So like I don't get to use my degree except to tutor kids and answer physics questions and the like, so indulge me for a second because this doesn't make much sense to me. So like the endgame is to create a new Pryor, OK facility in space, right? It doesn't make sense if we're not shipping something that's about the same as what you can build on Earth.
Epoch.AI says that this facility has 100,000 TPU chips, eats 370 megawatts of electricity, but designed to get rid of up to 480 megawatts of heat. Sun power is 1.36 kW/m^2, which is something like 7 megawatts per football field (assuming ~5000 m^2), but you have to put up something like 3-5x that because solar panels are only 20-30% efficient, right? Assume that you can arrange it so that the solar panels are one side, the radiators on the other, you can maybe get away with only 300-400 football fields maybe? So like 1.5 million m^2 or 1.2 km wide, right? At LEO, 650 km, you have (1.2km)/(650 km) * 180 degrees/pi is 0.11 degrees or 6.6 arcminutes of visual size. The moon is only like 30 arcminutes. And the paper talks about how they're going to not do one big monolithic construction but an oval of fridge-sized objects separated out 2-3 times this size -- so like I don't see how you get another Pryor, OK size data center in LEO without basically having it look like a second, smaller moon flying across the sky 10 times per day.
And if this were a wildly successful idea are we talking about having like 5-10 of them, a few for each big frontier lab? This just sounds like we're talking about the most profound shift to our night skies since we started having to deal with light pollution in our cities.
There's a lot of different designs besides this one.
So far as I can work out, literally all of the plans are bad. The "why" varies, but they're all bad.
I'm too tired to double check your maths, so I will assume correct: one likely difference even for this plan is a terminator following sun-synchronous orbit, which means you'll only see it twice a day despite the orbital period being about 90-100 minutes, and when you see it will be specifically at sunrise and sunset.
Visibility is also a question of reflection, not just size. Terminator following orbits are worse than normal satellites, because one of the tricks for reducing e.g. Starlink visibility is to tilt them as they cross the terminator and you can't do that if they're always on the terminator.
The SpaceX plans (a million small ones) becomes a glitter band in some parts of the sky and will appear visually contiguous in other parts, though I need to double check my maths and assumptions about visibility given this happens during sunrise and sunset so the sky itself is pretty bright.
I think there are at least a few problems with the math’s assumptions:
1) Don’t the radiators need to have more surface area than the solar? (Unless the chips run very hot.)
2) The datacenter will have the pesky earth between it and the sun some fraction of the time, and have to either shut down or run off batteries. ~50%, assuming LEO, right? If you leave LEO, then the latency sucks, so they’re training-only clusters. At 50% the solar doubles and you need 370 megawatt hours per hour of darkness, or you run the machines 50% of the time, rebooting for each orbit. If you make the orbit shorter (so you can have smaller batteries), then they wear out faster. The batteries also emit heat. Plus, you need to double the solar so they charge while the workload is running.
3) How do they cope with cosmic rays? The standard approach is still to duplicate or triplicate all computation, or use larger/slower processes, right?
The obvious answer to each question makes the engineering design at least twice as dumb, and they stack. There are many other problems like these.
Retail cost per watt for terrestrial panel: under 45c
Manufacturing cost per watt of space grade solar panel: up to $450
Annual performance degradation of terrestrial solar panel: under 0.5%
That of space-grade solar panel: up to 2%
Life span of terrestrial panel: about 2x that of space panel.
Total difference in cost per watt feeding a DC load in space vs on land: about x400
And that's ignoring launch costs. It makes absolutely zero sense. And given the scale of production and investment in manufacturing, terrestrial is likely to stretch even further ahead in the cost stakes.
The idea is still a real stretch, but according to Google's paper in Joule [1] they're putting satellites in low Earth orbit and only targeting a service lifetime of 5 years. Over that short time, with the correspondingly modest radiation exposure, you can use inexpensive silicon solar cells like you would use for terrestrial solar farms. The expensive space grade solar cells from e.g. Boeing Spectrolab are more resistant to radiation, and achieve higher conversion efficiencies, but silicon cells are fine for satellites like these.
Starlink satellites already use silicon solar cells, since they too are cost sensitive and don't have long lifespans:
On the Starlink V2 mini satellite, we predict that approximately 5% of the mass of the entire satellite could survive reentry. The biggest contributor (~90% of the surviving mass) is silicon from the solar cells, which has a high melting point...
Best cost to orbit I am seeing is $1500/kg. A GPU rack is ~1500kg. Let’s imagine you can take a terrestrial data center rack, no scaffolding, solar panels, radiators, radios, propellant, or propulsion. Fly it into orbit, kick it out the airlock, and let it work through magic. That’s $2.2 million to get into position.
Industrial power rates are cheap, say $.10/kwh, but pretend you sign terrible deals, and it costs you $.30/kwh to run and cool a terrestrial GPU rack. A 150kw unit will then be (150x24x365x.3) =$394k/year.
You can operate the terrestrial version for 5.7 years before the two hit parity.
Yes for now; Alphabet is optimistic costs will come down.
They're projecting the learning curve, that the more you do it the cheaper it gets, continues arbitrarily far.
However, they recon it will take SpaceX launching 370,000 tons to LEO to make the costs come down enough to be worth it: https://arxiv.org/pdf/2511.19468
Even my bull case put that 10 years off, which is so far away it lacks relevance just because tech moves so much faster than that timescale; my bear case says that's about 45 years off.
Even if getting to orbit is free, the numbers don’t work.
5 years is a standard depreciation cycle for hardware, but plenty of machines last longer than that, but these satellites will not. On top of that, power distribution, cooling etc in the data centers is a huge percentage of cost and lasts much longer than 5 years.
If you think you can make a space air conditioner for less than a terrestrial one, I have a bridge to sell you.
Smells like a cover story for military AI module validation testing.
The AI industry spends crazy money to reduce latency and increase density. The act of placing resources in space, to be used by people on earth, only increases latency, and reduces density. The performance metrics are ops/power/space/$, not G ratings, unless its being validated for flight.
What would a radar constellation look like that could see everything on earth to a metre or so resolution? Thousands and thousands of satellites all with big phased array radios and solar panels and batteries.
What do you need to distill a 24/7 stream of everything on Earth in roughly 1m^3 voxels into data like humans and vehicles? Bearing in mind the data torrent is gigantic and can't always be reliably downlinked: co-orbital AI.
Terrestrial data centers are having a hard time connecting to power. What’s the good of a $40/Watt compute asset you can’t turn on? Spending $10 or even $20/Watt extra to get power by sending inference load to orbit starts to sound like a bargain, and that’s doable with today’s launch vehicles. With high reuse, that number drops to below $1/Watt for 24/7 power, which makes it cheaper than any terrestrial energy source.
And the vast majority of compute load is inference. Part of the training process benefits from a large, coherent (ie low latency) data center, but that’s a minority of the load nowadays.
Also: low Earth orbit is already pretty low latency. A lot of people’s intuitions about space latency are from when satellite internet was based out of distant geosynchronous satellites, but even that is probably just fine for most inference workloads.
Maybe. But it’s also Google, a company famous for just spending money on random crap. Data centers are politically toxic right now, and Elon Musk’s bloviating made it trendy to discuss AI in space. Maybe they think that following this trend with a relatively simple launch will raise the stock.
I think it’s somewhat impressive they got it into orbit quickly since surely a year ago this wasn’t an earnest idea, and I think it’ll be interesting to see the affects of space on high end chips and connectivity fabric.
But also I agree that this just doesn’t make sense from a performance standpoint.
It's not really extraterritorial from the point of view of needing to comply with regulations though. Your datacentre company has registered offices and physical locations. Your downlinks are based on territory. You need to relaunch on a regular basis and you have very restricted choice of territories that you can launch from.
There's one other thing: it's essentially like the cloud but for DCs. i.e. need more compute? send up another rocket. 'limitless' space. expensive but a relatively fixed one-time cost (vs earth-based DCs which are the bare metal equivalent: fast, affordable, but pets not cattle).
All of that said, I'm skeptical we'll see this in a big way any time soon.
When you add in the VRT and the QML (even not including the ZB-PRP 5) you get far lower than what a naive USP calculation gets you. Horner Weisenbach talked about this recently on their blog.
These endless conversations about how it can never work are so boring. The only difference between scenarios where it will and won't work are (often unstated) assumptions about launch cost. Launch cost is the whole ball game, as any problem can be solved with more mass. So the whole discussion can be summarized as "I think Starship will fail to meet its launch cost targets" vs "I think Starship will work". It's fashionable to bet against Elon these days but historically it has been a bad bet on long timeframes...
If launch costs are assumed to be $0/kg, how does that solve the waste heat problem? Add more mass? What about the maintenance of failing hardware?
I'm genuinely curious, I thought basically heat radiation was a hard physical boundary, regardless of launch cost.
Arbitrarily large radiators can dispose of arbitrarily large amounts of heat. In practice the radiators don't even need to be as large as the solar panels.
If launch cost was truly $0 you could also do maintenance in space but I think it's more likely that you'd work to push up reliability instead. That's what SpaceX did with Starlink.
More material for the eventual investor fraud/scam lawsuits.
They did not even attempt to solve the main issues: energy and heat management. The GPUs can work only for 15 minutes and are cooled by a phase change material that then slowly sheds the heat.
Cool, we will finally have actual data to see if that’s possible or not. My understanding is that doing compute in space for ML is pretty absurd, but the discussion will be way more interesting with real numbers.
I still think it’s mostly a way to put pressure on politicians, « you either give us tax cut, fast tracked permits, and let us do what we want with the land or we go to space and you get no benefits »
Doing compute in space for ML is perfectly possible (i.e. https://en.wikipedia.org/wiki/Phi-Sat-2), making it competitive with terrestrial compute is a much harder question to solve.
Starlink is heavily regulated. Or at least as heavily as anyone else using radio spectrum
Not sure what interest the government would have in stricter regulation of the rocket launching side of SpaceX. Environmental impact and impact on air traffic and boats is regulated. And SpaceX existing is strongly benefiting the US government. They are their biggest customer
Is it possible that our current government is particularly prone to under regulation of Elon Musks companies and that a different administration may in fact behave differently?
This is obviously for the Golden Dome. Missile interception for ICBMs only works in space and they need the hardware and AI for classifiers (decoy/not_decoy) and for targeting in space, too.
For civilian purposes this would be even more dystopian. Communication satellites and propaganda from space have always existed, but AI manufactured mass slop and propaganda from space really raises philosophical objections.
When did they do this? I didn’t hear of any SpaceX launches besides the falcon launch yesterday for a government spy satellite. I wonder if this one was also bundled in.
I don't know what that Arianespace figure is referring to (perhaps just the commercial satellite market, or the portion of it they expect to capture?) but there were 317 successful launches to orbit in 2025. Last April there were six in 24 hours: https://www.space.com/space-exploration/launches-spacecraft/...
Three launches from one company is unusual though.
If US folks are watching a rocket launch and they're wearing light jackets, it's Vandenberg (which doesn't publish all their scheduled rocket launches.)
So like I don't get to use my degree except to tutor kids and answer physics questions and the like, so indulge me for a second because this doesn't make much sense to me. So like the endgame is to create a new Pryor, OK facility in space, right? It doesn't make sense if we're not shipping something that's about the same as what you can build on Earth.
Epoch.AI says that this facility has 100,000 TPU chips, eats 370 megawatts of electricity, but designed to get rid of up to 480 megawatts of heat. Sun power is 1.36 kW/m^2, which is something like 7 megawatts per football field (assuming ~5000 m^2), but you have to put up something like 3-5x that because solar panels are only 20-30% efficient, right? Assume that you can arrange it so that the solar panels are one side, the radiators on the other, you can maybe get away with only 300-400 football fields maybe? So like 1.5 million m^2 or 1.2 km wide, right? At LEO, 650 km, you have (1.2km)/(650 km) * 180 degrees/pi is 0.11 degrees or 6.6 arcminutes of visual size. The moon is only like 30 arcminutes. And the paper talks about how they're going to not do one big monolithic construction but an oval of fridge-sized objects separated out 2-3 times this size -- so like I don't see how you get another Pryor, OK size data center in LEO without basically having it look like a second, smaller moon flying across the sky 10 times per day.
And if this were a wildly successful idea are we talking about having like 5-10 of them, a few for each big frontier lab? This just sounds like we're talking about the most profound shift to our night skies since we started having to deal with light pollution in our cities.
There's a lot of different designs besides this one.
So far as I can work out, literally all of the plans are bad. The "why" varies, but they're all bad.
I'm too tired to double check your maths, so I will assume correct: one likely difference even for this plan is a terminator following sun-synchronous orbit, which means you'll only see it twice a day despite the orbital period being about 90-100 minutes, and when you see it will be specifically at sunrise and sunset.
Visibility is also a question of reflection, not just size. Terminator following orbits are worse than normal satellites, because one of the tricks for reducing e.g. Starlink visibility is to tilt them as they cross the terminator and you can't do that if they're always on the terminator.
The SpaceX plans (a million small ones) becomes a glitter band in some parts of the sky and will appear visually contiguous in other parts, though I need to double check my maths and assumptions about visibility given this happens during sunrise and sunset so the sky itself is pretty bright.
I think there are at least a few problems with the math’s assumptions:
1) Don’t the radiators need to have more surface area than the solar? (Unless the chips run very hot.)
2) The datacenter will have the pesky earth between it and the sun some fraction of the time, and have to either shut down or run off batteries. ~50%, assuming LEO, right? If you leave LEO, then the latency sucks, so they’re training-only clusters. At 50% the solar doubles and you need 370 megawatt hours per hour of darkness, or you run the machines 50% of the time, rebooting for each orbit. If you make the orbit shorter (so you can have smaller batteries), then they wear out faster. The batteries also emit heat. Plus, you need to double the solar so they charge while the workload is running.
3) How do they cope with cosmic rays? The standard approach is still to duplicate or triplicate all computation, or use larger/slower processes, right?
The obvious answer to each question makes the engineering design at least twice as dumb, and they stack. There are many other problems like these.
Capacity factor of orbital solar PV panel: 97%
Capacity factor of terrestrial solar panel: 23%
Retail cost per watt for terrestrial panel: under 45c
Manufacturing cost per watt of space grade solar panel: up to $450
Annual performance degradation of terrestrial solar panel: under 0.5%
That of space-grade solar panel: up to 2%
Life span of terrestrial panel: about 2x that of space panel.
Total difference in cost per watt feeding a DC load in space vs on land: about x400
And that's ignoring launch costs. It makes absolutely zero sense. And given the scale of production and investment in manufacturing, terrestrial is likely to stretch even further ahead in the cost stakes.
The idea is still a real stretch, but according to Google's paper in Joule [1] they're putting satellites in low Earth orbit and only targeting a service lifetime of 5 years. Over that short time, with the correspondingly modest radiation exposure, you can use inexpensive silicon solar cells like you would use for terrestrial solar farms. The expensive space grade solar cells from e.g. Boeing Spectrolab are more resistant to radiation, and achieve higher conversion efficiencies, but silicon cells are fine for satellites like these.
Starlink satellites already use silicon solar cells, since they too are cost sensitive and don't have long lifespans:
https://starlink.com/public-files/Starlink_Approach_to_Satel...
On the Starlink V2 mini satellite, we predict that approximately 5% of the mass of the entire satellite could survive reentry. The biggest contributor (~90% of the surviving mass) is silicon from the solar cells, which has a high melting point...
[1] https://www.cell.com/joule/fulltext/S2542-4351(26)00362-4
Just on a mass basis it falls apart.
Best cost to orbit I am seeing is $1500/kg. A GPU rack is ~1500kg. Let’s imagine you can take a terrestrial data center rack, no scaffolding, solar panels, radiators, radios, propellant, or propulsion. Fly it into orbit, kick it out the airlock, and let it work through magic. That’s $2.2 million to get into position.
Industrial power rates are cheap, say $.10/kwh, but pretend you sign terrible deals, and it costs you $.30/kwh to run and cool a terrestrial GPU rack. A 150kw unit will then be (150x24x365x.3) =$394k/year.
You can operate the terrestrial version for 5.7 years before the two hit parity.
Yes for now; Alphabet is optimistic costs will come down.
They're projecting the learning curve, that the more you do it the cheaper it gets, continues arbitrarily far.
However, they recon it will take SpaceX launching 370,000 tons to LEO to make the costs come down enough to be worth it: https://arxiv.org/pdf/2511.19468
Even my bull case put that 10 years off, which is so far away it lacks relevance just because tech moves so much faster than that timescale; my bear case says that's about 45 years off.
Even if getting to orbit is free, the numbers don’t work.
5 years is a standard depreciation cycle for hardware, but plenty of machines last longer than that, but these satellites will not. On top of that, power distribution, cooling etc in the data centers is a huge percentage of cost and lasts much longer than 5 years.
If you think you can make a space air conditioner for less than a terrestrial one, I have a bridge to sell you.
$1500 / kg seems incredibly low considering that launching a simple 1U cube sat costs roughly $200,000.
Recent and related:
Google's first Suncatcher orbital data center test launches October 1 - https://news.ycombinator.com/item?id=49837350 - Sept 2026 (96 comments)
Google’s Project Suncatcher to put ML infrastructure in space - https://news.ycombinator.com/item?id=49830606 - Sept 2026 (551 comments)
Smells like a cover story for military AI module validation testing.
The AI industry spends crazy money to reduce latency and increase density. The act of placing resources in space, to be used by people on earth, only increases latency, and reduces density. The performance metrics are ops/power/space/$, not G ratings, unless its being validated for flight.
What would a radar constellation look like that could see everything on earth to a metre or so resolution? Thousands and thousands of satellites all with big phased array radios and solar panels and batteries.
What do you need to distill a 24/7 stream of everything on Earth in roughly 1m^3 voxels into data like humans and vehicles? Bearing in mind the data torrent is gigantic and can't always be reliably downlinked: co-orbital AI.
Terrestrial data centers are having a hard time connecting to power. What’s the good of a $40/Watt compute asset you can’t turn on? Spending $10 or even $20/Watt extra to get power by sending inference load to orbit starts to sound like a bargain, and that’s doable with today’s launch vehicles. With high reuse, that number drops to below $1/Watt for 24/7 power, which makes it cheaper than any terrestrial energy source. And the vast majority of compute load is inference. Part of the training process benefits from a large, coherent (ie low latency) data center, but that’s a minority of the load nowadays.
Also: low Earth orbit is already pretty low latency. A lot of people’s intuitions about space latency are from when satellite internet was based out of distant geosynchronous satellites, but even that is probably just fine for most inference workloads.
Talking starmind.
150kw design Launch on 4000kg is 4 million. 150kw solar Radiator Electronics
20/watt is 3 million. We are above your budget.
Maybe. But it’s also Google, a company famous for just spending money on random crap. Data centers are politically toxic right now, and Elon Musk’s bloviating made it trendy to discuss AI in space. Maybe they think that following this trend with a relatively simple launch will raise the stock.
I think it’s somewhat impressive they got it into orbit quickly since surely a year ago this wasn’t an earnest idea, and I think it’ll be interesting to see the affects of space on high end chips and connectivity fabric.
But also I agree that this just doesn’t make sense from a performance standpoint.
The usp of data centres in space is extraterritoriality. That's all it is.
It's not really extraterritorial from the point of view of needing to comply with regulations though. Your datacentre company has registered offices and physical locations. Your downlinks are based on territory. You need to relaunch on a regular basis and you have very restricted choice of territories that you can launch from.
There's one other thing: it's essentially like the cloud but for DCs. i.e. need more compute? send up another rocket. 'limitless' space. expensive but a relatively fixed one-time cost (vs earth-based DCs which are the bare metal equivalent: fast, affordable, but pets not cattle).
All of that said, I'm skeptical we'll see this in a big way any time soon.
I think it's far more than the USP.
When you add in the VRT and the QML (even not including the ZB-PRP 5) you get far lower than what a naive USP calculation gets you. Horner Weisenbach talked about this recently on their blog.
If you're extraterritorial, who is protecting you from ASATs?
That and being able to pretend that they’re doing something noteworthy instead of treading water and burning funds.
These endless conversations about how it can never work are so boring. The only difference between scenarios where it will and won't work are (often unstated) assumptions about launch cost. Launch cost is the whole ball game, as any problem can be solved with more mass. So the whole discussion can be summarized as "I think Starship will fail to meet its launch cost targets" vs "I think Starship will work". It's fashionable to bet against Elon these days but historically it has been a bad bet on long timeframes...
If launch costs are assumed to be $0/kg, how does that solve the waste heat problem? Add more mass? What about the maintenance of failing hardware? I'm genuinely curious, I thought basically heat radiation was a hard physical boundary, regardless of launch cost.
Arbitrarily large radiators can dispose of arbitrarily large amounts of heat. In practice the radiators don't even need to be as large as the solar panels.
If launch cost was truly $0 you could also do maintenance in space but I think it's more likely that you'd work to push up reliability instead. That's what SpaceX did with Starlink.
How do you move the earth out of the way so it does not shade the solar panels?
Sun-synchronous orbit.
Four short videos (~10 minutes total) about Suncatcher:
https://www.youtube.com/playlist?list=PLAeFjVYtuU34
They couldn't stabilize their footage ? My Google phone's videos aren't shaky like that.
More material for the eventual investor fraud/scam lawsuits.
They did not even attempt to solve the main issues: energy and heat management. The GPUs can work only for 15 minutes and are cooled by a phase change material that then slowly sheds the heat.
Cool, we will finally have actual data to see if that’s possible or not. My understanding is that doing compute in space for ML is pretty absurd, but the discussion will be way more interesting with real numbers.
I still think it’s mostly a way to put pressure on politicians, « you either give us tax cut, fast tracked permits, and let us do what we want with the land or we go to space and you get no benefits »
Doing compute in space for ML is perfectly possible (i.e. https://en.wikipedia.org/wiki/Phi-Sat-2), making it competitive with terrestrial compute is a much harder question to solve.
What makes you think governments can't or won't regulate organizations operating in space?
The fact that SpaceX + Starlink exists and wasn't regulated in to oblivion shows the US government will mostly stay away.
Starlink is heavily regulated. Or at least as heavily as anyone else using radio spectrum
Not sure what interest the government would have in stricter regulation of the rocket launching side of SpaceX. Environmental impact and impact on air traffic and boats is regulated. And SpaceX existing is strongly benefiting the US government. They are their biggest customer
Is it possible that our current government is particularly prone to under regulation of Elon Musks companies and that a different administration may in fact behave differently?
This is obviously for the Golden Dome. Missile interception for ICBMs only works in space and they need the hardware and AI for classifiers (decoy/not_decoy) and for targeting in space, too.
For civilian purposes this would be even more dystopian. Communication satellites and propaganda from space have always existed, but AI manufactured mass slop and propaganda from space really raises philosophical objections.
When did they do this? I didn’t hear of any SpaceX launches besides the falcon launch yesterday for a government spy satellite. I wonder if this one was also bundled in.
There were three SpaceX launches yesterday. (An ISS crew, a Falcon Heavy spy satellite, and a big rideshare that included this).
There is a global market of only about 25 rocket launches a year (according to Arianne aerospace) and SpaceX launched 3 in a single day?
I don't know what that Arianespace figure is referring to (perhaps just the commercial satellite market, or the portion of it they expect to capture?) but there were 317 successful launches to orbit in 2025. Last April there were six in 24 hours: https://www.space.com/space-exploration/launches-spacecraft/...
Three launches from one company is unusual though.
There were 317 successful orbital launches in 2025 [0] and 229 so far this year [1]...
[0]: https://en.wikipedia.org/wiki/2025_in_spaceflight#Orbital_la...
[1]: https://en.wikipedia.org/wiki/2026_in_spaceflight#Orbital_la...
Edit: mistyped 317 as 313
Those were the Before Times. Now SpaceX is sending up sky clutter all the damn time.
If US folks are watching a rocket launch and they're wearing light jackets, it's Vandenberg (which doesn't publish all their scheduled rocket launches.)
https://www.spacex.com/launches/transporter18
In other news, Howard Hughes explored the deep sea for purely civilian purposes ...