money goes. At least in a market with competition.
Without competition you are talking about a monopoly and it is your government that is responsible for mitigating its effects. The government that you elected.
The mechanisms are all there, they just need to be used+embraced.
Yeah I know how the American civil religion gospel goes; don't patronize
There is inherently competition with entropy and just getting ourselves out of bed to deal with everything
Corporations and brands and tax codes are just culture war distractions on top of the raw physical competition; middle ages like words of power
The competition is with the physics that binds us to Earth; does it? Or are we just to distracted by rhetorical competition between legal constructs to embrace the levers of science
Rhetorical competition wrt to markets type blather is only seeking internal consistency with an arbitrary rhetorical framework. Not pushing the boundaries of engineering based on science.
Yall can have AI agents simulate Bloomberg terminals so actual physical scientists can focus on the necessities ground truth work you take for granted
> but highways and automobiles took over
Specifically because of massive government investment in road infrastructure in preference of railways. This would be about like the government deciding to funds DCs directly...
This is an analogy worthy of LinkedIn.
Your point is a good one, but it's constructed upon the premise that AI is a single discovery or invention or even a single tech wave.
AI is bigger than the Industrial Revolution, the Internet, than computers, than literally any innovation or invention that has ever come before, combined.
Today you see AI slop and vibe coding.
Tomorrow, all of humanity will be affected, forever.
Everything about the way that we learn, that we think, and that we invent will be changed forever.
I know where you're coming from, but we probably should not be talking about AI in this way. I don't think it's helping or serving its purpose.
I think civilization will have time to react and adapt, panic leads to haphazard action, we need deliberate consistent action.
Strange comment to make. We shouldn't be honest about AI's capabilities?
How do you think we need honest deliberate action if 95% of people believe AI is a scam and a bubble and mostly just hype marketing? Just take HN's comments for example.
You're defending one extreme by claiming the other extreme is the only option. Which isn't honest behaviour even if it was well intended. Truthfully, AI will cause real problems and we need to handle it, but it's also often exaggerated. It's a rapid technological shift, but like all the other ones anyone who works hard will still find a place to be useful.
"Tomorrow, all of humanity will be affected, forever."
This..is not being honest about AI's capabilities. Its just generic hyperbole. Also, simply not specific enough to be called honest or dishonest. It's just panic-inducing.
AI wouldn’t work without any of that above stuff though.
Their monopolies got broken and the profits disappear in a commodity market.
If nvidia or tsmc got split up i doubt its stock is going to stay at as a high a p/e.
We're discussing AI in general here, not individual companies (and none of those companies have a monopoly). Nobody's arguing that AI wouldn't exist if the previous revolutions hadn't occurred.
Investors get paid back by specific individual companies so the details matter.
This is exactly the kind of talk that led to all the other bubbles. Amazing how infected people get.
That's a fair point, but tulips are not AI.
I'm not even talking about tulips. I'm talking about The Information Superhighway in 1999, I'm talking about real estate investing in 2007.
Each event was so different with different motivations. The popularity of Tulips never back. 2007 was about hiding poor risk mortgages. 2001 every company had money thrown at them and today is about a product that costs more than the market can pay.
The motivation was the same for all of them. Money.
I dunno, we're throwing plenty of money at AI, just like dotcom got in 2001.
But it was thrown to everyone in 2001 vs 4/5 companies today.
Yeah, they at least look and smell nice and you can touch them.
The "internet bubble" didn't make the internet go away nor reduce its impact.
AI can be in a bubble right now and the bubble can explode and yet all what GP said can still be true
You know what was actually a bigger invention? Language. Farming.
Imagine a world without those. Oh, wait: tomorrow people won’t know how to use language, will they? And they won’t have water to farm with, so there you go.
AI bots trolling in an AI bubble thread says a lot about the state of AI.
I have not been able to workout if AI means demand for compute for most people will mean it's infinite since you can have AI to just work on many different problems and many different possible solutions in parallel. Or if for most people (developers included) once AI is good enough for most tasks people will reach a limit of things they could want to do and find some limit of compute enough and if so where would this limit be.
I am talking about average users and developers, there will always be tasks that want as much compute as possible I suppose but these will be limited and cash constrained as they where in the past, I can imagine most developers being satisfied with finite compute and not wanting any more for their AI agents and not needing any more intelligence for their LLM's so no more training apart from distilling and making models more efficent.
The thing I don't know is if where AI is basically good enough for largely all possible things people could want to do, if the lack of demand for training coupled with the satiated desire for compute will mean hardware falls massively in price as well as local AI getting better and cheaper to run. It feels as if some of these data centers are not going to find have the issue of massive over supply.
The logic here is so flawed it suggests to me a significant lack of financial understanding. For example:
>If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year. For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019. Adjusted for inflation, that's an even more impressive 213 billion dollars.
Alphabet's REVENUE was $162B, not earnings! You can't use all your revenue to pay off your debts because you have to use some of your revenue to pay operating and other expenses. So even if AI generates as much revenue as Alphabet did in 2019 (adjusted for inflation), there is no chance it could pay back $250B/year.
Your point is valid, but there is a flaw in the premise. They don't need to pay off the $5 trillion in 20 years. They just need to be able to service it. If they can use that debt to bootstrap the company, as long as they can pay the interest and issue new debt, they can keep that $5 trillion on the books longer than either of us will be alive
I suspect if you go back to look at say, Microsoft's books from when they went public to today they have never been debt free
Nearly every bit of logic here demonstrates a gross misunderstanding of the entire industry.
No acknowledgement of rising energy costs and problems accessing it.
Thinking you can just “turn off free users to save costs” (screw having a pipeline, I guess).
Not understanding the bespoke nature of data centers as it relates to the finite-living chips they’re designed for.
No acknowledgement of interest rates to service the massive debt they take on over the extremely long build times of the data centers.
No recognition of open weight models severely undercutting the American leaders.
This whole thing seems to be written on vibes.
The author talks about revenue growth on the basis of Anthropic's S-1 filing, but ignores the third quarter results which were posted later.
The growth slowed and operating profit has swung to a loss. The company can still make money, just not at a pace that has a chance to pay off the investments.
Also, the author conflates AI's technical potential with the market potential, naively assuming the demand will also grow as technology matures.
Looks like "Non-GAAP Adjusted Operating Income" is still +4bn$ ... if i read this correctly.
AFAIK GAAP Operating Income has stock compensation packages in it and is susceptible to valuation changes(?)
https://www.sec.gov/Archives/edgar/data/2130357/000162828026...
(I have not much experience with that stuff so please do not take that as reliable information and more as a question)
Liabilities in stocks are still liabilities. It's money they will not get if they go public. You can't just ignore these.
> naively assuming the demand will also grow as technology matures.
I don’t think this is naive at all; it’s quite rational that demand increases as price decreases, which does seem inevitable.
Most of the "laws" of economy in this form say often as a footnote "all else being equal", which remains to be seen. Otherwise you can justify any attempts at price decrease in any industry the same way.
Unless the number of users is already pretty much saturated. Which it probably is if we are honest by this point.
I absolutely disagree number of users is saturated. Fortune 500 companies, for all of their 'AI accelerating' shareholder speeches, are doing very little to actually integrate or spend on AI for the average employee. I've talked to a lot of people who work white collar jobs that could absolutely be accelerated by AI, and they just... aren't using it at all. Some of the companies make Copilot available. Some people just copy paste their tasks into the free version of chatGPT and spend a lot of time correcting the dumb model. A lot of orgs are concerned about data leakage and just have a blanket AI ban.
I recently joined a software dev team in which nobody is using any AI at all. They weren't conceptually against it, they just didn't feel the need to change their process.
There's a lot of room to grow.
> I don’t think this is naive at all; it’s quite rational that demand increases as price decreases, which does seem inevitable.
But that's not what OC is saying, isn't it? Frontier models becoming better and cheaper at the same time are not a given.
Near-frontier models that have 99% of the capabilities of frontier models, at a tenth of the cost of frontier models, might ironically be what pops the bubble.
Is there seriously a non-frontier model out there that can do what Opus 5.5 does? I can say: go look at a ticket in Jira and figure out how to fix it, write end to end tests for it, make sure it aligns with previous architectural decisions we've recorded in documents, and then test it, summarize all the testing you did, and open up a PR for review?
Yeah Opus 5.5 doesn't do that either a lot of the time you have to review it yourself or you're a moron.
Anthropic, for one, doesn't seem to be replacing their human engineers and data scientists with Opus 5.5 yet even when they would be better positioned to do so than any other company. https://www.anthropic.com/careers/jobs
But they're all using Claude (with unlimited tokens) every day, multiplying their effectiveness for sure over what even the best coders can produce.
A 10x will always be beat by 10x+ai. But ai+10x is better than just ai. The world has an over supply of ai (or soon will) but a finite number of 10x coders.
This is satire, right?
CPUs (and a million other things) got better and cheaper with time, I see no reason that AI would be any different.
The distinction between frontier/non-frontier models doesn’t matter much here, for the overall market. With CPUs, for example, Pentiums got better and cheaper while Cyrix/AMD did the same.
The real question is not if the technology is going to get better and cheaper, but rather if it's going to meet the market expectations to justify trillions of USD in valuation.
Demand will also grow as technology matures, if that maturing means more efficient and cheaper models. We're already at a stage where only a very small niche cares for better more complex models.
But which companies will capture that value? It could be like semiconductors where TSMC makes most of them but captures very little of the value.
People keep repeating this but is it true? In natural sciences, for example, there are endless applications. Anthropic models have only recently reached the level what I would call useful. Still, they are far from perfect and there's a lot of room for growth. Is science a very small niche? I don't think so.
As far as I know, Q3 for Anthropic wasn't the best because:
* OpenAI killed it with their models in Q3
* US Government ban on Anthropic models, including temp restricting Mythos/Fable
* They decided to dedicate more compute to training in Q3 instead of inference
* Opus 5 was a disappointment and Fable 5 was too expensive to inference
* They saturated all their available compute in Q3 and didn't have anymore to sell
Their growth rate has ramped up in Q4 since Opus 5.5 if OpenRouter data is a good proxy.
> But the AI industry doesn't consist of just Anthropic.
That is not the claim. Author uses Anthropic's filing as an example indicating growth of the whole industry. While questionable on its own, it is also not given in full, which I pointed out.
The competition is also a problem tho. Both of them are betting the house pretty much on winner takes all. If competition reduces growth, that's a sign that they both can't fulfill all their promises which ipso facto means one or both is overvalued.
Sounds to me like the market isn't big enough for both of them. So how much growth is left?
For the AI bubble, competition is good because it would mean better cheaper services and products.
I know this is a bit of an ad hominem attack, but it is hard to take the authors arguments seriously as they also argue that the imperial system is superior to the metric system because people in Italy apparently state their height as 183cm (in real life everyone would use 1.83m) and that calculating things using the imperial system is easier as well...
This is just general comment to all the excitable kids over here.
Anthropic is one company in a gigantic ecosystem.
You want to understand the eco system. Got to flyover the jungle, not just stare at one elephant herd.
Read annual reports of NVIDIA, atleast one from Microsoft/Amazon/Google and atleast one representing activities of Enterprise projects (Accenture/IBM/TCS). Then look at govts budgets/programmes for AI US/EU/UAE/S.Korea.
Do it with a group like a book club.
Assuming you have done that, what's your conclusion?
As a kid maybe I don’t get it, a giant ecosystem? Of frontier model providers?
Isn't that worse because you then get into neocloud companies like Coreweave and Nebius who have taken insane debt to build out data centers and if the demand sinks for those they'll go under
It gets worse. CoreWeave has more than $50 billion in debt. rising costs will make financing economically unfeasible. And a demand shock will push the company into bankruptcy. This will affect NVIDIA and others involved in this entire AI supply chain.
And what if the demand does not sink?
A "gigantic ecosystem" of two frontier model providers
Big red flags that I think point the other way:
- anthropic IPO repricing
- pareto frontier gap decreasing by the month between open and closed models
- SpaceX disappoinment
- enterprise buyers as main consumers of closed tokens are being displaced by OS competition for pennies
- job loss and entry level hiring collapse
- massive discounts of current data center contract debt
- consumer open weights model efficiency improving constantly
- the Chinese arms race
- a possible rapid downward spiral that may force governments to step in and subsidize
Am I missing anything?
- revenue growth at Anthropic slowing
- margins shrinking at Anthropic
- Google getting thwacked by missed earnings
- OpenAI missing funding target
- price war
- Microsoft revising capex projections down
- Plateauing of model capability improvement
What's weird to me is that you don't see a lot of people making deals that seem way too good to be true but yet the total numbers are absolutely gigantic. Throw on the fact that the stranded assets will have a shelf life, it really seems terrifying. Yet everything is being priced as pretty darn risky. I think maybe there can be a big slowdown without a dramatic pop. A sort of grind down. But there's not a lot of room for error to prevent a pop so I guess it will happen? I don't know its very weird.
Also of course this is all predicated on models not making another dramatic leap. The signals out of the cutting edge labs don't really suggest they have a model that would be as dramatic of a leap as Mythos was when it came out in the hopper. Their code quality seems about as mediocre as what the publicly available models can do. The OpenAI math results are very impressive but that's ripe picking for AI. Also they keep talking about a coordinated training slowdown which is textbook antitrust so they seem to want the government to force their hands off the "spend money on R&D" button.
So I don't really know. I guess I would say purely on vibes that this doesn't feel like a bubble but the size of the build out is just so massive that there really isn't a big window for a success case.
> stranded assets will have a shelf life
This is what keeps me up at night. That fiber that got overbuilt in 1999 is still there spitting photons. That Vera Rubin gathering dust on a shelf will be a paperweight in 5 years.
Even Anthropic is having their most recent debt priced at 10% interest. Wall Street is seeing them as pretty high risk.
The 10-year treasury yield is about 5%, national average 30-year fixed home mortgage is about 7.2%.
In that context I don’t think I’d describe 10% as “high risk.”
Well, this website https://usinterestrates.com/corporate-bond-yields says that currently, B-grade bonds go at around 8.5%. So bonds paying 10% would be below B, pretty much in junk bond territory https://www.investopedia.com/terms/j/junkbond.asp .
So maybe it's not high risk enough for you, but it is definitely high risk enough for the market.
But the best approach is just comparing rates. Go to https://archive.epa.gov/epawaste/nonhaz/municipal/web/pdf/bo... , and you will see a huge list of companies whose bonds are classified by armies of analysts as "less risky" than these 10% Anthropic bonds. Is Anthropic less risky than AT&T, or BP, or Firestone Tire & Rubber Co? The market says "no".
You're right. This isn't even speculations, Blackstone has said that Anthropic will literally be in the "junk" category: https://www.ft.com/content/59383c7f-8d82-493f-bf01-d2e862a10...
and you'd be thus mistaken
currently US long term bonds have very low yield and are being issued less as usual buyers move to gold/stock -- trust in US gov. has eroded especially in fiscal matters. The budget keeps growing, interest rates are still too low to combat inflation and debt is ever-expanding, without drastic measures taken to stabilize the situation: thus, high bond yield - we pay you premium to have at least someone buy it due to trustworthiness not being the sole attractor anymore
> margins shrinking at Anthropic
Why is this? Competing on price, or compute somehow getting more expensive?
Competing with chinese open-weight model which can run openly or on independent providers. Once the temporary hardware shock is over and becomes more easily available, there will be tons of independent inference providers hosting open-weight models. And I suspect we will see dozens of custom, niche distilations from open-weight models being offered by smaller groups.
Anthropic has no value proposition - just being 3-6months ahead in frontier models vs. open-weight doesn't justify trilion dollar valuation.
People keep saying this about open weight models but I haven’t heard of any real companies any of my friends work at willing to switch over to them.
My company buys Claude because it’s a trusted name with a user-friendly interface and integrations with a lot of business apps.
This “open weight models will take over based on price” is analogous saying that Linux and LibreOfficr will displace Windows and Microsoft Office in the workplace because they’re drastically cheaper.
Another analogy is that people talk a lot about the resurgence of audio CDs, but if you look at sales numbers it’s about 4% of peak audio CD sales from the year 2000. Sometimes in society a trend is more that people keep saying that a trend is happening rather than a trend is really happening.
> This “open weight models will take over based on price” is analogous saying that Linux and LibreOfficr will displace Windows and Microsoft Office in the workplace because they’re drastically cheaper.
You can't just literally drop LibreOffice into an organization and have it automatically and seamlessly work.
LLMs are, definitionally, fungible. Their input is tokens. Their output is tokens. They have minor variations in chat templates and can be trained to perform better in a particular harness (like training GPT in Codex).
Do you or your friends at work use redis? Postgres? Are they hosted at "Redis Inc" and "Postgres Inc"? No, they're commodities hosted at a huge number of places.
I went from using Anthropic via API daily to using Kimi K3 (now 3.1 pro) daily, only going back to Anthropic when a refactor is too much for Kimi to handle (once in a month's worth of work).
That surprised me. I didn't expect K3 to be as good as it was and now with 3.1, it's clear (imo) that the gap is closing. Anthropic still has a minor edge but from the 'feel' of these open models, that gap is negligible, not a moat.
The first. There have been massive waves of price cuts in Q3. OpenAI and Anthropic keep slashing prices to take market share from each other, it seems.
To me these are arguments _against_ the bubble popping. The market reacted to all these signals. The semis took the brunt of it but even there it wasn't really dramatic. These are signs of a relatively healthy environment.
The article's thread title (and proposition) is:
> I'm not convinced the AI bubble exists
You're arguing against the bubble popping. So you agree the bubble exists?
> What's weird to me is that you don't see a lot of people making deals that seem way too good to be true
You gotta be kidding. HuggingFace, a website to download files, was sold for 13B.
haha thank you :D
> Plateauing of model capability improvement
On what are you possibly basing this? The fact that there hasn't been a new model release in 6 weeks?
Have the models become significantly more capable in the last 6 months? I'm not convinced we're still seeing massive improvements as we saw at the start. The output certainly isn't getting better, the software these companies produce is actually pretty buggy.
I took a six month break from AI coding after leaving my startup and just restarted this week. I feel like Opus 5.5 is a step function improvement from 4.5 or 5, which I built a large codebase with.
Opus 5.5 is capable of completing very long running tasks and managing many threads of work with subagents without getting confused in a way that simply wasn't possible before.
For me personally? yes absoluttly.
Opus 5 vs. 5.5 is a clear stepup.
Problem is, you need to do a diverse set of experiments to recognize it. Blender support got relevant better,t he amount of corrections got less etc.
It does very much vary with the domain though. I haven't personally noticed much improvement, they're still useful but unreliable, in my view the idea of agents, harnesses etc has been a distraction and functions to cover up the lack of progress in the underlying models.
But everyone has a different view of this, it'll be interesting to see how it plays out.
In some YT talk it was explained quite well: AI is not just progressing in one field, its progressing in a lot of fields in parallel.
Before Opus 5.0, anything with 3d was not doable at all. Starting with Opus 5, it can build you cloth simulators and more.
Yes, but what's interesting here is that everything didn't get better. The original "interesting" for transformers was that a translation AI turned out to also be good at generating novel prose. But if we now need to focus to train new capabilities into the models and those don't yield better behavior everywhere then we're no longer in the "free lunch" period of LLMs.
Yes, without a question. Fable 5 and Opus 5.5 are a step-change above what we had before. On the open-weights side, GLM 5.3, DeepSeek 4.1 v1017, Qwen3.8-Flash-Next are all significantly more capable in basically any task you can give them.
You can now have a model work on a task non-stop for a day or more. This was unthinkable 12 months ago, and unstable 6 months ago.
If you are searching for evidence of capability, just look at the amount of projects, blog posts and videos on game remixes, hardware reverse engineering, new compilers and runtimes that has started coming out in the last 6 months. Slop? Sure, but this kind of slop didn't exist at this level 6 or 12 months ago.
Are you certain this is from model improvement and not improvements to agents?
And yet we have to contrast that with reports like this of model capability falling recently:
https://news.ycombinator.com/item?id=50422069
It's really very hard to pin down progress in this domain, companies are definitely targeting benchmarks.
One bit of evidence I would look at is the maths results this last year which have been impressive, that was a huge expensive brute force effort, but it is still good progress.
i am wondering in what potential world would average user or even business user require for model to work on a task for more than a couple of hours? Is assumption that models on average are used to build nextt SpaceX or something?
The web of circular financing that is going on leads to higher risk. If some ventures fail, and are unable to supply compute they are obligated to, or are unable to pay for compute they are obligated to buy, this could cause a series of financial problems for other related businesses.
I also think that some businesses may find that it might not (yet) be cost effective to focus AI use on every problem.
The pushback against data centers is getting bigger too. But america at least is usually pretty industry friendly so I'm not too sure how much it'll delay planned projects. I've read the limiting factor is sometimes just finding a source of power that can consistently deliver the electricity needed for the DCs that companies want to build.
Big enough companies who could be the best customers can run models on their infra for a fraction of the price. Eg: Revolut.
https://dealroom.co/news/169124-revolut-moves-2-000-engineer...
Yeah, seems like there is a GPU to engineer sweet spot with models like GLM. This could become a more popular pattern for some organizations. Likely also the reason gov. attempts to intervene
The biggest metric is that anthropic make ~$2 for every $1 of compute.
Two things:
1) it requires a massive money pit to make that compute worth anything
2) it requires charging a premium price
To compare to two previous bubbles; railways in the UK and USA.
running trains was really quite cheap and engines weren't expensive per passenger served. Compared to stage coaches, the number of staff needed was < 50%. the Massive fucking problem was that it needed GDPs worth of money to make the railway in the first place. Moreover as more railways were put in it significantly altered profitably of railways near them. Lines that were there for produce moving suddenly lost profitably because the produce they were moving fell out of fashion
> anthropic make ~$2 for every $1 of compute
That was true a quarter ago, when Elon still wasn't charging them. Now that he is they're making about $0.96 for ever $1 of compute.
Source?
Genuinely asking.
Just throwing numbers at eachother without back up isn't helpful to anyone reading
Anthropic's S-1 filing. I don't think the parent is necessarily posing contradicting numbers, but removing the benefit of the doubt I was giving anthropic when it comes to inference costs.
Inflation, house prices, gas prices, interest rates, huge public distrust of government and corporations, votes to restrict data center construction, endless war in Iran, etc.
100%, great additions. Someone should make a tracker site for this.
> pareto frontier gap decreasing by the month between open and closed models
Oh? This is the first I've heard that.
Some of these red flags aren’t real or backed up by much of anything.
- Is it uncommon or alarming for an IPO to be repriced?
- I’m not sure I see strong evidence to show that open weight models are making significant progress to catching up (and I’d argue that the model quality isn’t even where the business value is)
- What does “SpaceX disappointment” mean exactly?
- Are enterprise buyers buying raw tokens or are they buying productivity tooling for their employees? Not many companies switched to LibreOffice because it was cheaper than Microsoft Office.
- Look up BLS unemployment data, what job loss and entry level hiring collapse are you talking about?
- Does discounts on debt point to anything? Genuinely don’t know.
- What about the Chinese arms race is concerning? Isn’t an arms race evidence of value? There wouldn’t be an arms race if this technology wasn’t worth investing in. Doesn’t China’s desires to compete in this market bolster its merit?
- What is “possible downward spiral” and how do we know that’s real? Isn’t there a “possible downward spiral” with every industry at all times? This seems like pure speculation.
- job loss and entry level hiring collapse --> I will out that as in the group "Not a bubble" (Not that I am thinking it myself that it's not a bubble).
It looks like the end goal of replacing labor by capital is working.
With big technological shifts like this, it's important to keep in mind that the ability to do more with less resources can pan out in two different ways: (1) Keep doing what you're already doing, and lay off the people you no longer need, or (2) Do more than you did with the same amount of resources.
The answer depends on the company, but I think many companies are still figuring out whether and how they can exercise this new capacity. We don't know where it will land yet at a societal level. It's a parameter shift, and it will take some time. Pray that you live in a place with sensible labor laws and safety nets.
No. The biggest layoffs are happening at companies that need to free up additional capex for infrastructure buildout, because they're not getting ROI from their existing capex, and they're having trouble finding more money.
Layoffs are also happening at normal companies which want to compensate their AI spending.
Bingo: AI will probably take your job, just not in the way people normally mean that
That's a good point.
Political resistance against data center build-outs
I think that's actually going to moderate the bust, by giving the big companies the moat they need.
It performs the same function as regulation, replacing "the scale to navigate and manipulate the regulatory environment successfully" with "the scale to navigate and manipulate the political environment".
It's essentially grassroots regulation instead of federal regulation.
(I don't think it will stop the bust, but moderate it, allowing more companies to survive.)
I generally agree with you, but have to get this off my chest.
> pareto frontier gap decreasing by the month between open and closed models
We need a temporary moratorium on the words "pareto frontier". Every new model is the pareto frontier by some combination of metrics. Like "Bayesian Priors", people love saying / tweeting it because it sounds Smart.
It's like watching car commercials where every single vehicle has won something from JD Power and Associates. Which, (a), who the hell are they? and (b) you can only pass out so many awards until the gluttony of awards proves the award is worthless.
This isn't targeted to you! It's a pet peeve.
Agreed. This applies even to simple words like “velocity”. When you keep seeing the same key words/expressions used in support of the same argument, you know the person is parroting an idea as opposed to sharing an original thought.
> We need a temporary moratorium on the words
There is no such thing as a moratorium on words. It's another phrase that people like to repeat (a meme) until the excitement or prestige of using it has waned.
That said, i agree with you that word-salad investobabble is a sign that the bubble has put many people in a raving fever. This current bubble manifests like the cryptocurrency mania.
Also - all the gee-whiz AI companies are headed by people who don't have great business backgrounds, either by people from tech backgrounds (chasing the last of the VC), or by people with the kind of "interesting" background who you wouldn't lend your own money!
Agreed.
Also the sadder slice is all the X people token maxxing with no hope of distribution. Cool things are cool, but getting buyers is getting harder.
> Am I missing anything?
- the general lack of merit with regard to most of the words in the article
- this post being front page
> - pareto frontier gap decreasing by the month between open and closed models
This is a problem for Anthropic and OpenAI, sure, but why does it matter for e.g. Nvidia (or any of its suppliers) who will use their GPUs for training?
We need to separate the model companies from the shovel-sellers in these "AI bubble" conversations.
Maybe this is what you meant by "Chinese arms race" but there is, though not quickly, a massive wave of increased hardware supply incoming. Again, let me mention this isn't coming soon in terms of the AI bubble conversation. But China is pushing for a home grown EUV and fabs that should start coming online and having an impact in 2030.
The reason that is important is it will fuel the competitiveness of on premises and local open weight models.
Though maybe it's less useful in the "bubble" conversation since over four more years you could fit in various booms and busts before the supply side significantly starts changing.
Still useful. Especially if Chinese GPUs operating on Nvidia software become performant and accessible, thereby undercutting the price gouging that we're seeing with western chips.
The big issue will be access (Re: BYD)
one good thing about it is that GPUs don't have to be near you, so as long as they're able to be deployed somewhere, they'll be accessible
Access here means being allowed to participate in the market. BYD doesn't have a problem shipping cars, it has a problem of being allowed to sell them.
People will need compute. And if the Chinese economy can transition into providing not only factory output but cheap compute, then yes - the future is indeed theirs. Every single country on this planet will need compute.
so who is going to block sales of Chinese chips aside from the US?
I'm betting many countries will allow companies to buy and deploy them in their countries
The EU is also quite skeptical of Chinese exports, though it remains to be seen how that will evolve once the EU-USA relations settle into a new normal. Certainly the EU wouldn't be happy, though, if China managed to break ASML's monopoly on EUV machines, and at seek to retaliate with economic pressure on chips.
problem is if you can just rent cheap compute from elsewhere, you'd need to ban buying compute and all that which I'm not so sure would work. maybe heavy regulations could do it buuut....
unless I am mistaken, you missed that the datacenter infrastructure being a one-time cost. nope. those servers have about a 5 year lifetime and then are likely to be replaced. The buildings aren't the expensive part. The GPUs are.
so I think that's one big miss from this blog, or I don't understand datacenters as well as I think I do (which is always a possibility! hopefully someone will inform me if I am wrong)
> those servers have about a 5 year lifetime and then are likely to be replaced
This made sense in the past when they were not prohibitively expensive, but other than them breaking down why replace them on the 5 year schedule now? Yes, the newer stuff is better but the jumps from this point forward seem more incremental than revolutionary hardware wise. And even then, a company can simply swap out racks as parts fail.
Anthropic (iirc) said that GPUs have a failure rate of 10% per year, so you will have to replace them with an upper bound of something like 10 years.
Also, power draw is a huge cost factor, and newer models are more power efficient. If you want to stay competitive, you will have to upgrade.
Same for density. Data centers take time to build and face ever increasing opposition. By increasing density, you can improve competitiveness as well.
Musk's claim of Softbank defaulting on turbine orders for datacentres.
As the friendly article points out, the "AI is a bubble" hypothesis and "entry level hiring is collapsing because AI is replacing their jobs" hypothesis are not really compatible.
As always, reality is likely a muddled mix: there is real utility and there is wasted investment that will never pay itself back.
They actually are compatible, because the idea of replacing workers with AI in large numbers isn't compatible with the kind of economic growth these valuations are predicated on. We are not living in pre-industrial revolution times, where most people are wasting huge human effort on food growth and physical tasks, having room to move into more white collar jobs.
So, if you have massive unemployment because AI really is cheaper than labor, you'll get a massive reduction in revenue for virtually every industry, collapsing all of these valuations. Even for B2B companies, Google isn't going to be making much money selling ads to unemployed people who can't afford to buy anything.
They're compatible because they are both driven by hype, not reality.
My reality: Before AI Tokens I had to hire someone. External or normal.
After AI Tokens: I don't need to hire someone anymore and if, i need to know why. Because if i have an contractor for 1-2 years and explain a lot to them, i can just explain it to AI.
Also i prefered before already to not hire someone if i had the feeling they don't have potential, now i have even less incentive to do so.
Which means that those contractors (solo or companies) struggle to get work. For the companies that means laying off some/all employees.
The net result is unemployed or effectively unemployed people, who cut back on extra spending (games, patreon, subscriptions, etc.).
That puts pressure on those people/companies which then repeats as they have reduced revenue. So the cycle repeats.
Yeah absolutly. The question though will be if capitalism is able to update its gain function somehow or if the system collapses.
As long as robots are not there yet, there might be enough of buffer in the whole system for non computer jobs to catch but i'm not seeing this either.
It will increase pressure on every industry.
And it might also lead to an increase in speed of adoption because everyone getting fired, will have the tools to fill any niche.
And because of massive overhiring. In that scenario, the sequence of events is: Cheap money (ZIRP), companies hiring developers just to deny them to the competition (there were reports) or for empire building or as a signal to investors (etc), AI and nonzero interest rates come around, need to spend billions on AI with limited money -> lay off overhired developers and say it's because AI replaces them. It's great for companies because it allows to admit no mistake regarding the overhiring and claim to be fully in on the current thing.
Simple enough, no?
No one seems to worry about the demand of ai. That makes me wonder. My neighbors for example are normal working ppl they don’t know how to open a zip file (they call me to help them with something like this) and they are in their early 40s so not that old. Sure companies and workflows are going to become agentic but not everyone needs it or even wants it. Then there are smarter small models like Qwen 27b running on a Mac mini and this will also continue to get better and develop further. If the usage is easy I see ppl downloading a desktop app that’s dead simple instead of paying for ChatGPT. I just don’t see this growth curve going for 20 years for these companies to pay off their dept
I would say your neighbors are the perfect use case for AI. AI can understand human language. Your neighbors probably never were able to use Google for search, compare flight tickets online, Wikipedia etc. Give your neighbors something like Alexa powered with better AI, and they are sold.
Having dealt with many family members just like this, I can tell you they certainly can use Google for search, they just don't want to. When forced to they use it fine.
If I were to guess, a lot of people, particularly older people, see computer use outside the web browser as an amateur might a chop-saw; a dangerous tool not to be used by inexperienced hands, so an AI that can also directly do the tasks for them might solve that problem (assuming it came preinstalled or someone else did that for them). Though then comes the question of whether they'd be willing to pay for it
Plus hallucinations. Most "regular people" questions aren't inherently easy to verify. And when regular people do have to verify the answers, the value of "AI" drops like a rock.
> Having dealt with many family members just like this, I can tell you they certainly can use Google for search, they just don't want to. When forced to they use it fine.
Exactly! I noticed this when I moved out of my parents' house, when suddenly my mom became an electronics whiz who can hook up VCRs and TVs just fine. I always thought I was the smart person in the house, but now I realize I was just taken advantage of.
'Taken advantage of' is very negative framing. Maybe it was just a normal instance of specialization of labor!
> Give your neighbors something like Alexa powered with better AI
Didn't they literally try this and it failed miserably?
"Alexa but better" is not a profitable business. Amazon profits from Alexa being in people's homes, not from people buying Alexa.
Based on my experience, you’re right. I know a woman in her late 60s that is basically tech illiterate (can barely send an email), and she has been making effective use of the ChatGPT free plan. AI has been helping her cut through the noise of navigating the web in 2026 (which in her defense is often a bewildering experience). It can’t help her with everything yet, but it’s a clear boost to her independence online.
In the US, roughly 95% of AI users do not pay for it. So, they are sold, but only for the free tier. Technology is there, money are not.
I'd say the exact opposite. People who have no idea how technology works shouldn't touch AI at all. "AI can understand human language" is fundamentally false, and it leads people to actually believe AI is "AI" and not just a token predictor.
These people WILL fall victim of their own ignorance and upload all their private information to an AI chatbot.
There is no "honest" AI though. It is always biased towards its prompt and ultimately its training data.
Compare flights? You have to assume it is objective and not tainted by corporate bias or backroom deals(aka ads and sponsored content), unless you selfhost and find an objective model.
Search Wikipedia? You have to constantly check that it is giving you straight facts instead of being constrained by its community and safety guidelines.
The problem with so many of these usecases is treating AI as a super intelligent Vulcan who is always logical and objective aka Spock, when we are better off treating it as if it is a salesman.
This is all true, but which travel booking assistant is really unbiased? Online comparison sites might not be, the good old 'brick and mortar' travel assistant most likely isn't either.
You don't need an unbiased one. Only someone who cares about the consequences of getting caught while being blatantly unbiased.
Your neighborhood brick and mortar, or an online one may care about that while a faceless trans-national big tech will not even register that you left.
agree. That's why i love models (agents) that just search the web get the answer and use tools to their disposale instead of guessing based on training data which is outdated anyway...
Also, their neighbors are also the exact exemplars of what two generations later will look like. There isn't a lot of hard data on what AI usage might look like, but there is data on tech literacy, and the numbers are clear: We trained precisely -ONE- generation that actually understands computers. The incoming one will happily gobble up all the AI, same way they gobbled up willful ignorance of filesystems. Much like there is a button that corresponds to "gimme shot videos", there will be a button that corresponds to "Listen to what I'm about to say, and do it".
Demand will not stop in the next 20 years.
i dont know about that, they could've adopted using google which you could already search things in pretty much natural language and even got youtube snippets how fix your problem... so i'm not convinced they will adopt
Now tell them it costs $100/month and, eh, not so much.
They need less than $20/month though. Probably just a free tier with ads. That gets you google.
Right. So my worry – inasmuch that I worry for the state of the global economy – is that this doesn't generate valuations that match the current hype.
There's so much ad money to go around. It currently fuels Meta and Google. Let's say Anthropic and OpenAI enter the field and take a slice. That's a fine business, sure. But does it justify the state of the market?
> ad money/free tier
And at what cost to the underlying quality of the product and its utility, and will that cost be acceptable to the end users?
> AI can understand human language.
It is a bit comical how the HN crowd spent the last number of decades trying to convince the audience that a programmer's value was actually in their system thinking, not their arcane knowledge of a particular language, but now that the language barrier is removed it turns out it was just the arcane language knowledge after all...
Although I'd posit that the old school HNers were the ones actually right. From what I see in the average Joes is that, while they can communicate with AI, they don't really see how to leverage it. Granted, the younger generations grew up in a school system that tried to promote system thinking more so than in the past, so that gap will likely shrink over time.
This year, a lot of big companies are building the agentic layer.
This has to be build and takes time.
But this is the major thing were ai will be introduced into every single layer of software we have.
The HR System which might have needed x people? The agent has solved 90% of tickets already before they hit a human. Agents will do the repetitive but not that easy to automate tasks.
This is the point. hmontazeri is imagining where AI still has left to grow as a consumer product, but AI is currently more of a business product. Even without a new consumer use case, it's easy to see demand for tokens keep going up.
> But this is the major thing were ai will be introduced into every single layer of software we have.
How is that even possible? I thought they already put AI chats into everything. Maybe AI chat inside an AI chat.
Not just chat but an agent.
The agent will be running 24/7 and be triggered by different things. Chat input, scheduler, actions someone executes inside the software etc.
The SEO campaign tool? You might need to fill out a lot of different things, configure stuff then execute. In the future you will start the SEO campaign tool, the agent will ask you missing information, it then fills out everything, gets a budget and runs it. It will adjust the campaign based on feedback from the SEO tools.
Kind of a workflow engine but smart, so smart that it will handle A LOT more than what an traditional workflow engine could ever do.
>My neighbors for example are normal working ppl they don’t know how to open a zip file (they call me to help them with something like this) and they are in their early 40s so not that old. Sure companies and workflows are going to become agentic but not everyone needs it or even wants it.
Maybe not everyone (yet) but your neighbours clearly do need it.
Is using AI to open a zip file and not interact with the neighbor supposed to be an improvement? Maybe they just like interacting with their neighbor.
It's definitely an improvement. Most people would not want to depend on their neighbours for opening zip files. I know I wouldn't.
And if one of my neighbours ever asked me to help them open a zip file, I would not leave before I'm absolutely sure that they can do it themselves next time.
Feels like collectively we are under indexing on LLMs as user interfaces. I use a CLI with an agent to things I once did in a GUI. Your neighbors should never have to worry about zip files again.
Exactly!
I said it two years ago already: An LLM is the perfect user interface.
The problem though: It will make it a lot easier to do harder things which will devaluate jobs.
having to write "hey LLM-du-jour please unpack these files!" os strictly less efficient/more time consuming than clicking "extract" on a menu. I have hard time what usecase you have in your mind. (Unless the user is blind that is)
I'm sorry but if somebody born in the 1980s cannot or will not double click on a zip file to open it they also cannot or will not ask Claude to open it for them.
> companies and workflows [] going to become agentic
What makes you so sure? Some would call it "automated running scissors" and want tools firmly in responsible hands. And could assess "automated (indeterministic) reliability" as still far away in time.
Are there any contexts we may have missed that would surely, safely replace individual deliberate action with "the ANN can do it"?
price?
As in, "machines cost less"? Well, this general principle has always been present, as in "sucking rocks costs much less then cordon bleu", but there are "human costs" (U. Eco said for an equivalent area) involved.
Of course spending less is alluring: but you have to be able to afford spending less. Also considering that giving jobs to people who deserve it prevents other general "boomerang" social costs.
Im not convinced by the claims of infrastructure only being expensive to build initially. Hardware doesn't last forever especially in a data center, maintenance isn't free, electricity is not going to get significantly cheaper. If models get more efficient we won't need super expensive and expansive hardware build outs, if models keep getting more powerful and needy and are hardware limited, each new hardware generation will stomp over all this expensive and soon to be old hardware in electrical efficiency requiring another massive round of funding that players already holding debt will struggle to afford.
> Im not convinced by the claims of infrastructure only being expensive to build initially.
The claims seem to be true.
> A typical one-gigawatt AI data center requires $38 billion in up-front capital expenditure (CapEx) and $0.9 billion in annual operating expenses (OpEx). If CapEx is annualized over each asset’s lifespan, the total cost of ownership equates to $8.5 billion per year. Servers dominate this cost at $5 billion per year, or 60% of the total. Operating costs are small by comparison: even energy, the largest OpEx category, costs only $0.6 billion per year.
https://epoch.ai/data-insights/ai-datacenter-cost-breakdown
If new hardware comes online that doesn't allow you to payback your billions in hardware, what then? The level of rigor in the article is low.
If we steel-they the point in the essay, it still doesn't matter because a bubble is setup in the economics of the system, not whether the underlying tech works and/or is profitable. It can both work, be profitable, but have been overcapitalized before the knock on effects are in-place in the rest of the economy. This is a crazy analogy, but what if 1700s earth figured out, theoretically how to build a fusion reactor that could supply all the energy needs of the whole planet. But it cost in their equivalent dollars, 1 quadrillion dollars? You can't even integrate under their money curve to scrape together that many dollars. The only economy you can make about that is in the research side to push the cost down. Then when its dialed in, you can launch that rocket AND the rest of the system has had time at many different levels to integrate AI at various levels. OpenAI dry barebacked its way into the economy and literally screwed everything up. If they really were an altruistic AI safety lab, they should have kept working on ChatGPT, but demoed it to the world including government, so we could as a planet decide where we are taking this.
It is a huge bubble, AI is over capitalized, we have no idea what the knock on effects are and really the only benefit right now is in programming, which is huge and civilization changing. But it also has massive downsides in that it has given the most powerful even longer levers to move the earth.
I just don't think the author really has any clue.
they seem to think that debt markets are infinite. "There's Still Plenty Of Capital To Raise"
> The ones that have issued public debt, like Alphabet and Meta, have excellent bond ratings. If they need more capital, they won't struggle to find people to give them a loan.
Yes, and that bond rating is based on the idea that they really don't have that much debt, and have a net profit of something like 30%. If they hoover up a load of debt, the rating will plunge.
> I don't think the rhetoric about the AI bubble is really about finances
Its exactly about finances. The market participants that are not bollocks deep in fake Private Equity valuations are worried about circular debt obligations that are off books. The SEC has been rather lenient about reporting of non-traditional obligations (hence why spaceX was allow to do what ever the fuck it did)
For example anthropic has about 450 billion in non-cancellable obligations. the S-1 puts it at 895 billion. all of that is someone's debt.
If openAI are doing the same thing, thats pretty close to 1.6trillion in debt, both for the same customers, at the time when customers are going "hmm this isn't giving us what we hoped for the price"
Agree. CDS on the debt of several players here is getting expensive. Debt markets are getting less promiscuous quite quickly, just when even more money is needed.
It could be that a slowing of data centre build (and hence a slowing in the finance needed) takes the heat out of the situation avoiding a dramatic burst. Even then there will be casualties. If someone knows who and when ... they could do very well. More likely, Microsoft and Google will buy up anything distressed that's worth having.
Similarly, the claim that "Once all the data centres are built, AI spending is going to plummet" feels incongruous with "it's hard for me to see AI consumption not being 10x what it is now" and "Anthropics revenue has to grow to 250 billion a year to pay off debt". Scaling up seems like a constant need, especially in a market that values growth and hype about frontier models so much. (that said, the author raises things like P/E ratios that I had no awareness of, so I'm not claiming to be an expert)
The debt thing is amazing since it ignores the interest on the debt and cherry picks Google's revenue from 2019 for some reason (which is then adjusted for inflation for the comparison).
If you have 20 years to pay off 5 trillion dollars, it doesn't come out to a neat 250 billion per year as that suggests there's 0% interest and no servicing cost. Servicing costs would probably be at least 3 times the stated figure, corresponding to 10% interest paid. Some of the debt may obviously carry less interest, but the financing is opaque (as the author admits to), we have seen interests rates creep across the economy and big tech has had hits to their credit ratings.
I hope the Liberal Party can find a better candidate than the author of this piece ...
That is a weird take and after that I sort of just skimmed the rest. Yes building the datacenter is a huge expense, but you need to replace the hardware to stay competitive, and it doesn't last forever. Previously companies ran three year cycles, now five seems to be the standard. If hardware prices come back down, then.... maybe, if not those refresh cycles are going to put some datacenters out of business.
Another issue is that while e.g. Anthropic and OpenAI seems to be interested in building their own datacenters, they are mostly just renting capacity. If they can shop around, then they can get the price down, but there's a base price, due to power, hardware and staff.
The author also suggest that there's still capital to raise... no there's not. Anthropic is going for an IPO, that means institutional investors, not venture capital or borrowing and Anthropic, CoreWeave and Oracle are rated as junk. That doesn't mean that they can't raise capital exactly, but it will be very expensive and the interest payments needs to be factored into the price of the products.
> electricity is not going to get significantly cheaper
The DC push may be the thing that finally gets renewables over the finish line in the US. For example, TX is around 35% renewables right now. If that continues to rise, we may see power fall in price. It’s all about incentives.
Without a requirement for DC's to add more power to the grid I don't think the price is going down, they're essentially adding to demand and the supply isn't particularly elastic. That being said this is a good opportunity for renewables, they're the cheapest form of power.
And we just saw a pretty strong repudiation of that type of DC buildout. I have a little hope politicians can communicate it’s not DCs that are the issue, but the externalization of all the costs.
It hardly seems to be a one time expense. GPUs and networking equipment may have a useful lifespan of only about five years before needing replacement, and it could be expensive to do so.
So much this.
The current-gen hardware is expected to have only 2-3 years of useful life. New tech, and especially new architecture is going to make inference much faster and more efficient in the near future, and the big vendors are going to be more limited by datacenter floor space and power/cooling, so that the hardware will be replaced very quickly by better models.
For the time being this thesis hasn’t worked out though. H100’s are only going up in price and becoming more scarce [1]. At what point would you update if now going on 3 years into H100 lifecycle this doesn’t seem close to happen.
But they are becoming more scarce while staying on Nvidia’s shelves buy-and-hold, or being leased back in some cases. They are not being installed by customers at anything like the rate the sales figures suggest. So yeah, these H100s will last a long time in the packaging.
Meanwhile, Meta is selling their spare AI data centre capacity.
Isn’t that what SpaceX is doing too? Selling spare compute while they work on their rocket side hustle.
Where are all these unused H100s sitting and where are the people who have been fired for buying them?
They are in Nvidia warehouses. Nvidia is selling H100s and other hardware in advance to firms who do not have data centres to put them in.
Nobody is getting fired for thinking they need to buy scarce H100s in case they aren't able to get them in the future. Nobody is even getting fired for keeping their companies out of Jensen Jail — the fear that if they don't buy the whole server stack from Nvidia they might not get future GPU allocations.
There really are analysts who think Nvidia is channel-stuffing.
If the market deflates there will be a glut of very cheap compute, both in terms of data centres (though many of them will be able to offer managed model training services instead, I suspect) and in terms of unused hardware being made available for sale.
Nvidia won't lose out; everyone else is going to.
The market for compute or the stock market? People seem to have got lost conflating these two things.
Supply and demand bring what it is means that (alleged) cheap compute won’t stay cheap or go unused for very long.
People are buying v100s on ebay to run LLMs and Chinese manufacturers make custom PCIE boards for them. Older cards might not be useful for training but they're perfectly capable of doing inference work.
Look at DeepSeek! They have done full system optimization on their model, architecting it so that it would be cheap to serve. I am proud to call DeepSeek a world class tech engineering firm in a way that 99% of the tech world is not. They are a 1% company.
Jevons' Paradox. If it becomes more efficient, cost goes down, demand goes up, use rises, revenue rises.
> The current-gen hardware is expected to have only 2-3 years of useful life.
Expected by whom and why?
Google hasn't retired any of its now ten-year-old TPUs, or any of the subsequent generations. Maybe TPUs are special or Google doesn't understand compute economics but either argument requires data.
If your argument is depreciation, you're confusing a tax concept (depreciation) with useful life. Yes, they're supposedly related, but in practice, the relationship is rather loose. (Residential housing depreciates in 28 years yet typically lasts 2-3x as long and 4x is not rare.) Google's V1 TPUs
no, it's 10y. you just repurpose.
also, power and cooling agreements are usually 10y too.
> Profits Can Grow, By A Lot
Profits are mentioned but then only revenue is discussed, Revenue != Profit. Revenue can grow yeah but experimentation/token-maxing stage is ending and price wars have started, so I'm not sure about rising profits, AI Labs won't necessary be able to capture the margin(>>China<<), it might be companies/consumers who will capture most of it, if so there is a bubble - it is in AI labs with their trillion $ valuations and they might end up in a race to the bottom, BigTech is not in a bubble for the most part, they will endure a correction and fall back on their existing business.
The bigger picture is that they are willing to risk a bubble bursting before reaching ASI. Once ASI is reached this bubble will not mean anything.
The entire premise on whether it is a bubble depends on if LLMs can be more than a next token prediction machine or not. Adding reasoning agents, agent swarms, and so on to enhance the results from what is essentially just predicting the next token still means this technology at its core only predicts the next token. It is not AGI and unless something fundamental changes, it never will be. There are certainly valuable applications for this technology as we have seen in coding, math, and other sectors. But it is still not more than taking in a lot data and predicting the next token.
Meanwhile, the major tech companies first financed this build out using the cash flow from their existing revenue. Once this became more difficult to justify, they literally set up companies to keep this off their books directly, and these companies are ultimately designed to take a loss and fail. But before reaching this failure stage and defaulting on debt and obligations, they went to the criminals on Wall Street and designed a financial product structured not all that different from mortgage backed securities. Also, as the "AI" aspects of their existing businesses expand, it eats into the traditional revenue generated by their core business, which if we remember, is the original collateral being used to justify all the debt in the first place. Now these core tech companies have to win "AI", whatever that ultimately means, to not only pay the debt from the build out, but to also replace their existing collateral from their original products. Think in terms of Google Search being replaced by AI. Maybe a clean transition, maybe not. Really just depends on the technology living up to the hype. If not, the hyperscaler companies designed to take a loss a fail, will fail and with it, the circular financing will pop the very real bubble that exists.
Speaking of IPO valuations as well, it's not to say that LLMs are not a valuable technology. Just look at what they have done to computer science already. But there may be increasingly limited returns on investment if every valuable problem to be solved isn't as well defined as logical computer science problems. Without solving the hallucination problem (as a result of the next token prediction limitations), the wide application of LLMs as being sold will never happen.
I never understand the “only a next token predictor” argument. I do share the view that something fundamental about intelligence might be missing in LLMs, but I don’t see how this argument captures anything. If you can predict the exact next word Einstein is about to write, every time, you’re as intelligent as Einstein! It seems very clear that the LLM is also “thinking” about where the output is headed after this next token.
I do hear this position quite a lot though so I’d be interested to hear more about what I might be missing.
The main criticism is, imo, that these models cannot think and have no logic. This is proven time and time again. Every day something new comes out, like counting Rs or the ”drive or walk“ test etc. These are the ”fixed“ or rather band-aided on a case by case basis, but nobody seems to be able to fox the underlying problem.
Having done my master‘s thesis on Transformer LMs, I think this is a fundamental limitation of the architecture.
These models can solve math problems and then fail to count the number of problems they solved. Not being able to count to 68 is not intelligence.
Saying it is a 'next token predictor' should not diminish what LLMs are capable of. It is incredible what can be achieved by a simple rule like this.
Look what we built with only 2 digits: 0 and 1.
Next token predictor is not a very good argument. A submarine does not need to be a fish to swim.
It's a statistical machine, so you only predict a statistical Einstein, ie. something that might appear quite like Einstein and even can be fairly accurate in predicting the next word he types, but which is not him, and will be unable to predict the words that are produced not just statistically but with certain intelligence, direction and intuition.
“It” is not “thinking” for christ’s sake, it’s a massive statistical regression curve fit to a massive, frozen library of text. Unless you’re thinking of “thinking mode” which still is just a template that tells the model to ask questions, which are asked one token at a time, turtles all the way down.
LLMs cannot, by their core design, predict next word correctly every time. That is the main thing you are missing. Another thing is that I know that I am not predicting next word based on othr words in my "trainin data", there is much deeper understanding of most subjects that I am using to "predict next word". This may not be true for everyone, I cannot know for sure, but it might be possible that there are people out there who are just biological LLMs, and in that case I am certain those people will not be able to predict next word Einstein would say either.
> If you can predict the exact next word Einstein is about to write, every time, you’re as intelligent as Einstein!
Einstein is not hold in such esteem because he was really could at spewing out text or formal proofs or theorems, but because he was able to invent something no-one else could. Some people say he advanced physics 100 years wuth his theories in relativity, whether it is true or not.
So even if you can predict what Albert is going to write next, you are still merely copying old Albert. They are amazing at processing large amounts of data, naturally, but creating something new and surprising we have yet to see evidence of (the solutions of the millennium problems seem to be brute-forcing a solution based on fuckton of data to me, but I am not a mathematician so I guess it is up in the air still).
> I never understand the “only a next token predictor” argument
Yeah, that's typically the case for people who have no clue about LLMs or AI.
Is there any evidence that you yourself are more than a 'next action predictor with a multi-modal context window'?
I don't think the nature of 'predicting the next token' is the problem. One could argue that any organism works the same way. The problem is that we've had million years of evolution to stabilize the platform such that the important parts of it work without that many hallucinations -- and even then that sometimes fails and you get failure modes like e.g. schizophrenia, where the context window gets confusing and the 'next actions' don't make that much sense anymore, seen from the outside.
An LLM never experienced evolutionary pressure the same way. There's no fundamental difference between how it was incentivized to answer correctly when asked about how many Rs there are in strawberry vs whether it is safe to leave a house through the second story window, even though for any physical organism, those two would be _very_ differently weighted during training.
We have a world model.
> Is there any evidence that you yourself are more than a 'next action predictor with a multi-modal context window'?
I love the age of LLMs because people will post stuff like “the fact that I absolutely, unqualifiedly, fundamentally do not understand two separate things indicates that I understand at least one and maybe even both of them more than you do“ and then do that cool move where they pretend they were gonna high five you but don’t
it is a golden age of Posting
Is your post any better? You attack GP's character without addressing any of his arguments. You could have provided evidence for why you think his speculation is wrong, or tried to point out how his logic falls apart.
It's because of the proliferation of "insight-posting" - people thinking their little insightful "aha" thoughts are unique or have any kind of value at all.
However, insight born of knowledge ignorance is worthless. I really don't care for your great insight about nuclear fusion if you're not a nuclear engineer or physicist working on nuclear fusion, because your insight is just a hypothesis in the void that is your ignorance.
Ironically TFA falls victim to the same thing. It doesn't seem like the author is an economist or an expert of corporate finance. Their assessment of whether something is a bubble or not has absolutely no value, and carries with it the arrogant implication that the hundreds of thousands of people who are actually experts at these things never had the little 2AM epiphany that they had.
It's a smaller version of the same delusion that produces crackpots.
> entire premise on whether it is a bubble depends on if LLMs can be more than a next token prediction machine or not
I don't think current valuations need AGI to work. They just need enterprise and personal spending on AI to keep rising exponentially for a while.
We could absolutely wind up in a case where these companies are fairly close to a trillion dollars each, but not multiple trillions, at a reasonable multiple, all without AGI.
> The entire premise on whether it is a bubble depends on if LLMs can be more than a next token prediction machine or not
A theoretically perfect next token prediction machine has limitless utility. Predict the weather, stock market, lottery number, the next step of groundbreaking cancer research, etc etc. That's not the argument against current LLMs being capable of being AGI.
A theoretically perfect next token prediction machine would be uncomputable.
A theoretically perfect next token prediction machine is just an LLM with a temperature of 0, which performs worse than one with a non-zero temperature (i.e. imperfection)
The "prediction" has nothing to do with predicting the future, it just means choosing the most likely next token for existing human text
This is not how LLMs work.
I think you misunderstood the point. The point is that "next token prediction is not enough" is not true; in fact, it's plenty enough.
The best next chess move is ___
The best stock to buy now is ___
If that is all you have then the next chess move will be random (or whatever token is more likely in isolation), or the stock prediction will be the same.
For the chess example to work you need the current board state in the context or the move history. How an LLM performs will depend on what it has learned across the transformer layers and in the residual stream, as well as how much chess data it has been trained on. -- NN models like Leela and Stockfish have been trained on thousands/millions of games and have a game-specific state space representation.
The stock prediction example is similar, but doesn't have a consistent state space in which to make that prediction. It will need information in its context (current news articles, stock history of all stocks, etc.) in order to be able to make a prediction. Its performance will depend on whether it has been able to learn factors that make it possible to make accurate predictions.
> The stock prediction example is similar, but doesn't have a consistent state space in which to make that prediction. It will need information in its context (current news articles, stock history of all stocks, etc.) in order to be able to make a prediction. Its performance will depend on whether it has been able to learn factors that make it possible to make accurate predictions.
This also presumes that accurate predictions are possible.
You are completely missing the point.
The point is that next token prediction is potentially very powerful. Nobody's claiming that LLMs can predict the market or win a chess match.
But the argument against next token prediction as a powerful tool is void.
It's the transformer architecture (token embeddings and cross-token encoding) and residual stream (cross-transformer encoding and token differentiation) that make LLMs powerful tools.
The key question is not whether this architecture is powerful. The key question is how well these models can represent things outside of their design.
With chess for example, are they actually learning the rules and strategies of the game (like AlphaZero/AlphaGo, Leela, etc.) or are they just relying on a large corpus of chess games? For example, a lot of early game moves will have been played out and the best moves will likely be the more frequent moves in the training set. So LLMs could do well early game based on that game corpus.
I think that you fundamentally understand "predict" differently from those making the argument.
Predicting the next token well is indistinguishable from intelligence. The bubble is financial, as you point out, just like railroads, radio and dotcom bubbles were. The technology didn’t die, only the companies did. LLMs are in the exploit phase, that doesn’t mean there isn’t more to explore and there may be an architecture-hardware combination that takes us an order of magnitude or two higher in a step-function-like fashion just like LLMs did 2024-2026.
But despite getting better and better at not making stupid mistakes, the models are still just next-token predictors. There’s plenty more room for them to improve without a totally different architecture.
The entire premise on whether computing is a bubble depends on if computers can be more than a bit-flipping machine or not. Adding operating systems, compilers, distributed clusters, and so on to enhance the results from what is essentially just flipping bits still means this technology at its core only flips bits. It is not a general-purpose machine, and unless something fundamental changes, it never will be. There are certainly valuable applications for this technology, as we have seen in ballistic tables, census tabulation, and other sectors. But it is still not more than taking in voltages and toggling transistors between 0 and 1.
> Adding reasoning agents, agent swarms, and so on to enhance the results from what is essentially just predicting the next token still means this technology at its core only predicts the next token. It is not AGI and unless something fundamental changes, it never will be.
A lot of people have dismissed LLMs four years ago for this reason. "It cannot possibly do coding since it doesn't think, it just predicts the next token." On the other hand, Scott Alexander did say in 2019 that next token prediction is a brilliant idea and might cover a lot, including math proofs [1]. So I wouldn't say it is impossible for either 1) LLMs to be all the way to AGI or 2) LLMs to become creative enough to tell us what the AGI architecture is.
[1] https://slatestarcodex.com/2019/02/19/gpt-2-as-step-toward-g...
Software will never take off until we can show that it’s not just a bunch of simple instructions like load, store, add, and multiply.
> Just look at what they have done to computer science already.
I do. And it’s not good.
McDonalds doesn't have a moat? Eh - yes they do.
They've global procurement, the franchise model means the franchisee shoulders a lot of costs, they own the land the restaurant is on in 50%+ of the locations, and they own the technology system that underpins their app + ordering.
It's a massive moat.
That's not a moat, those are advantages.
Anybody can open a franchise restaurant business, of course not initially having the margins of a well established international behemoth.
A moat is what ASML has, having achieved something others spent many years failing.
A moat is what NVIDIA has with CUDA, against which so much software is written.
An economy of scale, which McDonalds surely has as GP suggested in various areas, supports durable profits. Seems like a moat to me.
Land is the classic cornered resource. Again a source of durable profits. You can’t open up a franchise restaurant on land owned by McDonalds; the profits from traffic earned in that location are not accessible to you.
CUDA is not a moat because it happens to be unique. CUDA provides a moat for NVIDIA due to the high switching costs and network effects it provides.
McDonalds to my knowledge does not have any network effects dynamics going on however they have entrenched switching cost mechanisms in the form of loyalty rewards programs.
If you think McDonalds has no moat you’re dreaming.
Under your definition of a moat, OpenAI and Anthropic also have ones.
Maybe we just have a slightly different definition of moat, mine focusing more on intangible assets and switching costs.
I don't think the land of McDonalds is a moat. A factor increasing profit, sure. But anybody can by land across the street immediately, only needs money.
Even with infinite money, it is questionable when and if you can cross the moats of ASML and NVIDIA.
MCD has a huge real estate moat.
Can you describe how that is a moat? A structural advantage that prevents other businesses from taking their customers, or their customers from switching. For example, how does McDonalds real estate holdings make it harder for diners to choose Burger King instead?
It gives McDonalds a lot of intrinsic value. It’s definitely a core part of their business. But it doesn’t prevent people from going to Burger King.
They own restaurants in key touristic areas that have massive amounts of potential customers going around. This is a structural advantage because people often just pick the easiest option (especially when it comes to fast food).
For example, in Paris they have two restaurants on Avenue des Champs-Élysées. While there's a Burger King right around the corner (Rue de la Boétie), a lot of hungry tourists will simply not walk in front of it.
I dont know if the valuations make sense, but the usability is all there and so are the potential profits. Over a decade, that's probably all that matters. If the valuations pop or dip (anytime soon), I think that's mostly irrelevant. I get that the press probably doesn't like to hear this boring conclusion.
P.S: A Matlab/Physics simulation guy running for a Liberal party seat in Canada? Say more!
The decade after the dotcom crash brought about as many if not more long-term successful products as the dotcom boom: Youtube, Reddit, Spotify, Google Maps come to mind
A bubble that funds all kind of crazy ideas and big investments into infrastructure work is great for laying the groundwork. But the end of the bubble won't mean the end of progress, or even necessarily a slowdown in progress
indeed, very happy to see scientifically minded folks running for office.
Is the usability all there? I just don’t see it at the moment.
The usability and potential profits of housing were there in 2007 and 2008, too.
In 2007, It was regular people who lost their homes and their leveraged investment properties. Regular people are not leveraged into AI startups today. If the AI bubble pops, regular feel people will feel the direct financial consequences significantly less than the '07 housing crisis.
> Regular people are not leveraged into AI startups today
this isn't accurate. as an example, anyone with a 401k will be affected by the stock market and the stock market is mostly tech, and tech has bet the house on ai
"Your 401k should never go down at all" is not a realistic expectation.
Hopefully, people who need to pull from their 401k's soon have wound down to less risky positions. 401k exposure is just not the same thing as your house valuation tanking.
i used 401ks here as one example. but if your company has to downsize due to exposure, that means you're out of a job.
Temporarily. The market goes up and down for all kinds of reasons. Long term though it’s never failed to go up and to the right.
For certain values of temporarily. On human time scales, we do see bear markets which last a decade or longer, and that can have a serious impact on someone's earning ability or retirement.
Markets do include significant risk and we would do well to remember that.
Unfortunately, they are, the market (even the global market) is heavily concentrated in just a few US listed companies, and other indexes closely track US indexes. A crisis in AI would impact everyone.
For proof, look at the fall in VWRL last week correlated with concerns over AI which hit nasdaq.
> In 2007, It was regular people who lost their homes and their leveraged investment properties.
And the debt on which these owners defaulted had been converted into CDOs and mortgage-backed securities, it was financial institutions and investors who lost enormous amounts of money on these.
Today there are GPU-backed securities and compute futures that are helping back all the debt that is financing the datacenter boom. Maybe it’s a bubble, maybe it’s not a bubble, but either way I think its parallel to 2007.
And the valuations made no sense in 10 yea....oh.
I'm not imaginative enough to see more use cases and who will pay. You can burn a lot of tokens and dollars on math proofs. Same for white hats finding vulnerabilities in all the software. Generating video.
Personal assistants dealing with your email, filling out paperwork, and booking reservations don't use a lot of tokens.
If every software developer paid $200/month, every academic paid what they pay for Matlab or Mathematica licenses, every YouTuber paid for whatever video processing, and every normal person paid a subscription comparable to their TV subscriptions and cell phone plans - does that justify the valuations?
The government pays for surveillance? Companies pay for customer service replacements and target advertisements?
It seems like there are a few small categories that use a lot of tokens, and a large category that doesn't need many at all.
> If every software developer paid $200/month, every academic paid what they pay for Matlab or Mathematica licenses, every YouTuber paid for whatever video processing, and every normal person paid a subscription comparable to their TV subscriptions and cell phone plans - does that justify the valuations?
> The government pays for surveillance? Companies pay for customer service replacements and target advertisements?
Yes definitely. All together that's probably within striking distance of a trillion dollars in annual revenue. Just TV subscriptions + cell phone plans in the U.S. alone is about 400 billion (leaving aside any other country).
Most large, mature tech companies are at 5-10x revenue to valuation ratios (and 10-20x EBIDTA to valuation ratios).
Of course whether you get there is a separate question, but any company that fulfilled what you're asking would pretty easily justify multi-trillion dollar valuations.
You make it sound reasonable then.
That $400B for video and phones though. That's like a $1k average for every adult and child in the country. It's twice what I spend, and I'm not on a family plan.
If every software developer paid $200/month
That's significantly underestimating the potential. Companies typically spend $200k/year for developers (salary, benefits, taxes, equipment, office space...). If AI can make them even 25% more productive, that would justify $50k/year in tokens.
I more than agree on the value, but not on the price. There's currently a competitive market, including free models that are pretty impressive if not "frontier".
I don't think LLM inference is ever going to stop being a competitive market, though. Unless regulation or other factors make the existing players form a cartel.
I think it's pretty safe to say that, unless the government steps in to prevent it, they will form a cartel or some sort of oligopoly. The incentives for them to do so are very high.
We're also one high profile cyber attack, possibly false flag, from the government granting that cartel.
Regulation wouldn't solve the problem, but many frightened people would be in favor of it anyway.
If they do for material sciences what they did for math proofs, the world is going to get weirder!
Yeah, we need a lot more kinds compilers like LEAN. Give the accountants, physicists, chemists, biologists, and so on the tools where they can verify the results.
if we were able to make a LEAN for hard science we would all be living to 200 by now
Why so short? :-)
The labs probably foresee things looking more like OpenAI solving Navier-Stokes: a company identifies a strategic opportunity locked behind an extremely difficult problem, throws a huge amount of money and compute at it, solves the problem faster than their competition, snatches the opportunity.
Current companies are paying way more than 200$ per month per developers. In startups, SMEs and big tech, you can multiply it by x10
It seems like there are a few small categories that use a lot of tokens, and a large category that doesn't need many at all.
It won't be about tokens in the long run, and I think the labs know that. It will be about long-term memory and user relationships, and who gets to own that data.
The disruption and upheaval has barely even started, IMHO.
You mean they hope to be the next Facebook/Google that knows everything about you and call sell that data to advertisers? Yay :-|
I don’t think the question is so much of AI is a future market, it is has the capital been allocated appropriately. And no one knows. The challenge Anthropic has is they have a narrow lead on the use case that makes a lot of money today, but new models come out every month and one tiny slip or wrong decision and they are toast. There is also a reliance on the idea that two things continue to happen:
1. Revenue growth increases 2. There are more highly profitable uses cases like code.
Neither are guaranteed.
I also think the narrative that these companies are profitable on compute is erroneous. You can’t ignore the other costs and no they don’t magically go away. Anthropic can’t stop training new models now or in the foreseeable future. Gemini 4 Carbon is about to come out and is apparently competitive. Google has other revenue sources. They are in the ball game. They can wait for Anthropic to run out of money.
The amount of profit they make on compute is also not particularly attractive (at least right now). It ‘s only slightly more than 2x for Anthropic, far off from many big cloud services which sell compute at something closer to 20x.
Is there evidence they are unambiguously making profit off inference when all the supply costs of that inference are considered? It does not seem likely.
Even if Anthropic and OpenAI go to zero, the bubble popping still won't hurt many people like it did in other recessions.
Actually there are two different things going on here
1. model layer companies: OpenAI, Anthropic
2. hyperscalers, data centres, hardware layer
When people say the bubble will pop, they usually mean the second one.
> I also think the narrative that these companies are profitable on compute is erroneous.
I don't think you read the Anthropic filing. It is substantially profitable if you exclude the one type of debt which was actually equity. It so happened that Amazon lent some money to Anthropic but it was tied to the valuation of Anthropic, which became a "problem" because Anthropic's valuation grew 10x. You wouldn't call this a loss in any other circumstance.
> I don't think you read the Anthropic filing. It is substantially profitable if you exclude the one type of debt which was actually equity.
Anthropic filling shows negative EBITDA, how is that profitable? Revenue < (computing + personal + selling expenses). That is not a business, that is a money burning furnace.
That before even taking into account the huge amount of money they promised to datacenters (i.e. "build that for us, and we would rent it for 5 years", non cancellable).
Sorry I meant in Q2 it is profitable excluding loan converted deductibles. There’s close to no chance that q3 and h2 is not profitable as well, unless you disagree and want to make that bet?
It heavily depends on the definition of "profitable". I bet you that "net earnings after everything distributable to shareholders" (i.e. EPS) would be negative for the foreseable future.
But this is a bet we can settle right now: https://www.sec.gov/Archives/edgar/data/2130357/000162828026... , the ammended S-1 with Q3 data (that for some reason they omited from the S-1 that made big news), page 23:
But lets compare this with S-1 non ammended (https://www.sec.gov/Archives/edgar/data/2130357/000162828026...) Hmm, seems that doubling their revenue also doubles their expenses, causing double the operating loss.First 9 months minus first half:
You are lucky you didn't bet actual money.Did you read the amended filings with Q3? Because they are not profitable in Q3 either. And their debt due is about to increase substantially. This is all before they inevitably have to issue billions of dollars in bonds with insane yields. 450 billion in commits due over the next 5 years or so.
I think there's just some core arguments here that you can totally reasonably disagree with
>In theory, an AI company has to be worried about their users switching over to a competitor or using an open source model instead. In practice, this isn't a major concern. For starters, the average person...
There's two things to disagree with here. First, this isn't theory, I work at a company and use AI heavily, switching model is as simple as changing a line in a json file. And would I do it? I've done it like 5 times this year. Claude was good, then openai was for a while, then there were times gemini was great for quick trivial tasks, now I'm mainly back on opus-5-5 with astra for some more complex tasks. Secondly, I'm not the average person, no, I'm the person openai and anthropic actually make money from. Sure, a consumer might be stickier, but so far they're not really making money from consumers, they're mainly doing B2B.
The core thing that bugs me is not so much that AI is going to be massively unprofitable, I think there's definitely going to be profitable companies. I just think if you're Google, the best case scenario is you're transitioning from fantastic economies of scale in a monopoly business to a business with costs roughly proportional to revenue and no moat. That is really bad for Google and it's not unreasonable to think it's motivation enough for them to try and undercut Anthropic and OpenAI until they're bankrupt.
But you are switching between all of the same names. You are not switching away from AI. So why would AI bubble pop?
It means that the companies have to be highly cost-competitive, they can't win market share and then raise margins. If (as is the case today) astra is bad value for money compared to opus-5-5, then I'm switching, and if (as is the case today) for lots of stuff gemini can do the low hanging fruit, I'm switching. And all of this is happening before my company has really put any boundaries on our spend, in the future I'm going to be on a budget and that's really going to impact my decisions.
Depends if the vibe changes, right now people are switching based on capability and not cost (unless it's entirely unreasonable like release Fable).
If we start to see the main switching factor be cost or a plateau in capability then I would be worried.
But according to Anthropic they are getting the healthy margin right now. If they raise margin it would be “even better”, not just “finally making any money”
But it's not clear that's sustainable. If you look at what happened in the first half of the year, they were very clearly ahead and that translated into both massive growth and great margins. But that looks like it's atleast slowed or maybe even reversed in Q3. And again, that's while there's a shortage of compute and so everyone who has compute can charge a premium for it. Shoot forward 1-2 years as this massive glut of compute hits the market, can Anthropic sustain their margins or do they have to compete with Meta and Alphabet dumping almost free inference to claw back costs on data centre investment.
> Most (or maybe even all) programmers [...] are worse at their respective field than an AI swarm equipped with a good harness
Is this the consensus view now? I'm not any kind of denier and I do recognise that motivated thinking may cause me to overestimate my own value.
But trying to be brutally neutral I'm still not sure this is true. AI doesn't make certain classes of mistakes that humans do, granted. But I'm still surprised daily even in 2026Q4 at things it does seem to make a mess of. Usually, changes that are locally good but subtly globally poor.
I thought the world of measuring programmer productivity by lines of code was dead in the 90s but clearly we see now that people forgot all of that when evaluating AI output.
I'm only slightly joking...
> These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
What is the most expensive thing in an AI datacenter? The GPU! And what it's its lifespan? 3 - 5 years!
Bubble or not, there is no way AI spending is going to plummet. It will decrease, obviously, but thinking that after 5 trillions of investment everything is profit is so naive.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet
Unless nobody is working on miniaturisation, then this reasoning might hold true.
As seen throughout history, any sufficiently absurd tech starts largescale (refrigerators, computer, radios) and eventually become portable.
Amortising data Center cost will take a long time. Within that time, if someone figures out how to make ai run on cheaper hardware, these data center cost will Become harder to recover.
> AI companies can also cut costs by no longer allowing people to use their products for free and temporarily suspending expensive AI research
That will just kill the industry.
> Amortising data Center cost will take a long time. Within that time, if someone figures out how to make ai run on cheaper hardware, these data center cost will Become harder to recover.
Or datacenters are the old mainframes, and when they stop building them and squeezing the market, GPU prices go back to what they were before the frenzy or even lower, so anyone really interested can buy a decent GPU for a fair amount of money, and run models locally for peanuts. Meanwhile, dataframe owners go bankrupt.
Per the article, _somebody_ is overvalued. On the one hand, at some point they will stop needing new hardware, and can drastically reduce costs; in this scenario, NVIDIA will hit a wall because they no longer need to buy hardware. On the other hand, NVIDIA can keep selling; in that scenario, Anthroic &c is doomed to continue purchasing hardware, so their costs can't come down.
>For starters, the average person isn't going to use a locally hosted model, ever.
Let's talk about that in 2-3 years. There is a very clear incentive to have models running locally, from data privacy to cost effectiveness. The software will pop up to easily power this, and if it is cheaper than non-open frontier models it will be used. This can be done today with a bit of hassle, and will only get easier. The consumer hardware vendors are clearly heading this way.
I'm heavily biased here. We are building a ChatGPT/Cowork like solution for data privacy concious enterprise customers.
This is already starting today, not in 2-3 years. Currently customers are using our solution with cloud-hosted LLMs, and for the meantime costs are acceptable, but we are also building capabilities to do organizational rollouts of local LLMs, so that the majority of tasks can be offloaded to the local machines of the employees (which are most of the time underutilized anyways). For a lot of typical office worker tasks, this works great. This also has the nice side-effect of allowing C-level to use AI locally for highly sensitive documents, without a risk of leakage.
I would expect that the cost-concious customers will push the broader market of harness providers to allow for solutions like this going forward. I am however not sure how much this will take hold on the consumer market.
You're forgetting every iPhone and Android user for a start!
If you don't talk about competition from DeepSeek you haven't really thought it through all the way. You might argue it's like Windows vs Linux, but operating systems have huge natural motes. Any person can just switch to a provider of an open weight model instantly. Most people don't know about it but we may very well see people learn about this broadly in the next few years. Changing models is not super complicated. Anthropic is unlikely to make $2 to the dollar when competitors are taking only pennies of profit.
> Changing models is not super complicated.
Don't the harnesses try to lock you in?
The assumption here is that having "the best" model is not worth that much of a premium over a "good enough" model at a lower price. This remains to be seen, though, and people's behaviour paying for Claude Code and Codex suggests at least for now people are willing to pay.
> This remains to be seen, though, and people's behaviour paying for Claude Code and Codex suggests at least for now people are willing to pay.
I feel like this will depend on the user.
A lot of people use Opus more than Fable, so yeah, a "good enough model" definitely exist.
That's just because opus 5.5 is arguably better than fable 5.1. I'm sure once a better model comes out, people will want to use it
Even before opus 5.5(which is opus 5), most people i know still used it more than Fable. Fable is kind of big gun that you use when opus fail.
It also assumes Anthropic can maintain the lead. If anything the competition both in the USA and in China seems to be getting closer and closer to parity with Claude.
I use Claude a lot at work and Deepseek for personal stuff, and they feel very comparable to me. I haven't found a task using Deepseek that makes me wish I had Claude - especially given how expensive Claude is.
This is important. I think people would pay a large premium for a marginal increase in intelligence. Primarily because intelligence is correlated with exponential utility which is often ignored.
For example, I'd much rather pay for the highest intelligence model when designing something or developing something because the problems it can preempt will save me more money in the long run.
The whole issue is that intelligence (as measured in AA benchmarks for example) is not linear utility
>Any person can just switch to a provider of an open weight model instantly. Most people don't know about it but we may very well see people learn about this broadly in the next few years.
I mention this in the article...
> You might argue it's like Windows vs Linux ...
Sorry what? There are 10x if not more Linux installation vs Windows installation in the world. Linux powers everything, including billions of smartphones, where Windows is nowhere to be seen. Same for routers. And actual proper servers.
If it's like Windows vs Linux it means open-weight models are going to be incredibly more successful and touch just about everything. All the while being cheaper and working not a bit but much better.
Two things can be true at the same time:
1) There's a huge demand for AI.
2) AI is way too expensive for the companies building it.
The conclusion I've come to is that AI is still revolutionary, and the buildout is probably necessary to some degree, but it is way overestimating the demand and pricing power providers will have.
SW can use a ton of AI. Lots of other applications can't, and they don't need the super smart frontier versions either. We just don't need all that much compute to serve AI for these applications, and people won't need to pay the margins necessary to get access to that AI. SW alone is not going to be enough to pay for all the investment.
I don't buy brand loyalty either. If I can pay 1$/m for the same thing from someone as I pay OpenAI or Anthropic 10$/m for, I'll do that, which almost zero barrier to entry.
The buildout is paying stratospheric prices on the assumption they can charge that much on the other end, but if demand tapers off (or model efficiency doubles again), prices will start falling. Everyone talks about Jevon's, but I don't really see why I'd spend more on AI than I need to do the things I want. I already run out of stuff to ask it to do on a basic subscription, I hardly need more.
The crux is this: I can spend x for y results. If I spend more, I probably don't get 10x for 10y results - I probably get, say, 2y results. So on that equilibrium I just don't spend all that much more because the reward for all that extra spend is not worth it. And if models get better this tradeoff starts to look worse, since I get more of the super-effective benefits for cheaper. Now I'm spending 0.1x for y results, and spending even x for 2y might not make sense, let alone 10x for 3y.
The problem is, that its not just AI for a niche thing, its AI together with compute.
People are already using AI to build them some things with CAD. Electronics too. But now what does AI need to be very good in it without an expert? Simulation and Optimizers. This again needs more compute.
And the agentic layer will be added to all big products out there. Everything you will do will consume AI tokens
To be fair, Jevons suggests that as prices fall, novel use-cases are unlocked. The things you or others currently do with AI might not need frontier models or loads of tokens, but if prices fall or capabilities increase, new applications might arise to exploit that.
Other than the fact that we’re due for one, this one is differently shaped in that there isn’t really widespread “irrational exuberance” and constant reporting about AI fortunes. I think your general investor is quite skeptical about AI and worries about exposure in their portfolio. Sure there is plenty of irrationality happening, but there is also a lot of real growth and changes in assumptions. Like the job market is way stronger than expected for even AI-threatened sectors (not entirely and I’m sorry if this affects your circles) but like we’re really just getting started here. I don’t think AI in today’s form will be the same as this time next year and that real growth is just beginning. All of the destruction of value that PE has wrought will only serve to motivate a new generation of entrepreneurs to build fundamentally better, more nimble and customer-oriented institutions.
McDonald’s is considered to have one of the biggest moats of any business. They have an extremely valuable brand, excellent locations, optimized supply chains, and so on.
Their moat is only for people who feel like eating a McDonald’s. That’s what the article is saying - when analysed they don’t actually have much of a moat because there are so many other options and competitors including cooking at home. Yet they still survive and look as if they have a moat. A moat for McDonald’s would be if their customers ate there every day or even every meal, and few people actually do that.
McDonald has a giant moat: Consistency
I can pop into any McDonald across the world and be 99.9% sure that the meal will be clean and taste like what I except McDonald to taste.
I don’t love fast food but I’ve been to a lot of them, across a wide enough range of country , to say that every other fast food chain struggles hard with that.
Again that’s the point - the moat exists if you feel like a McDonald’s, because you know what you’re getting and they are pretty good at delivering on that. But if you’re hungry now and feel like a burger you’re probably not likely to walk 30 minutes across a city to go to a McDonald’s if you’re already standing outside a burger or pizza place that looks alright.
You'd be surprised. Kids want McD's, not some random burger restaurant. They want the real thing. This value proposition exists with any restaurant with franchises though. It isn't specific to McD's. According to my wife (who fancies burgers), Burger King is way better anyway. I wouldn't know. For me, it is a scam. Since what they sell as cheddar isn't cheddar (it has nothing to do with that) but processed cheese.
Going back to my original point, kids don’t eat McDonald’s for every meal, or even 10% of their meals.
If it is my son's birthday, he gets to choose. He'd pick either 'pizza from a box' (with which he means New York Pizza for kids) or McDonald's. If I'd pick up a meal from the local snackbar, from a Burger King, or bake my own fries and burger it wouldn't cut it for him. Would it cut it for me? For sure! Cause I like my fries tasting like potato without much salt, not the garbage they dare to serve at a McD's or Burger King. But not for him. It isn't about me; it is about the kid, my son in this case.
Furthermore, if we were on the go, the prolific yellow 'M' can be seen from afar, after which kids yell: 'I see McD's! Daddy can we eat at McD's today?' They want the real deal, and its standardized throughout the country (which has adapted McD's to local cuisine; example: https://news.ycombinator.com/item?id=50423291). So in short, the kids already saw the 'M' sign that day, they know it is around in the vicinity, and you won't get away going to some alternative however good, proper, healthy, or tasty it is: it isn't McD's!
> except McDonald to taste.
'[…] expect […]'
I'm Dutch and lived near the German border. The first time eating at a German McD's was about 25 years ago, and it was an experience. Because they changed the mayonnaise to German customs. In Germany, mayonnaise is sour, not sweet and the specific McD's version (which normally is yellow, and contains, I don't know, pickles?) was rather different.
As such, your 99,9% is overrated since there's even lots of local flavours and customs, but yes it is generally good enough, and if you consider intra culture, you're flat out right.
Corals has no moat. The foo business is notoriously competitive with thin margins.
If that’s what Ai looks like they are in trouble.
Napkin math: if the 5 trillion investment is paid back by 1 billion customers, every customer pays $5000. If they want to write off the investment in 5 years, that's $1000 per year. Even with the current AI capabilities, with the right products that seems doable.
IIRC, the estimates are something like:
Estimated user share: Free 60%, Plan 35%, Consumption/API 5% Estimated revenue share: Free 0%, Plan 21%, Consumption/API 79%
So, in other words, almost all revenue comes from a small group of (B2B) customers.
I ain't paying 1000 USD per year (83,33 per month) for this garbage. I pay about 18 EUR per month for a Mistral subscription, and if it cost 3-4 the cost, I wouldn't pay for it.
The beauty is the open models: the margins are going to be thin on proprietary models. The only thing they got going for it, is creating artificial scarcity. But the Chinese are able to buy them from secondary markets anyway, or via Jensen Huang's girlfriend.
Even if the US proprietary models are 6 months ahead, there's going to be a time when the competition is good enough. I believe we are already well past that point.
> When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations, which makes the burst obvious in hindsight. With public AI companies, we don't see that happening. Alphabet, Meta, and NVIDIA are all AI companies that have seen their market capitalization explode over the past few years, but all of them have stable P/E ratios in their 20s. If the increase in their valuations were irrational (or even just in expectation of future growth as opposed to current earnings), you would expect the P/E ratios to go up. Instead, their valuations look exactly what we would expect from a healthy trillion dollar company. When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds. I'm not going to pretend that I'm an expert on financial bubbles, but I've never heard of a bubble where the companies involved had such a modest valuation based on their earnings, even if their earnings (and therefore their evaluations) are enormous.
My worry isn't so mcuh the large players ability to raise funding, it's the long long tail of "AI" in everything else and especially small startup companies that looks more worrying to me.
At the moment you have young grads getting $20-30 million with basically limited evaluation of their skills, all over the place. The area I am interested in is physics and engineering, and the number of Physics AI startups for where there there are not physicists or mathematicians with at least PhDs in numerical computing as founders for example looks to me like a total mess. And I keep seeing things where very clearly solvers haven't converged then getting used as training data and publicised in press releases from companies like this. I assume that most technical areas are seeing similar. This is where the dotcom bubble comparisons are most apt.
The only modification I'd make is to amortize the investments over 5-6 years instead of 20 because most of the cost is GPUs not the buildings. This increases the required revenue by 4x.
Well, if you buy a bunch of GPUs and don't have buildings to put them in (a very real situation right now), then you're just .... holding hardware and not making a return off it and *the building (and related power infrastructure)* becomes much more important than how many GPUs you have.
Compute hardware is ~50% of the cost of a datacenter. If we can assume a 5 year lifespan for the GPUs, that puts the ammortization timeframe closer to ~10 years.
Another weird trick (that seems to be in play) - don't unpack them yet, so they are not yet depreciating assets! Listen to https://open.spotify.com/episode/1QjM5Ud5kjHzMIY5G7hjhW?si=o...
There's one weird trick to preventing your GPUs from losing value. Buy all the ram.
The trouble that that idea is that there comes a point where the generated AI needs machines to test the code that is generated... and it actually puts a limit on the growth of AI.
I enjoyed that article, thanks for sharing. It is nice to hear that people are still having a "rage against the machine" in these heady times. For me, I think a lot of us at high market cap companies think AI is going to do xyz, but most folks are employed way outside of that boundary.
Correct me if I am wrong on this, I am only a student of mathematics and by no means a researcher.
For the programming part, I think author already says that automated solutions require a good harness, and this is indeed correct, but in my experience building good harness is hard and agents can not do this currently without human closely guiding them. And in pre-LLM programming practice we did not need harness at all, or needed only small part of it: so, solving programming problems with LLM involves more work than without LLM, even if big part of this work can be automated.
The rest of the article seems to be a speculation mostly based on the premise that LLM can really replace human in the loop, which seems to be not the case currently.
> in my experience building good harness is hard and agents can not do this currently without human closely guiding them.
I think you are right here. In the end it's probably more or less a software specification? I would love to have more formal verification in coding.
Still a Master student in CS. I attended some courses about defensive programming or programming by contract and also formal verification.
Your post repeatedly confuses evidence that the AI industry is commercially viable with evidence that the AI industry's current valuations and investment levels are economically sustainable.
I’m not convinced that something has to be overvalued for it to be a bubble. A lot of the dotcom stocks that went bust probably could’ve taken some piece of the pie, even if small, and made a business out of it if the money hadn’t run dry. Or if they had just come along later. Turns out the internet pie is enormous and it’s still growing. And now people are much more willing to use it for payments. Even some previously bad ideas might be workable now. But investors don’t have infinite patience, and the market can spook them.
Also, great products don’t always make great money. Some companies just never figure out how to monetize and eventually go belly up. Users may not be willing to pay what you expect, and the amount they pay may not be able to pay back your debt. Not every LLM chat app is going to survive the next few years. That’s just how it is when there’s sudden growth. Our economy tends towards boom and bust cycles.
> These expenses are one-off purchases, not regular operating expenses.
But no one really knows what the operating expenses are, Anthropics s1 does not mention them as far as I read them.
Also the author is wrong about that. They are ongoing costs.
I mostly use Chinese models now. Last 30 days the average token price has been 98% lower than OpenAI. This from my dashboard today. So I now use 10x the tokens on the Chinese models. More people will discover this. I can’t understand how OpenAI can recapture what they have spent.
Unless I misread it, this article simply doesn’t mention training or the cost of it.
Training would fall under "compute and infrastructure investments", which I mention in the 3rd paragraph.
You mention that training can be suspended, which is true, but don't acknowledge that puts them behind in the arms race against both other frontier companies as well as second line cheaper providers.
Without a moat, training is an ongoing cost that they can't afford to not pay
Imagine if Apple decided to "stop developing new iPhones". That's what "stop training" sounds like.
They'll stop training when they're dead, not before.
Training is not infrastructure. It’s not capex, not “investment”. It’s really an ordinary running cost of what they do, for the foreseeable future, and it’s absolutely enormous. They couldn't even dial back training in a winner-takes-it-all scenario, because training of smaller, domain specific models would remain a way to compete.
Yes, and the training data behind a lot of models is already stale and rotting fast.
Maintenance is just as important and the cost is often overlooked. The sober reality is that an LLM is just a natural language search index. For now, Google might be the only one with the capability to maintain "training" at the pace most people expect.
>The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
First mistake. No. Data centers are depreciating. And they're one or two hardware changes away from having most of the hardware in them worthless (because even if you're contend to keep running those, your competitors would move to the next tech).
>For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used. This partially offsets the risk of losing customers to a different product.
Never helped browser vendors maintain their base. Not even social networks, where the graph serves as an incentive to stay. And there's also generational churn. Facebook is a boomer wasteland now, for example.
> your competitors would move to the next tech
All of the fabs are over-booked. Where do you get allocation to print the next 10 GW of chips to replace the current ones?
And except Anthropic, all the other AI labs are producing their own specialized AI chips. They are the next tech
Is it me or is this a lot of extremely dense information?
Is this worth reading? I have to ask because of the anime avatar, which strongly discredits the author all by itself.
No, it isn't. Most of the arguments are weak and buried under a wall of text.
What kind of neurosis is this comment?
Some people have a weird bias against anime profile pictures. It's mostly ageism.
I like the sentiment, but the author doesn't appear to be accounting for some harsh truths:
- Anthropic is already expected to have junk bond status; this is going to make dealing with their existing debt rough while costs likely continue to rise
- Datacenters are 'cheap' once they come online, but that assumes they will. Power infrastructure hasn't been able to keep up, and several pending datacenters (I've seen estimates in the 50% range) meant for AI growth have been significantly delayed because of it.
- The US has a lot of debt coming up in 2028. The AI investments have been far more lucrative than the traditional 30-year bond. The AI bubble doesn't have to pop at all for the global market to plunge.
Datacenters being delayed/canceled is actually good because it reduces overinvestment.
Not if they've bought the land, started/finished building them and bought the GPUs already, but just can't get power. See the case of Oracle.
> Anthropic is already expected to have junk bond status; this is going to make dealing with their existing debt rough while costs likely continue to rise
Exactly the opposite. This is good for a tech company IPO which opens up all kinds of bond options. Most dont even have a bond rating! BB+ is a get.
> - Anthropic is already expected to have junk bond status; this is going to make dealing with their existing debt rough while costs likely continue to rise
If he is not accounting for this, then he is not saying anything of value whatsoever. So many people are still operating on vibes instead of these companies' actual financial positions, and "because Claude does cool stuff for me!".
> When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations, which makes the burst obvious in hindsight.
I think this highlights a big issue with both sides of the "AI is/isn't a bubble" debate - the idea that there's an established pattern that economic bubbles cohere to. There really isn't, and economists haven't been able to find a consistent pattern that past bubbles have conformed too (that isn't also the case of non-bubbles). "Bubbles for Farma" is a cool paper that tries to do this[0].
Like everyone else, I have my own opinions on whether there is/isn't a current AI bubble. But we should all stay humble and try and avoid stating anything as if it's an absolute fact.
[0] https://www.nber.org/system/files/working_papers/w23191/w231...
Just because a technology is going to be successful in the future doesn't mean that the current beliefs are accurate of which company is going to capture the most profits from that technology.
The internet is successful and cisco is having a blast but its valuation has still been enormously wrong during the dotcom bubble. Same goes for pets.com, etc.
So in other words, maybe the valuation of some companies is totally wrong. Maybe Anthropic is not worth two trillion, or even a hundred billion, but Nvidia is worth 20 trillion. Maybe it's the other way around. Maybe the legal framework changes and the value will be captured by the major rights holders, reddit, twitter, bertelsmann, etc.
I think the biggest problem is people are conflating current AI being useful with current AI companies being financially viable. Dot com bubble had its winners and losers but the biggest thing is that many of the current AI companies are valued at prices the winners were decades after the dot com bubble. My prediction the big 3 infrastructure AMZ, MSFT, GOOG will lead openai and anthropic to a point where they owe them trillions in payments and use it to take them over.
What exactly are the author's credentials?
Some indications in the text suggest that their insights are somewhat limited.
> In other words, once you ignore the cost of building infrastructure, AI is very profitable.
Once a provider stops such investments, they stop being "frontier" which makes them unattractive immediately. These two things are tightly connected.
> If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year. For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019.
Revenue vs profit
> Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness
Anybody doing serious software engineering will wholeheartedly disagree.
> People can theoretically switch from model to model, but people tend to settle in with a product and develop brand loyalty. For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used
LLMs are just one piece of the tool chain, a mostly interchangeable one. Which one a concrete product uses is not of concern to the end user. This quote very much misses the role of LLMs and their providers in the big tech picture.
> In 2023, a memo from Google called "We Have No Moat, And Neither Does OpenAI" was leaked, which outlined how the open source community was keeping pace with AI labs
Labs releasing the weights to their models is not what is typically called the "open source community".
I could go on but it's just not a very well-argued opinion piece. Unclear how it made it to the HN front page.
I'm actually convinced that an AI Bubble doesn't matter at all.
The quality of models like Opus 5.5 are already here, smaller labs can mimick these with a little bit of a delay.
Alone the current status has massive impact.
The status quo will get just faster and cheaper.
Then we have labs like Deepmind, companies like Google and then countries fighting over it (US America vs. China). At this point its not just private VC investing in it, its also countries aka military budgets or Research Budgets.
In parallel universities are doing their research anyway and if an AI bubble would burst tomorrow, Open Source and community driven would finally have a little bit of time to coordinate.
There was just last week an article about open post-training data. I think we want this but as long as frontier is as fast as it is, its not fun to play catch up with a moving target.
The dotcom bubble was still a bubble even if the internet still exists.
Sure, but we are a lot smarter now aren't we? So everyone can prepare accordingly.
But whenever people talk about AI Bubble, it sounds like they expect AI to stop working and impact getting reversed.
Yes, people dream of going back to the way things used to be, and irreversible steps have been taken regarding artificial intelligence. But while a certain segment of society is getting richer economically, humanity is heading toward irreversible sociological destruction.
To use an analogy, just like the scenario in post-apocalyptic movies where people eat each other, we are heading toward a grim end as a species.
What people don't realize is the exponential runaway growth when AI can make its own buyers.
I am increasingly open to the possibility that AI is undervalued. There have been recent applications of GPT 6.1 to the client business that genuinely threaten to fully eliminate their development team and put the business in the captain's chair again. I'm inches away from convincing them they have what it takes to prompt the robot to manage their codebase 100%. AI adoption is entirely a people problem now. It's a much better developer than me in most ways the business cares about.
Everyone keeps pricing this as a technology thing for some reason. This is more like a new kind of thermodynamic cycle than a new kind of software. The downstream applications are inevitable. We can't continue milling grain by using live animals walking in circles if we wish to remain competitive.
I think the bubble narrative is a necessity for some technology people. It now comes off as a fantasy that this will all blow over soon and we might go back to some semblance of normal. I'd argue that we've burned the proverbial ships at this point. There is no going back.
Think about how far we've come in such a short time. It's barely been one year of competent coding models. Some parts of tech/media seem only impressed by instant and globally successful results for some reason. Something this disruptive takes time. It took two fucking decades before I had a broadband internet connection. There are massive crow eating buffets that are still attended by those who missed the eventual consequences of the internet.
I agree that AI as a technology is probably undervalued right now. But that doesn’t mean the (current) AI companies are undervalued. I think it may well turn out that AI is a bit like electricity; transformative for the economy, but not spectacularly profitable.
> As one final note, I don't think the rhetoric about the AI bubble is really about finances. Conversation about the AI bubble is too widespread for it to be a contrarian position, and too gleeful to be driven by investors' fear. More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be.
I call this “SchadenFraud”: reveling in and playing up every little negative thing that happens in the AI world to make it look like it’s all just hype or completely irrational and will come crashing down any moment now.
There are enough people aggrieved by GenAI -- traditional media, creators, artists, coders, Mathematicians, and the list is fast growing fast -- that this kind of narrative will have an audience for quite a while.
They're not looking for solutions, they're just looking for rageviews. Worse, it gives folks the hope that things have not irrevocably changed and will go back to normal after the bubble pops.
Me, I’d rather get a stark but realistic picture of what’s happening now so I can prepare better for the future.
so many variables here. Do anthropic and openai as "AI" companies last longer term? If LLMs curve off in terms of what they can do, and local hardware becomes capable of running local models, seems like they won't last that long. But if they keep improving such that massive amounts of compute will always be needed beyond what can be done locally? Then sure. But then does it just become a commodity market where all the margins get stripped right down? Do they invent AI tech well beyond what LLMs are capable of at the moment? I would imagine, no matter which way it goes, hardware companies will win.
Great post. Entirely agree
Well, first of all, these are not one off purchases. GPUs do break and newer generations come out rendering the old GPUs outdated. I don't know if the 2.08 figure is very valid. This is based on costs of leasing GPUs from companies that may not be making money on the lease deal. What I do know is that Anthropic weren't profitable in Q3 and I'm hearing rumblings they are losing even more money in Q4.
Also Popeyes never had over $800 billion in uncancellable compute agreements. And I'm glad you're having fun with GPT, but I've moved all my projects to use DeepSeek 4.1 Flash. There are plenty of different providers. And as I use OpenRouter the change is one line.
The reason I think the PE ratios are lower than previous bubbles are because everyone already knows its a bubble. Also all these tech companies already have massive businesses and AI is not going to change that.
The P/E ratios are lower because the "E" in that ratio is completely fake
>There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for. [would -> should]
There is no logical contradiction even, just another race to the bottom, drying out of markets. The good ol' disruption playbook, like Amazon shrank or destroyed the markets to capture them, only now they are targeting all white collar jobs.
The conclusion starts: >I didn't write this article as some big takedown of the AI bubble theory. Even with this evidence, there are serious risks to the financial sustainability of the AI industry.
But the presented "evidence" is largely only: I don't know, nobody can tell the future. I mean there are also good reasons to be skeptic of skeptics like Ed Zitron, but Idk "modest valuation" seems like a big stretch. Also, I'd challenge the notion, that you'd only ever have to build that data centers _once_. Especially with the AI pace the chips are said to be replaced every three years, while the building itself could last 25 yrs. So even, if the run at the frontier cools down, I don't think running your data centers and models would come for free.
>Shameless Self-Plug I’ve applied to seek the Liberal Party of Canada nomination in Sudbury East—Manitoulin—Nickel Belt
Why do I always jump to the HN comments before opening the link
> Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well.
We can START discussing this when someone defines "worse" and we all agree on it. For SW engineering, we've been trying to define "good" for decades and we still don't know shit.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
No, that's a total straw man. For example, I am afraid of how AI will change the world, but for that it doesn't have to stop producing slop. I'm afraid of how the slop is CURRENTLY changing the world, people and societies. Frankly, I don't even care whether someone ever pays for the slop. This isn't a "I'm going to lose my job" kind of fear, it's a "we're running education into the ground and rapidly transforming into a low-trust society" kind of fear.
It’s orthogonal to the point of the article but:
“AI can't replace lawyers and doctors, but they can automate their paperwork and free up time for them to do other things”
So my question is - in software engineering (where this has bitten hardest): does anyone know _anyone_ who is less busy and has time to “do other things”? Nope, nor me.
I do many more "other" things that I previously wouldn't even dream about doing in the same time. Work that would have taken weeks now takes minutes or hours.
So yes I am doing many more "other" things. But those other things are still building features for my product.
Yeh - same. But I read “other” as “leisure things”. It’s not like we’re all having a nice time elsewhere, we’ve just compounded the amount of stuff we can deliver so we deliver even more. It’s the polar opposite of the dream where the tech does the nasty shit so we can all go chill.
The bubble is not about whether AI is profitable or whether profits can grow.
The question is how much? If they grow 5x, then these companies are very profitable, changed entire industries,... etc. Yet, the bubble still pops because investors assumed 10x at least.
I liked his point about bubble theories being more about antipathy towards the tech firms rather than rational thought. I definitely experienced this myself, where acquaintances are like "the bubble will burst any time now" and when talking to them, it's very clear that it's an anti-AI stance that motivates them to say it rather than any sober economic analysis.
My thoughts has been: what would happen if GPU prices fell 50% over the next weeks?
My gut feeling is that there would be such a massive demand to buy them to run our own models.
I am also not sure there is a (structural) bubble.
I do speculate whether some companies will survive, also some of the big ones.
>Anthropic's S-1 filing says that they make $2.08 for every $1 of compute. In other words, once you ignore the cost of building infrastructure, AI is very profitable.
If you ignore the cost of building cars, they are also very profitable.
>Profits Can Grow, By A Lot
They can also decline by a lot, and Anthropic's S-1 suggest that is what is happening, as their losses are widening, not narrowing.
>Google Has No Moat, But Neither Does McDonald's
That's why no one is sinking several trillion dollars into a new fast food chain.
>Alphabet, Meta, and NVIDIA are all AI companies that have seen their market capitalization explode over the past few years, but all of them have stable P/E ratios in their 20s.
The "E" in their PE ratios is circular revenue that is ultimately derived from debt and equity, not real end-user demand from profitable companies. For some of these bubble companies, if you subtract the debt-fueled private stimulus money, their actual PE ratios skyrocket.
> $2.08 for every $1 of compute.
sounds pretty good. if I made 2.08 for every dollar of my spend I would make more money. Plenty of businesses operate on much much thinner margins, and thats not utility-like at all.
I could only find these numbers in articles on something called crypto briefing dot com, not sure how the author got this number from the S-1.
When using traditional business metrics, such as operational net income, we'll see that Anthropic has been in the red for the entirety of its history, at times to the tune of $3B per quarter, except in Q2 2026.
When looking at amended S-1 (with Q3 numbers), page 23, "Consolidated Statements of Operational Data", we can see that for 2026 YTD (Q1 - Q3) net loss was $103B; operating loss was $1,260B.
Anthropic also disclosed up to $895B (multiyear) commitments for infrastructure build-up.
Basically, the company's financial results do not allow for any reasonable valuation (which would be more than 0), leave alone $2T or even $1T. If this is not an indication of a bubble I don't know what is.
If your hyper growth startup is not in the red for its entire history, it fucked up. One can argue their "profitable" Q2 was a mistake, startups at this stage should be far from profitable
While this is true, most of Anthropic's expenses involve locking up compute for the future. But they have no idea how much compute they will actually need in the future, or how much they're going to be able to charge their customers for it. Per token prices are dropping like a rock, and performance optimizations are making model serving more efficient. Combine that with the comparatively slower growth in organic AI demand (compared to debt-driven/investor subsidized demand), and you have the very real possibility of a 2000-style "dark fiber" problem.
That’s 52% gross margin. Not much higher than the overall S&P 500. If you break it down by sector, Real Estate, Health are, Financials (and of course Tech overall) have higher gross margins at the sector level.
That $1 of compute is not taking into account R&D spending, infrastructure, etc.
And that's their only revenue stream
>That $1 of compute is not taking into account R&D spending, infrastructure, etc
People assume this, but it is wrong. Check the S-1 for Q2, the $2.08 in revenue is for every $1.00 spent on "compute and infrastructure expenses" which also include R&D compute. (Chart is on page 137)
"Compute and infrastructure expenses primarily consist of costs attributable to operating our technology infrastructure, which supports the full range of workloads for our models, including research, model development, serving customers, and internal usage of Claude." page 146 in the S-1/A-1
It is not really clear. In the graph on page 137 does the "dollar of compute" includes only compute required for inference to provide the answer? Or does this include training as well?
It would be logical to assume that they show expense of compute required to run inference. Leaving training compute (which is huge expense) out.
It's damning that out of the only two companies in the world that absolutely must dogfood Anthropic's products, one of them, the investment bank behind Anthropic's IPO (Goldman Sachs), didn't find enough of a use case for to justify its continued use by most of the bank. Now LLM use is heavily restricted to specific departments and even in those departments spend must be approved. They've also repeatedly dropped the IPO target by almost half in just 4 weeks. That's WeWork levels of adjustment right there.
It's almost as if they know something about the finances that the rest of us don't...
You assume these companies telegraph their motives, they could just as easily be feeding this narrative so they can buy more of the company at a lower price.
This should be a top level comment
Or it's like no one wanted to believe Ed Zitron, who has been telling everyone everything they need to know about this real-life, actual AI bubble....
Along with the previous 34 AI bubbles.
I mean it's all the same bubble
What a weird thing to say: a business is profitable if we ignore a key part of its cost base. That’s true of any business, you know. Even if they make a 100% coverage of every request served, it’s still a far cry from the usual internet business model where responding to a request is almost free and most of the expense is the initial investment in software and its upkeep. This is why Alphabet used to be flush with cash and we get free services.
The trouble with Anthropic and OpenAI is, that not only do they need to get operationally profitable, they also need to be sufficiently profitable to provide a return on capital on all the money they have burned to date. Much of which has gone (and keeps going) not to assets but to expenses like the electricity bill. Thus, once the models are efficient enough to be served profitably, a relatively small amount of assets needs to be generating income for a relatively large amount of capital. It is my understanding, that currently the efficiency of the models does not grow even fast enough to outpace the need for new capital.
One more thing, that gets conveniently ignored when talking about AI cost is the cost of training the models. From one hand, there’s but a few institutions capable of building them and, from the other, it’s like the models grow on trees. Yes, you could run your open-weight model on your own hardware and pretend it’s profitable because you don’t have a comprehensive costing model and the infrastructure costs from security to electricity to risk are mixed in with your general cost base. But someone has spent astronomical amounts of money training that model you run and would probably like to get it back with some interest. Let’s hope that return is just money and not, say, the ability of a certain country to control the global software development process.
"If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year", 20? more like 5 years befroe they degrade and be useless
> Google Has No Moat, But Neither Does McDonald's
Sounds like a valid argument until you realise McDonald's is a real estate business.
There are many problems with this article. I think the bubble is real and at some point the investor money will fall off a cliff.
The value is there but the cash burn rate is extreme to say the least; and the moat is only in what you haven't released yet.
So is Google! They buy real estate on your phone to make themselves the default browser. They pay other browsers and phone manufacturers for integration.
The only reason why Google search is doing well is because it's a real estate business like McDonald's. And lately their search quality has been similarly crappy.
reminds me of Dow 36,000 [1]
1. https://en.wikipedia.org/wiki/Dow_36,000
I think one of the things that people analysing the finances of the AI sector often miss is that the capital expenses associated with data centre buildout involve spending a lot of money on compute hardware which depreciates and becomes obsolete very fast.
Large companies have often changed the depreciation schedule for server hardware to 6 years rather than 3. I think it remains to be seen if this is justified as it’s not as though the pace of change in hardware has slowed down by 50% arbitrarily.
I think it’s pretty obvious to me that a data centre full of ampere GPUs is worth less than a data centre full of Blackwell GPUs. I suspect this is also why AI companies prefer to talk in terms of gigawatts of compute, as if efficiency per watt does not exist.
Data centre costs are not one-off. Some are, for buying land and building buildings and infrastructure. The actual compute however is a constant investment. I’m not familiar enough with the market to know what proportion of expenses are ongoing but I bet it’s higher than people think.
> Large companies have often changed the depreciation schedule for server hardware to 6 years rather than 3. I think it remains to be seen if this is justified as it’s not as though the pace of change in hardware has slowed down by 50% arbitrarily.
Anecdotally, I bought a new desktop over 7 years ago and it still feels blazingly fast with plenty of room to grow. And the cost of a new machine is prohibitive, so a replacement at this time doesn’t feel very compelling.
The 16% yield on Coreweave bonds probably indicates otherwise.
and those CDS...
His analysis fails to consider trend lines in inference costs. A couple of years ago inference required a supercomputing cluster. Today, you can run quantized frontier tier models on video cards with a $1600 MSRP. At current trends we'll be running frontier models on a typical gaming computer setup in a couple of years. That is going to destroy centralization possibilities.
It also fails to consider the impact of 'good enough.' A lot of the build-out right now is in anticipation of trying to push the frontier level. But there comes a time, a time we're already arguably past, where models being used for a specific task will already be able to complete that task efficiently. Going way beyond the present should enable LLMs to solve ever more esoteric problems, but in economic terms if a company's goal is to harvest corn, and your corn harvester works well enough, they're not going to care whether or not that corn harvester can also solve Fugazi's 7th Great Conjecture in exomathematics.
Hey, I run local inference on a few RTX 6000s and will say it's pure delusion to think you get "quantized frontier tier inference" on a modern $1600 video card.
I get that you want this to be true. It is not true.
It is not frontier tier, it is not anywhere close, and if you think this you are out of touch with the frontier
Here [1] are some various benchmarks for a RTX 4090 on models like Qwen 3.8, where it usually hits around 43 tokens/s. You can find similar vids on YouTube showing it live for comparison, 'proof' if you doubt, etc. And yes, it is frontier tier.
[1] - https://getrigbase.com/gpus/rtx-4090/
I don't think these companies would be calling for regulation if there wasn't a bubble. I think they know that these models cannot keep the same progress forever, and potentially not even a lot longer. Instead of admitting this to the shareholders, they instead want to go the government and say "REGULATE ME DADDY", and blame them for not being able to keep all the promises that they're implying.
I'm not convinced by this article. It's as weak as it is long.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses
Companies are investing in scaling models. Doesn't that mean that they'll need to keep investing to stay on the edge? After all, when they lose the edge they lose customers.
> For starters, the average person isn't going to use a locally hosted model, ever
If open weights weren't a problem for these companies, ram altman wouldn't have accused the people of stealing his compute.
When are we getting an AI_OS? I just tell the prompt what to do and the entire OS is run by an agent (likely Linux with sudo permissions). Everything including browsing and copying files and what not is an agent driven task.
This sounds like a nightmare but so is using windows and office everyday and everyone is already using it.
This is a remarkably weak article to be getting as much attention as it is. It would be easy and boring to shred the whole thing systematically, so here's one example:
> When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations [...] Alphabet, Meta, and NVIDIA are all AI companies that [aren't that]
Alphabet, Meta, and NVIDIA are not AI companies, and [I don't really need to explain this, right?] ignoring the inflated elephant here is a hell of a thing.
Honestly, criticising this is beneath me (sorry) and I don't really want to be here, doing that. Oh, but the irony. This article ("I'm not convinced the AI bubble exists") -- this copium -- being a busy front page submission today is such a neat way to demonstrate the bubble's existence. Truly, a thing of beauty.
Just replace AI by block chain or Internet and you'll see that none of these arguments hold water to claim there's no bubble.
> I don't think the rhetoric about the AI bubble is really about finances
But that's purely what bubbles are about though. People see something become popular, want to be in, over invest, and the thing becomes overvalued.
Internet was a bubble. I don't think anyone would claim the value people saw in it doesn't exist, and there's way more investment in internet today than there was when the dotcom bubble burst. It was a bubble nonetheless.
> Nomination in Sudbury East
Go Bulldogs!
I am convinced for a long time. Just the way the companies CEOs talk and move suggest that they are looking for infinite investment without delivering AGI or singularity and we are always "almost there", though it does not arrive.
AGI and singularity are the Nuclear Fusion of energy.
To write such an article with zero mention of interest rates tells me everything I need to know
Like others have said, it is not that AI/LLMs//ML is not perfectly valid technology, it is a auestion of whether the current investment into it is justifiable relative to the demand and profit. We still buy and sell tulips y’know.
> Outside of generative AI, self-driving cars are becoming a reality, with some evidence to suggest that they're even better than human drivers.
Haven’t they have been just around the corner for 10 years now? And the problem is not technology as such, but the fact you are inventing trains that use roadnetwork hardly large enough for cars?
> AI adoption takes time, but there's no reason to think that AI usage isn't going to massively skyrocket, even if AI doesn't get any better than it is today (and it certainly will get better). In other words, there's massive unmet demand for AI companies to profit from.
[Emphasis mine]. This might a good example of a person writing and thinking with an LLM and letting the model enchant you with its sycophantic bullshit. We might speculate endlessly about AI or any technology, but none of that is a basis on which you should throw money at, especially not as a retail investor. We can see value production of AI in software development, but the productivity increase promised is either not in the data, or it is too early to say yet. So the real valuation of AI is really difficult say still, a lot of things needs to happen, before it can be seen to produce concrete value increase. People need to stop really really fucking hating it for one.
It kinda pointless to talk whether there is a bubble or when it will pop if there is. It might happen today at market open or never. Timing it is a fools errand. Also it is kinda pointless to shout into the void how much you love/hate AI. Better use of time would be to develop the different systems and applications that will produce value for somebody. It all sounds too similar to crypto to me still…
Going in reverse:
"The AI Industry Doesn't Look Like Any Bubble We've Seen Before" - It looks exactly like the two railroad bubbles. Potential profit is there, in the future. But you have to invest in the infrastructure first. If you can't make a profit and pay back the loans fast enough, the infrastructure goes bye-bye, along with all the funds used to finance it.
"There's Still Plenty Of Capital To Raise" - IPO is not how you raise a trillion dollars. A million, sure. A billion, maybe? But a trillion puts retail investors - mom and pop - at risk. IPO is not how you raise funds for a railroad.
"Once AI companies become profitable" - there's the big red flag. This isn't Uber or Lyft that can float along on tens of millions in corporate hand-outs every quarter. They need 100x that much money, and nobody has it to hand out. So you can't wait around for profits. There's no magic cushion to float on.
"For starters, the average person isn't going to use a locally hosted model, ever" - They will in 4 years. "normal users don't want to have to deal with the hassle or expense of setting up something" - What hassle? You download a program with a fancy GUI and run it. There are a half dozen of these projects already, that provide ChatGPT on your computer from one binary. All you need is a GPU.
"Profits Can Grow, By A Lot" - Only if they remain cheaper than local hardware or open-weight inference providers. Which won't happen because they owe more money than they can make.
The sign of the bubble is all the investment made not based on stable returns, but on hype. The weird geniuses from San Francisco said the future will depend on this one technology, all work will be changed forever, and only their company can deliver it. If you don't get in now, you'll miss out on a fortune! That hype just doesn't work out. And the proof is here: the superhuman machine sucks at writing, the documents it creates are lies, and we seem to be spending more on it than we get out of it. Because it's not magic. It's just a machine.
> It looks exactly like the two railroad bubbles. Potential profit is there, in the future. But you have to invest in the infrastructure first. If you can't make a profit and pay back the loans fast enough, the infrastructure goes bye-bye, along with all the funds used to finance it.
That's just not true. Profit is there now. If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!
Totally unlike the railroad bubbles which built out years and years ahead with no customers and then collapsed.
> "There's Still Plenty Of Capital To Raise" - IPO is not how you raise a trillion dollars. A million, sure. A billion, maybe? But a trillion puts retail investors - mom and pop - at risk. IPO is not how you raise funds for a railroad.
You're totally confused about what an IPO is. You're confusing the valuation of Anthropic with what they will raise. They intend to offer about $100B in stock, their valuation is at 1T.
> "Once AI companies become profitable" - there's the big red flag. This isn't Uber or Lyft that can float along on tens of millions in corporate hand-outs every quarter. They need 100x that much money, and nobody has it to hand out. So you can't wait around for profits. There's no magic cushion to float on.
Most AI companies are profitable aside from expenditures to build new infrastructure to reap more profits. That's a very healthy sign.
> "For starters, the average person isn't going to use a locally hosted model, ever" - They will in 4 years. "normal users don't want to have to deal with the hassle or expense of setting up something" - What hassle? You download a program with a fancy GUI and run it. There are a half dozen of these projects already, that provide ChatGPT on your computer from one binary. All you need is a GPU.
Uhuh. A GPU. In your phone? In your laptop? The average laptop in the US is sold for about $700, desktop for about $1k. A good GPU to run a reasonable model is $10k+. It makes zero financial sense for most people to run their own inference. The issue isn't, and never was, the software, it's the hardware. And one thing we've seen from the past years of growth is that people don't look at the models from 3 years ago and say "Oh well, that's good enough, at least I can run it slowly on my $5k+ desktop". Until that happens this is a fantasy.
> If the hyperscalers stopped building new infra tomorrow, they would just roll in the profits available from what exists today!
They owe hundreds of billions in loans on datacenters that aren't built yet and don't currently have capacity for the new users and additional tokens they promise are coming. If they don't build capacity the increase of demand would cause more outages and higher prices which would crater both their use and valuation. They have to grow or die.
> They intend to offer about $100B in stock, their valuation is at 1T
That's one company, the aggregate debt of the US AI industry approaches 1T.
> Most AI companies are profitable aside from expenditures to build new infrastructure
Which they have to complete to make the profit they promised which they need to pay off the loans they already have.
> A good GPU to run a reasonable model is $10k+
Today. Due to supply and demand. Funny thing, supply and demand. It's very reliable. High demand makes supply short for a while, but then causes supply to go up. When supply goes up, eventually demand goes down, which brings prices down. The chip factories and economies of scale that will be complete in a few years will send prices crashing, as they have before. Which might bankrupt the remaining chip companies.
"Reasonable model" isn't a thing. Many different models of different size have capacity to do many tasks. I have a laptop from 3 years ago that runs local AI models. My phone from 5 years ago runs local AI models. And any strong CPU can run models (no GPU needed). Speech to text, image/video generation, image classification, research, analysis, computer use, troubleshooting, etc, all works on tiny models. I think you don't understand the tech here. Remember that many people spend a grand every year just for a newer iPhone/Samsung. They'll spend the same for one with better AI and more RAM.
You've been saying so many incorrect things.
> They owe hundreds of billions in loans on datacenters that aren't built yet
Covered by their current revenue plus 10-20B.
> to make the profit they promised which they need to pay off the loans they already have.
No, they don't need to make profit for this. Their loans and infrastructure expenses are mostly covered by a small increase in revenue
> The chip factories and economies of scale that will be complete in a few years will send prices crashing, as they have before.
All signs point to demand increasing to match. Economies of scale let you better capitalize on existing demand, so the companies stand to profit even as prices decrease.
> They have to grow or die.
This is the story of every startup. Even large companies are subject to this in the stock market.
Meta, I have constructive feedback on your choice of titles with this post https://news.ycombinator.com/item?id=50415422
There are very few people in the world where when they declare that they have an opinion on it, that the declaration itself is title worthy.
Yeah, this "just my opinion" humbragging style is an unfortunate antipattern.
Talking this way is a big part of the "tech bro" smell.
tech bro is leading indicator, they are emulated. we gotta change course.
It's more accurate to say that there's an OpenAI-Anthropic-X bubble. Financially sustainable investment in LLMs is useful and not going anywhere, as China shows. Dumping trillions of dollars into companies that make models that are 1.2x better than Chinese models at 100~1000x the cost, ehhh... good luck with that.
Im in China, and scared. Everyone is broke, losing their jobs while on the other hand there are these giant datacenter campuses being build in the warmest and most remote places. if the bubble bursts (there is 1000% a bubble, its the same like mass maker and innovation, sharing bikes and now ai policies)… it will be bigger than any financial crisis before. The only shit I see coming out of ai hackathons and competitions here is ai tools for marketing (ai). The rest is ai slop microdramas.
To what extent do people “on the street” as it were relate this to the real estate bubble?
No one really, they just fear their jobs.
Constant anxiety is widespread around the world. Although I'd accept it as a symptom, I'd not trust people to find the correct root cause for it.
I think this is correct except that it ignores the hidden bubble in the SPVs: a lot of the firms who look only partially at risk have simply moved debt off their balance sheets- but to whom? Pension funds, sovereign wealth funds etc.
And “OpenAI bubble” alone has one massive implication: the likely demise of Oracle, a blue chip business that is deeply embedded into the western world.
On the one hand, if the air is let out of the bubble before those SPVs get themselves too deep into building things that won’t be needed, that is good. But nevertheless the failures of the SPVs will harm people who are not in the exclusive billionaire club.
Also there is a considerable money-go-round; the money for everything will dry up fast when the cloud LLM businesses can’t convince people to give them money to buy Nvidia GPUs anymore l
This analysis marginally better than Rob Ford smoking crack, I'm not convinced you should run for office, but if you succeed, please get us more Shorsey.
Putting data and rational analysis aside, the style of this post is so similar to what I was seeing during the crypto bubble.
The first reason I became suspicious of the AI bubble is it is all the same people hyping it up. The grift just switched to the new thing.
A part of the bubble that people don’t talk about are all the small AI products and integrations that don’t do much of anything. But they took VC money to build an AI project management program and now they have to cram it in there even though it is useless and actively costs them more money. There is going to be a lot of pull back at some point with these pointless integrations and I think that is going to be felt on total AI usage.
Sure, the largest IPO in history was slashed by fifty percent before it came onto the market because no investor wants to underwrite its original offering, but it’s not a bubble
I love the AI doomerism in this thread
How much do you think these same people will be complaining about AI companies in 5 years when they realize they missed it? AI is the greatest wealth building opportunity of my generation and people are just dooming themselves away from it.
"If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now."
-------------
If everyone, or even the majority, believed that there is a bubble that's going to pop with near certainty, that would tank stock prices immediately. Why would anyone hold a stock they know is going to go down precipitously? If the majority of people know it's a bubble, the bubble immediately pops. AI stocks still have worth, so most people believe there is no bubble or that prices will go up before going down.
People talk about bubbles. They worry about bubbles. For the most part, though, they go on investing as though there isn't going to be a bubble. Every bubble looked like a solid investment to most people right up until it burst, or it wouldn't have been a bubble.
So, no, most people do not think AI is a bubble. That doesn't necessarily it's not though.
> If everyone, or even the majority, believed that there is a bubble that's going to pop with near certainty, that would tank stock prices immediately.
That seems to assume:
1. Those people are the invest-class and have control over what gets kept/pulled.
2. Those people don't think it'll pop later and that there's Great Fool Money to be had for just one more day at the tables.
if i plug my ears and close my eyes the 16% neocloud bond yields suddenly don't look so bad
am i getting this right`?
dump n pump. day trading. long term strategies. build up for the next big thing or delusion or hysteria or lie.
if you read mainstream narratives like "AI bubble", translate it to several millions of people making lots of bucks via a lot of fraud mixed into some good, genuine, valuable work of productive people.
investors and wall street people doing anything financial is not tied to bottom up signals but to insider knowledge and top down supply chain LEGAL manipulation by people not directly involved in the actual work done.
or it's a wrap up ( made money, time to shift or prep something that will hypergrow again, maybe already re-pre-hyped ).
or it's reactions to having maxed out some part of the supply chain which had negative consequences to some other market which reduced overall gains enough to realize "oops", if we don't re-act now, future gains in other markets will be reduced almost catastrophically.
bubble talks always feels like a mix of hyped people really getting fucked over by fraud runners
the details are not too relevant for the moment as I can't have play money to invest but the narratives are interesting because they directly translate into what I read in the suggestions and advice, subliminal or direct, from people to other people.
Profits will drop from competition. McDonalds has a moat.
Those are errors just from the subheadings.
It’s the unemployment that will disrupt not the AI.
This person lives in a bubble. IPO of Antropic is literally a sign there is no more capital in private markets and it’s up to the public to take the shit bucket.
Let’s not even talk about the second headline that costs are constrained, while one company spends few dozens off billions per quarter.
The thing is, the bubble exists because even though there is no guarantee of any big profit coming from the big techs, there are billions of dollars funneling into these companies. If you look at their current numbers, they are drowning in losses. Yet they are forcing to keep this bubble alive
The discussion of the “AI bubble” often conflates two different things: (1) is AI going to justify the valuations we are seeing? and (2) are the specific AI companies we have going to justify their own valuations.
The rapid progress of multiple closed frontier models, along with the rapid progress of open models, suggests that LLMs will be like electric motors, not like social media. Are electric motors worth a huge amount of money economically? Sure. But none of the companies who invented electric motors are especially valuable today, because the technology became a commodity.
> The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.
Because hardware doesn’t need updates. Newer hardware is more capable which gives competitors with newer hardware an advantage.
> Outside of these expenses, AI doesn't cost that much to run
Outside the expensive stuff it’s cheap.
So if a shoeshine boy tells you which stocks to buy, it's time to cash out.
But if the shoeshine boy tells you it's a bubble, it's time to buy?
If only shoeshine boys still existed, I would know what to do.
> If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now.
You complain that people take the bubble as a given. Yet you continue by doing what a lot of people on the other side do:
> Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well.
This isn't established. Like, at all. There's serious doubts as to whether AI-driven codebases remain maintainable, understandable and trustworthy over time. You look at a snapshot and conclude, while riding the hype wave.
The technology itself is here to stay i believe, but until they figure out a way to make it more efficient and not just burning money, a part of it does feel like a bubble
I think AI is fine. I suspect there is a data center bubble.
ok
“Its a bubble”
“Its a bubble”
“Its a bubble”
“Its a bubble”
“Maybe its not a bubble”
*bubble pops*
every time.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
One thing to keep in mind when you're "debating" things online is you're actually debating with a tag team of people, with a whole spectrum of views. And at any point people with stronger arguments (starting from different precepts) will reply while people with weaker arguments will fall silent.
That's why "the average hacker news poster" is simultaneously rich and poor, powerful and powerless.
Deluded
title is an attempt to switch from wild singularity tone, hype, to a mature reasonable tone, hype.
> More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be. People are afraid of what's going to happen with AI, and rightfully so.
> So, people wishcast. They tell themselves that this is all a giant bubble, and one day soon, everything will crash down and immediately go back to how it used to be.
+1 to this. I think this is why Ed Zitron gets so many readers - people want someone to tell them things will go back to “normal” someday.
I see it as the opposite. I think the this idea, the "seeing the LLMs as a bubble representing the hope that things can go back to normal" is itself projection from people who know their future is going to suck thanks to "AI" but really wish differently. Whether the LLM companies crash and burn, or succeed wildly, the outcome is going to be negative for the average American (and probably for the average human being).
It will eventually blow up of course but it wont be anything really expectacular, a bunch of larpers will be talking about giant loses for CEOs and surely, hopefully anthropic, all while it doesnt have a big effect on the people that know little about all of this. But still, nothing will be back to how it was, ever
Do people believe that? If AI is indeed a giant bubble and indeed the bubble pops, do people believe that AI will just... disappear from society?
I am pretty sure the biggest thing in question is if OpenAI and Anthropic are as valuable as they say they are vs any other AI labs. I guess we will see, but I still feel it doesn’t take many ex-employees from them to build something that can wrestle them, and also it’s weird seeing other players that also have comparable AI models (Alphabet, Meta, whatever company owns Grok these days, …) do not get as much in value. There is also the “Can local models disrupt this?” question (and the Chinese labs), and there is also the question of hardware availability (will some player change this?).
I feel the projected value of both OpenAI and Anthropic requires that nothing changes in a space of technology where things seems to change a lot and reasonably suddenly.
Does Open AI or Anthropic fall behind? I think if either falls behind people are going to panic and the eventual stock price will tank. Both companies have a ton of commits so much in fact that one of them failing might take one or two other companies with them. That is what I think the big risk is.
Not people on here or who use it regularly, but on other parts of the net.
Maybe they know on some level it won’t go away, but the “wishcasting” leads to rooting for the bubble in the hope that maybe it actually will.
No, no one believes that. Everyone can see that AI is going to change how many industries work completely. The parent is constructing a strawman argument "the reasoning is motivated and that motivation is nostalgia (or similar)".
Even Ed Zitron rarely talks about AI outside financials. I have no idea what OP is talking about.
I guess they mean as far as AI being this utopian cure all for the ills of the world that some might have us believe is coming, its BS?
Yes, quite a few people believe AI is a big scam and it’s not actually good for anything.
I know this might seem crazy to software developers who handed all their dev work over to Claude 6 months ago. But most people are not software developers.
I think a lot of software developers are convinced it must be transforming other jobs and it really isn’t and may never to the same extent. LLMs are just particularly good for technical writing and code.
But that is the potential issue. It is valued as it will and that just isn’t proven.
I have also seen this from many SWE and hobbyist programmers as well, actually, as not all workplaces have AI mandates.
Those people usually bury their head in the sand and have such strong opinions about the capability of these models as they haven't ever used anything but the free tier chatbots, if any at all.
I mean if one of the big labs collapse, then yes it mostly will? Like sure, the tech will stick around in some way or another as long as it makes money or provides a good user experience, but the forcing a garbage chatbot into every product surface, the constant flavor of the week hype, the endless doom from every CEO will all cease to exist.
> Do people believe that?
No, but it is incredible for somebody to use the word “wishcasting” and then describe a group of people that would need to materialize in order for their point to make sense
I think a lot of people who post on HN don't want to admit that, apart from chatbots, AI is completely invisible to the user. A lot of users already hate chatbots to begin with and avoid using them.
So, it has already mostly disappeared, and the lack of practical applications means it will continue to bleed money. Software devs aren't using AI either for the bulk of smaller tasks that require teamwork with non-technical stakeholders.
Yes, it totally can disappear from society. The remaining use cases go with local LLMs and specialized software built around it. From the user's perspective, that's the way things always were and we'll be back to "the way it was" when the bubble is done bursting.
-1 to this. I think this is why Ed Zitron gets under the skin of AI evangelicals. They want everyone on this surrealist funny bus on a journey to whatever-the-fuck. Fingers crossed! Ignore the weird AF obfuscated financials. Ignore the crippled hardware market for everyone except data centre overlords. Nothing to see here!
I think a data center bubble exists and may have just been forcibly popped. But that's never really been an AI bubble. It's a real estate bubble wearing a shirt that says AI.
I stopped reading after the 10th sentence or so demonstrating AI-pilled psychosis and hand-waving. That it is obviously this or obviously that and obviously some other thing.
I look forward to these types of posts fading in frequency much more so than AI use itself.
Meta’s margins are apparently getting crushed as a result of AI investments.
The lack of a clearly profitable operating model and huge (trillions) in commitments for buildout is what concerns people.
I am not convinced by this person who is not convinced there is a bubble. The bubble, by the way, specifically refers to the valuations. Is Anthropic worth 1.2 trillion when it’s not clear they will have positive cash flow in the foreseeable future?
> Is Anthropic worth 1.2 trillion when it’s not clear they will have positive cash flow in the foreseeable future?
Maybe think of it as a bimodal thing. I can easily imagine a future in which Anthropic builds a sustainable profit model, grows into new sectors, and is credibly worth 7T in ten years.
I can also easily imagine a future in which AI capabilities continue to increase but the amount that enterprises are willing to spend on Anthropic’s services to do, and in which that, plus ever-increasing competition, cause Anthropic’s revenue to grow slowly, stall, or even contract. And that Anthropic would be unable to raise sufficient capital to make up for it and would therefore forced to default on their obligations. (Or that they would be able to raise enough capital but on poor enough terms that existing investors take substantial losses as a result.)
> Maybe think of it as a bimodal thing. And how does that justify the valuation that they are planning to IPO at?
It's just a total gamble, is that the claim?
> I can easily imagine a future in which Anthropic builds a sustainable profit model, grows into new sectors, and is credibly worth 7T in ten years.
I can easily imagine a future in which Anthropic fails to find a market fit that results in the kinds of profit that will cover their current spending commitments, and where the company is another casualty of progress.
If it were a sure bet, then it would be worth 10 trillion today.
Of course it is a Gamble. Anyone who knows anything understands there's risk
“A risk” is not the same as a gamble. Of course there is risk. But companies don’t have trillion dollar valuations when there isn’t a clear path to profitability.
Every one of the few companies with that kind of valuation generates tens or hundreds of billions per year in free cash flow.
> Meta’s margins are apparently getting crushed as a result of AI investments.
They are down because of capex investment, but Meta is showing what’s possible with AI. They have been shrinking their workforce and steadily growing.
A fair critique. Even I, someone who started as a (righteous) angry doomer over it, have come around for the most part.
My viewpoint also shifted around the shape of the bubble. The technology isn't going away, obviously, but the valuations of these companies like OpenAI and Anthropic feels divorced from reality. Inference continues to go down, most folks don't need frontier models for common tasks (including coding), and local models provide a competitive edge in several domains the frontier labs cannot (namely value and privacy). Combined with demand still largely subsidized and the companies being deeply unprofitable, and I just don't see the investments in Anthropic or OpenAI or the sort actually paying off. That is the bubble I'm concerned about, not the tech itself.
...now when the debts get written off and all this compute is just sitting around idle, available for bottom-dollar prices to try and eke out any sort of ROI for whichever creditor(s) inherit the assets? That's going to be pretty cool times.
AI can both be a world transformative technology and OpenAI and Anthropic can go bankrupt.
>In other words, once you ignore the cost of building infrastructure, AI is very profitable.
Its not clear that the infrastructure will ever be paid back. Its also not clear that once models are trained and distilled that the infrastructure will be useful. Ignoring 1.9 trillion dollars of debt is probably not the greatest financial analysis.
>However, there's no reason for why this level of spending has to be maintained in the first place.
Exactly. So Anthropic and OpenAI might be holding the bags for latecomers who distill their models and get ahead. Sort of the Commodore64 and Atari of the AI space.
>Once all the data centres are built, AI spending is going to plummet.
Well data centers need maintenance, and GPUs and similar do have a very short lifespan. But even if spending plummets repayments on existing debt needs to be accounted for.
>In a pinch, AI companies can also cut costs by no longer allowing people to use their products for free and temporarily suspending expensive AI research.
Research keeps them in the headlines, see above about the latecomers.
>For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019. Adjusted for inflation, that's an even more impressive 213 billion dollars.
Yep, so all Open AI or Anthropic need to do is get acquired by a much more profitable business to pay back their debt.
>at some point, the rate of increase has to level off, but that doesn't have to happen anytime soon.
The rate of increase of revenue in the AI sector also doesnt need to be apportioned to OpenAI or Anthropic. If it wasnt for their generous loss leading subscriptions or their rnd, theres not much there to hold on to people.
>it's hard for me to see AI consumption not being 10x what it is now.
Right AI consumption will increase very likely. But will it increase at the cost of an already subsidised subscription, or increase the use of cheap Deepseek 4.1 tokens.
>For starters, the average person isn't going to use a locally hosted model, ever.
Isnt this expressly where Apple Intelligence is heading? Building small low cost models that run on Apple hardware? If I was nVidia I would have both plans being run down at the same time, looking towards selling GPUs to hyperscalers on one side, and working with OEMs to deploy AI coprocessors in laptops on the other side.
>For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used.
N = 1. OpenAI is already what halfway to their trust thermocline anyway?
>A potential way for a bubble to pop is for the companies involved to run out of capital to spend while they grow, causing them to fail.
Isnt Anthropic already down to BB+?
> NVIDIA are all AI companies that have seen their market capitalization explode over the past few years
NVIDIA is a hardware company that can has so much demand on their products they will likely survive 3 AI market collapses.
>Alphabet, Meta, and NVIDIA
Alphabet, Meta and NVidia all have other sources of revenue. They are companies building AI, not AI companies.
>When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds.
Anthropic and OpenAI dont have "E" (as in earnings)
Anthropics Debt to Equity is 363%.
>A potential way that this could be a bubble is if AI proves to be a fad.
A potential way this could be a bubble is if the makers of consumer devices point customers at cheaper or local models as part of their onboarding experience. OpenAI and Anthropic get thoroughly zuckerberged by the people holding the primary relationship with the customer base.
Or perhaps, the funding to the top 1% of token purchasers also turns out to be loans, and that dries up too.
https://www.pymnts.com/news/artificial-intelligence/2026/ope...
>As one final note, I don't think the rhetoric about the AI bubble is really about finances.
The people bringing it up every 10 seconds are doing it to spread FUD about AI. That doesnt really change the financial picture.
A bubble can't pop as long as most people, or at least most dollars, still think it's a bubble. Bubble pops can only take the market by surprise.
That is not true. The dot com bubble was widely discussed a year+ in advance of it popping.
> The AI bubble is a speculative concept that the AI industry is extremely over-valued. If true, the value of most companies involved in AI (several of whom are worth well over a trillion dollars) will collapse, devastating the entire economy. If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now. I don't think that this is a reasonable assumption, though.
Matteo's analysis is very flawed though. For instance, he projects that the outcome is "devastating the entire economy". Why would that have to be the case though?
The AI bubble can collapse partially. There are tons of examples of how this has happened in the past. Not everything is dot-com 2.0. What we see right now is a "consolidation" bordering on partial collapse of some of the companies' projected revenues. It makes sense to assume that there is not a complete collapse, and some companies will end up dominating more than others.
> If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now. I don't think that this is a reasonable assumption, though.
To say that it is (essentially) taken for granted isn’t a great start to a rebuke. You have to engage with the arguments that they have, and they have presented arguments. Just see born-again clanker Dan Luu respond to Zitron’s claims.[1]
I don’t see any engagement with skeptic’s arguments in this piece.
[1] https://danluu.com/zitron/
> In other words, there's massive unmet demand for AI companies to profit from.
Of course. There are hundreds of millions of white collar workers that are jealuously hoarding their unrightful wages.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for. You can't really have it both ways here.
But you can. A bubble of this magnitude will negatively impact the world, not just AI investor cretins. You can take the extreme position that AI is useless and still be worried about the fallout.
> It's completely valid to be afraid of AI taking your job, but money is going to change hands when that happens.
The conjunction here implies a contrast which I fail to see.
> As one final note, I don't think the rhetoric about the AI bubble is really about finances. Conversation about the AI bubble is too widespread for it to be a contrarian position, and too gleeful to be driven by investors' fear. More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be. People are afraid of what's going to happen with AI, [...]
The always-amazing observation that your intellectual opponents have motivated reasoning. Likewise, most AI positives have motivated reasoning, like that they can enrich themselves or that they are getting so much code out of their current subscription prices.
When I write about AI it’s clear when I argue against AI itself and when I am whining about my boo-boo thoughts about losing all of my economic prospects.
> I’ve applied to seek the Liberal Party of Canada nomination
That tracks.
I think the last paragraph puts it pretty well
>>> As one final note, I don't think the rhetoric about the AI bubble is really about finances. Conversation about the AI bubble is too widespread for it to be a contrarian position, and too gleeful to be driven by investors' fear. More than anything, I think the supposed AI bubble has become such a large talking point because it represents the hope that things can go back to how they used to be. People are afraid of what's going to happen with AI, and rightfully so. Beyond fear for the future, AI in the present has become uncomfortably common, with AI slop everywhere online and increasingly even IRL. This has all happened so fast, it's easy to forget that ChatGPT came out a little less than 4 years ago. Our world has changed in such a scary and odd way, and while I have a pretty optimistic view of AI overall, I get why people would want it all to go away. So, people wishcast. They tell themselves that this is all a giant bubble, and one day soon, everything will crash down and immediately go back to how it used to be.
And on a personal note, this probably goes against all economic theory, but I think the larger the scale gets of money, the less "real" it becomes. The US national debt is $40.5 trillion. I personally don't think a fraction of that will ever really be paid. There are rich and poor, but once you become rich, that added richness is essentially useless, and the debt we see is a mixture of money missing from the poor person's wallet, or the money missing from a billionaire's portfolio. One of these is more sensually palpable than the other.
If your product is good enough, shows potential, and has meaning, it will get funding. I have never done a startup or gotten venture capital so it's pretentious for me to say, but I think the money just comes, through effort (and I think 100's of pitches for the new startup), and we can't always know from where. It's just there. My brain is too small to consider where that money comes from. We're not valuing money at a certain point, but just sheer utility, and so investors will be there to back you up. In decades-long terms, sometimes centuries-long.
The debt will keep building up, but so will be our infrastructural changes. Early adoptees will get rich to degrees, and investors who held a conservative view on major progress like AI (will still be rich but) will be slightly less rich. Maybe there will be unpaid debt, like the 500 billion incurring a $400 billion loss during the dot-com crash. But the technology will continue growing.
After all, money doesn't mean anything anyway (in the theoretical sense). It buys us soup, clothes, and a place to pay rent to, but it is just an agreement. And with the changes in the world economy that some of us are expecting to see after the AI revolution grows even more, I really don't know what the meaning of money will become.
Large data centers never made sense. If an additional pc with a graphics card was put into every home, we wouldn't need power grid upgrades. Our power grid is designed around human distribution i.e. homes. Large data centers cherry pick where it may be possible to quickly connect power connections up. Data centers need to secure their power connections before the utility rate hikes on everyone else generate public opposition.
If large data centers cannot secure free power connections paid for by everyone else, they cannot make any of this profitable. The whole thing was a scam from the start. This is why cloud AI companies bought up so much pc and server hardware. Artificial scarcity was the only way to slow down local llm adoption and sell the false narrative that cloud AI has value.
Plucking a graphics card out of every home and centralizing the cards to a handful of large data centers that require a trillion dollars in power grid upgrades is an idea so dumb, we should all be laughing at it. They even have to use active cooling or evaporate drinking water instead of free air cooling. The consumer doesn't benefit in any way from this. Consumers would essentially be paying more just to be datamined.
LLMs are not a bubble, large data centers are a bubble. Look at how large data centers are now the first target in any war. A graphics card in every home, doesn't give someone a concentrated target. National security makes large data centers extremely bad.
Large data centers are dramatically more efficient at utilizing both the power and the hardware.
At it's core, running a top-tier LLM means linear load of trillion+ weights from RAM for every token. The cost of this operation absolutely dwarfs the cost of doing the actual math. If you are only serving a single user, you can only amortize this over a handful of token streams, times like 2-4 for speculative decoding. If you are running in a datacenter, you can batch thousands of streams, and get correspondingly much more tokens out of the same expenditure of energy and hardware.
Be slightly careful with this. Large datacenters use power to operate the facility (this is measured as PUE > 1). Datacenter-oriented hardware often contains low-diameter, high static pressure fans that consume quite impressive amounts of power that are not counted as overhead in the PUE calculation.
A home- or office-oriented workstation will have much less power-hungry fans (go compare a Noctua datasheet to a server fan datasheet - the Noctua wipes the floor with a server fan if measured by CFM/W, and the internal temperature rise is mostly controlled by airflow and not by static pressure). And in some climates and times of year the marginal facility HVAC cost associated with running a computer may be zero or even negative.
And lower utilization hardware may run in on a more efficient point on the power-vs-performance curve. [0]
That being said, your overall point is correct.
[0] For a dramatic demonstration of how much this curve matters, you can run a non-CPU-saturating, non-latency-sensitive workload on various Intel CPUs with and without turbo on. Seeing a 2x or greater difference in reported power is quite possible. (Intel is terribly at writing platform firmware and the default behavior may be wildly suboptimal.)
None of that matters when there is literally a 100x multiplier on efficiency because of batching. It's all a rounding error.
In addition to the batching issue, "a graphics card" is not going to cut it for the multi-trillion parameter frontier models. You're looking at a 512GB Mac Studio to support anything close to those model sizes, and that's with heavy quantization and much worse performance.
agreed, gets some mid range hardware, some click installing agents with descent customisability, and im talking boring customisability like 'help me plan what to buy this month for my home' kind of tasks. Thats what we are going towards
There's also the fact that the US and China are in an AI arms race. So there's an inherent existential benefit to being ahead in AI regardless of where the financial feasibility/profitability discussion lands.
And at least currently, keeping up in that arms race means a maximal push on growth at the expense of profitability. Realistically, this will go until we hit some bottleneck.
> There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for.
Fair enough. But the real problem, it seems, is that people love slop.
My family is trying to sell a house in the South of France; the real estate people are sending us 500-words emails every so often, that are so obviously AI it hurts. Well written, and content free. Nobody cares. I'm the only one having issues with it.
At work, a new team let Claude do all their work and write all their status reports. Everybody seems to love it. It's so fast! And mostly correct! And so fast to correct when it's wrong! And here too, I'm the only one having issues with it.
A Canadian author wrote a book in French that turned out to be 100% AI ("C'était ça ou mourir"). Before that was revealed, every critic loved it, in Canada, in France and everywhere else (25 translations were started). The public loved it too: it's one of the top best sellers, and success continued even after the "scandal".
We are drowning in slop and nobody cares.
The quote is not a contradiction at all, it can both change the world and be low quality.
If that's true for every AI output is up for discussion, but it's reasonable that if someone is skeptical about the technology, they will also be skeptical about the quality.
That the tech is taking over many areas is not a stretch at all, that's just something that can be observed. Big companies (not just big tech) have been trying to get away with lowering (cost cutting) the quality of their products for a long time, AI lends itself to the same purpose, and will (and already is) obviously be used for it. Starts with customer communication as you mentioned, and customer service, and believe it or not I don't think every company should generate their own tech stack and software for their little products.
I think the problem is calling all AI output “slop”. If people genuinely like the content then is it still “slop”?
And, playing devils advocate for a moment, does it matter if a robot or a human authored it if people like it? People loved Magic Eye pictures in the 90s and they were computer generated too.
The problem is also with the creator. You learn nothing when AI produces things for you. When it's art, to each their own, I guess.
But when I ask my realtor questions, he doesn't know anything. I get the same answers by typing into ChatGPT.
When I ask the Claude team at work about why this or that, they don't have a clue. They "will get back to me", which means they'll forward the question to Claude and Claude will output another 500-words email with a lot of things that are obvious and not part of the question, another couple of assumptions that are plainly false, and in the middle of it, sometimes, yes, an interesting insight. But the insight doesn't come from the people sending the email, they didn't produce it, they don't see it, and they never act on it.
This is all just exhausting.
> But when I ask my realtor questions, he doesn't know anything. I get the same answers by typing into ChatGPT.
Not going to comment on your workplace specifically, but I’ve found this to be an issue with other industries like real estate long before ChatGPT were a thing.
The problem is the people who pickup the phone, greet you at the shop door, or take the photos, they’re the lowest paid juniors and thus know nothing.
So in most industries, you were always lucky to get through to someone with knowledge on the first attempt.
>someone with knowledge
The problem with AI is that "someone" may not exist.
That's highly unlikely because then the company wouldn't exist. At some point in a (to use our example: a real estate business) you are going to need skilled professionals.
I think the bigger risk is what happens when they retire if we are not training the next generation of professionals.
>At some point in a (to use our example: a real estate business) you are going to need skilled professionals.
This is exactly what we were saying with programming during early day of LLM. But that does not stop people from doing 100% vibe-coded projects.
Omarchy is one example, just ask the "developer" how stuff works in detail, without using AI, can they answer ?
> This is exactly what we were saying with programming during early day of LLM.
Yes I know. I was there too. ;)
> But that does not stop people from doing 100% vibe-coded projects.
And why should it? Let people "code" however they want to "code". Just like some people will find creative ways to play the guitar while others will just play covers.
By the way, you haven't addressed the core point posed: if people like AI output then is it still "slop" (in response to comments where people describe all AI output as "slop")
Well, people also like junk food, does it mean it's not junk food anymore ?
To me AI generated content is exactly like that. Cheap, fast and easy access.
It's fine for occasional consumption, and in some setting it's definitely not suitable to be served
Here are they contacts: https://omarchy.org/teams/
You can try asking.
High chance i will just talk to a meat proxy.
The only way for them to prove themself to the mass is a live coding interview session, without AI of course
> Behind it all is the Omarchy Doctrine: ten principles for...
That's ridiculous... Let's vibe code everything, but with a doctrine and principles!
If I'm going to do vibe coding, I don't need principles. No rules! And no future.
"slop" is less so about if something is a popular product or not, it's more about the quality. You can think that AI can produce high quality outputs, but it's unreasonable to think that this applies to the majority of it's use. There has always been corner cutting, and using AI to cut corners/safe money in product design/production inevitably results already, and will more so in the future, in a ton of bad products. That does not mean that there can't be good ones too.
> "slop" is less so about if something is a popular product or not, it's more about the quality.
But we are talking about popular culture here. That was never an industry where quality was measured by artistic excellence. It's an industry that is entirely motivated by mass appeal. Thus something being popular IS the measure of success.
So if AI is slop, then so is 99% of the human-made content it's replacing.
Slop exists because the cost of creating content has dropped to zero, which made the signal to noise ratio crater as well. Concision and density used to be a virtue. Now it's the opposite: content is measured by weight.
People absolutely cared about "C'était ça ou mourir", his publisher ended up dropping him. The reason the success continued is that people denied that he used AI and had all sorts of shaky political reasons for this (e.g. claiming that Haitian writing trips up AI detectors). Even the reason it got popular in the first place is to do with the themes of the novel rather than its writing.
> his publisher ended up dropping him
That's not really the point. Critics the world over LOVED the book; they only distanced themselves from it when it became clear it was AI, but when they didn't know, they couldn't stop raving about it.
because of its politics. There were plenty of people who thought it was dreck from the beginning.
most of the infrastructure will be garbage in 4 years. As before, it will become a monopoly of 1-2 big players and everyone else's datacenters will become ... crypto miners or something.