This was a completely vapid article. I expected some technical details at least, but instead it's an LLM generated poem confirming it worked in a short interval, which nobody doubts it would. All eyes are on long term heat dissipation and the impact of radiation and success of radiation hardening meterials and processes.
Putting AI in space is a silly idea. How do you cool it reliably? How do you handle rad-hardness? How do you even get the gear up there? It's one of the most expensive places to put mass.
Using geothermal power or floating solar arrays in the ocean makes way more sense. It's much easier to access, you have water for a heat sink, and you're far away from anyone or anything that would really care.
More controversially, if geothermal or solar doesn't pan out, you could power it via nuclear microreactors (just what I want floating in my world's ocean) or gas that would otherwise be flared off from a deep sea oil well (CO2 non-withstanding).
Satellite cooling is long figured out, the problem is that it requires a ton of expensive mass.
Radiation hardening is not a huge issue in LEO either. It is an issue, not the issue.
Overall, it's like choosing international flights as your choice of a restaurant. It's...fundamentally...doable...? But nobody knows why you would want that, unless flying GPUs is the point rather than to offer a competitive service using the server sats.
No, I mean, you can fly daily from NYC to London and back every single night for 5 years straight just to use the business class seats to sleep and to have inflight meals as your dinners and your breakfasts. That would be a 100% feasible solved problem from technical standpoint. But why would you do that.
They can probably build like, a gigantic satellite that fits in an expendable Starship that has like just couples of RTX5090 32GB backed by a Tesla battery pack and a gigantic multi panel actively refrigerating radiator system. Or use bunch of such Starships to assemble a whole rack worth of GPUs on orbit that costs billions total and minimum half a year to build out. Satellites and radiators are solved good enough for that.
Or you can put just the computer part on a desk indoor on Earth and plug it into the wall and have it cooled with fans. Or have a guy roll up an air cooled rack full of GPUs next to it. Why involve the rockets and satellites and all. What's the point. That's what everyone is puzzled with.
The public doesn't need to. If these companies actually try to build datacenters in space they will bankrupt themselves. It's a ridiculously stupid idea.
A chief reason to put orbital data centers in space is the free, limitless power from the Sun. Based on the schematic SpaceX released, each of its AI1 satellites would have solar panels encompassing about 600 square meters, or about 1.5 times the size of a basketball court. These solar panels would generate 150 kW of peak power and 120 kW of average power for computing.
---
It's "easy" (as far as launching a rocket is "easy") to put a satellite into space. The scale needed to replace a single data center is what's being questioned.
The proposals talk about tens of kilowatts for a satellite - a fraction of a rack.
Individual data centers get measured in gigawatts.
xAI has its Southaven data center under construction and estimated to consume 2.0 GW. If we assume the SpaceX planned satellites and the 120 kW of average power - that's 17,000 satellites to replace one data center.
... and that's assuming that you're putting all of them in the sun synchronous dawn dusk orbit so that you can get 100% solar availability. If you put in in a different orbit you're getting 50% sun and need some combination of more capacity (more satellites), more battery (more mass), and more solar (more mass to charge the battery to supply power).
Scott Manley's video is an excellent "can you" given the power for a single satellite based on Starlink in space. The answer is yes - you can put 10 kW of compute in space without too much difficulty.
Can you put 2 GW of compute in space to offset one data center on Earth?
The fact that this comment is being downvoted says so much about so many things. And so many people. This community--not just HN, I'm talking about tech at large--has totally lost its way.
A competing team at Google already exfiltrated the weights of the model running on the satellite and uploaded them to huggingface! https://huggingface.co/google/gemma-4-E4B
The standard motivation is 24/7 unfiltered sunlight to power the AI with, where you're not in anyone's backyard and don't need to convince a planning commission.
> The standard motivation is 24/7 unfiltered sunlight to power the AI with, where you're not in anyone's backyard and don't need to convince a planning commission.
The orbit that this is in doesn't meet that criteria. While it is in a near polar orbit, it is not in a dusk dawn orbit ( https://en.wikipedia.org/wiki/Sun-synchronous_orbit ). Half of its orbit is in shadow.
As I wrote this, it is above Campo Grande, Brazil. The time there is 10 pm.
You do, though. The FAA/AST govern launches. I'm no rocket scientist but building a rocket and launching it is probably much harder than building a small data center.
It’s not though. There’s existing rocket launches that fit within a well established federal regulatory framework. Building data centers has overlapping state and municipal regulations as well as local opposition.
It’s probably a lower regulatory burden to get the approval of the FAA than it is any number of municipal bodies to get approval for building and powering a new data center. There are a lot of people that need to sign off on things. When you’re talking about space, if you can afford and build it, that list is pretty short.
I’m not convinced that the economics are favorable to space based data centers, but ironically, it still might be a lower hurdle than terrestrial data center approval.
The FAA is federal, and may pre-empt state/local laws (see: CFTC, Kalshi/Polymarket/etc).
It's also pretty well accepted that the FAA is very cozy to SpaceX, and partially because NASA is basically dependent and beholden on SpaceX at the moment.
Making it difficult to shut down AI’s is very important. There’s a real chance we end up with a regulatory capture cartel that’s enforced because of “safety”.
But there are solid engineering and logistical advantages. Not having to deal with nimby’s, permitting, data center blowback is a big advantage. It’s difficult to spin up new power infrastructure in western countries, and putting up a satellite with direct 24/7 solar power means you can skip the headache of building new power infrastructure. My electricity bill has increased by 50% in the last year and most of that increase is likely because of data centers powering AI. If we can instead directly power that compute in orbit from the sun then it will not increase demand on existing power infrastructure.
Is this the piece of junk that can only run for 15min at a time before it overheats? They probably could have done some more analysis before launching it into space.. Like put the radiator in a vacuum chamber maybe? Or do some thermal modelling? Or did they know it would do this and launch it anyway? Would love to know why.
> The chips can operate for about 15 minutes in space before needing to be shut down so they can cool off, said Travis Beals, Google’s senior director of product management for Project Suncatcher.
Did they actually have to launch the thing to do that testing? I think a much simpler and more direct explanation is that it is simply a deceptive marketing stunt. "Hey look we're solving the power problem, don't worry we'll be able to build all these megawatts of compute we promised" when in reality it will never happen and they had no intention of even trying. It's a short term con to try to calm fears in the market that the promises exceed their capabilities.
That is a pretty cynical take. I think it is likely that a lunch would be appropriate for a legitimate development process. They did testing and vacuum and testing with high energy particles on Earth. The cost isn't great to validate those models with the real thing.
I think most people are overestimating the cost and difficulty of compute in space, and underestimating the advantages
What a load of horseshit. A model running on a ESP32 running off a questionable powerbank supply with dodgy coding and worse storage reliability can do that.
The only thing anyone cares about is whether the (economic)math works on put AI in space.
This was a completely vapid article. I expected some technical details at least, but instead it's an LLM generated poem confirming it worked in a short interval, which nobody doubts it would. All eyes are on long term heat dissipation and the impact of radiation and success of radiation hardening meterials and processes.
That was a month ago. https://news.ycombinator.com/item?id=49830606 This is just a status update, which the title is pretty clear about.
[flagged]
Putting AI in space is a silly idea. How do you cool it reliably? How do you handle rad-hardness? How do you even get the gear up there? It's one of the most expensive places to put mass.
Using geothermal power or floating solar arrays in the ocean makes way more sense. It's much easier to access, you have water for a heat sink, and you're far away from anyone or anything that would really care.
More controversially, if geothermal or solar doesn't pan out, you could power it via nuclear microreactors (just what I want floating in my world's ocean) or gas that would otherwise be flared off from a deep sea oil well (CO2 non-withstanding).
Satellite cooling is long figured out, the problem is that it requires a ton of expensive mass.
Radiation hardening is not a huge issue in LEO either. It is an issue, not the issue.
Overall, it's like choosing international flights as your choice of a restaurant. It's...fundamentally...doable...? But nobody knows why you would want that, unless flying GPUs is the point rather than to offer a competitive service using the server sats.
I think the cooling mass is so expensive and fragile that it’s basically an unsolved problem.
No, I mean, you can fly daily from NYC to London and back every single night for 5 years straight just to use the business class seats to sleep and to have inflight meals as your dinners and your breakfasts. That would be a 100% feasible solved problem from technical standpoint. But why would you do that.
They can probably build like, a gigantic satellite that fits in an expendable Starship that has like just couples of RTX5090 32GB backed by a Tesla battery pack and a gigantic multi panel actively refrigerating radiator system. Or use bunch of such Starships to assemble a whole rack worth of GPUs on orbit that costs billions total and minimum half a year to build out. Satellites and radiators are solved good enough for that.
Or you can put just the computer part on a desk indoor on Earth and plug it into the wall and have it cooled with fans. Or have a guy roll up an air cooled rack full of GPUs next to it. Why involve the rockets and satellites and all. What's the point. That's what everyone is puzzled with.
> Why involve the rockets and satellites and all. What's the point. That's what everyone is puzzled with.
Corporate sovereignty.
How dare those elected governments have demands and the means to enforce them?
Someone made a good point on a previous comment like this: the main selling point is the public can’t destroy them
The public doesn't need to. If these companies actually try to build datacenters in space they will bankrupt themselves. It's a ridiculously stupid idea.
That sounds like a fun challenge! DIY ASAT when?
So? Feels like a real xkcd.com/538 situation to me. An angry mob would be more of a threat to the people than the tech.
It is expensive, but the radiators won't necessarily push the state of the art. https://arstechnica.com/space/2026/07/how-hard-is-it-to-buil... The current test is to see how big an issue rad-hardness actually is.
Scott Manley has an excellent video breaking down the math on cooling.
It’s totally doable.
https://m.youtube.com/watch?v=FlQYU3m1e80
It is... he's working it off of a Starlink satellite getting 20 kW of power.
A traditional data center's rack is about 10 kW of power. A rack in an hyperscaler datacenter averages at 60 kW. (Source: https://www.nlyte.com/blog/data-center-rack-power-costs-a-co... )
While cooling is theoretically possible, being a third of a rack in a data center...
https://www.racksolutions.com/news/blog/how-many-servers-doe...
There are about 35,000 racks in a data center.
But also... that's not the scale that's being talked about.
https://arstechnica.com/space/2026/07/how-hard-is-it-to-buil...
---It's "easy" (as far as launching a rocket is "easy") to put a satellite into space. The scale needed to replace a single data center is what's being questioned.
The proposals talk about tens of kilowatts for a satellite - a fraction of a rack.
Individual data centers get measured in gigawatts.
xAI has its Southaven data center under construction and estimated to consume 2.0 GW. If we assume the SpaceX planned satellites and the 120 kW of average power - that's 17,000 satellites to replace one data center.
... and that's assuming that you're putting all of them in the sun synchronous dawn dusk orbit so that you can get 100% solar availability. If you put in in a different orbit you're getting 50% sun and need some combination of more capacity (more satellites), more battery (more mass), and more solar (more mass to charge the battery to supply power).
Scott Manley's video is an excellent "can you" given the power for a single satellite based on Starlink in space. The answer is yes - you can put 10 kW of compute in space without too much difficulty.
Can you put 2 GW of compute in space to offset one data center on Earth?
The fact that this comment is being downvoted says so much about so many things. And so many people. This community--not just HN, I'm talking about tech at large--has totally lost its way.
So their weights are just floating up there?
I suppose it’s beyond hard anyone would hijack their satellite but still. Wonder if they put any physical security on it anyway.
A competing team at Google already exfiltrated the weights of the model running on the satellite and uploaded them to huggingface! https://huggingface.co/google/gemma-4-E4B
Sarcasm?
You’re saying they just deployed an open model so no security concerns for this at the moment, yes?
I still can't think of a use case for this besides making it difficult to shut down illegal AIs.
The standard motivation is 24/7 unfiltered sunlight to power the AI with, where you're not in anyone's backyard and don't need to convince a planning commission.
> The standard motivation is 24/7 unfiltered sunlight to power the AI with, where you're not in anyone's backyard and don't need to convince a planning commission.
The orbit that this is in doesn't meet that criteria. While it is in a near polar orbit, it is not in a dusk dawn orbit ( https://en.wikipedia.org/wiki/Sun-synchronous_orbit ). Half of its orbit is in shadow.
As I wrote this, it is above Campo Grande, Brazil. The time there is 10 pm.
https://satellitemap.space/sat/100902
You do, though. The FAA/AST govern launches. I'm no rocket scientist but building a rocket and launching it is probably much harder than building a small data center.
It’s not though. There’s existing rocket launches that fit within a well established federal regulatory framework. Building data centers has overlapping state and municipal regulations as well as local opposition.
It’s probably a lower regulatory burden to get the approval of the FAA than it is any number of municipal bodies to get approval for building and powering a new data center. There are a lot of people that need to sign off on things. When you’re talking about space, if you can afford and build it, that list is pretty short.
I’m not convinced that the economics are favorable to space based data centers, but ironically, it still might be a lower hurdle than terrestrial data center approval.
The FAA is federal, and may pre-empt state/local laws (see: CFTC, Kalshi/Polymarket/etc).
It's also pretty well accepted that the FAA is very cozy to SpaceX, and partially because NASA is basically dependent and beholden on SpaceX at the moment.
You don't have to build your own rocket anymore. SpaceX and others sell launches at fixed prices.
Making it difficult to shut down AI’s is very important. There’s a real chance we end up with a regulatory capture cartel that’s enforced because of “safety”.
But there are solid engineering and logistical advantages. Not having to deal with nimby’s, permitting, data center blowback is a big advantage. It’s difficult to spin up new power infrastructure in western countries, and putting up a satellite with direct 24/7 solar power means you can skip the headache of building new power infrastructure. My electricity bill has increased by 50% in the last year and most of that increase is likely because of data centers powering AI. If we can instead directly power that compute in orbit from the sun then it will not increase demand on existing power infrastructure.
Unlimited power and real estate with zero local opposition.
None of those things are true.
How so? Space has no locals, and the sun never sets up there.
In orbit around Earth it does! In LEO it’s like 45 minutes of day and 45 minutes of night IIRC
There are Sun Synchronous orbits that will be used eventually
Or just go further out. A box doing training breaks even faster even with the lift penalty.
They are if you can get launch costs cheap enough. I hope that is possible but I don't think that it is.
The entire concept of an illegal AI is something I hope never comes to pass.
AIs are already doing illegal things
AI doing something illegal and AI being illegal are very different things
Well, that's what I meant primarily. An AI that is doing something damaging.
No, people using AI are doing illegal things with it.
Did they finally figure out, while possible, it is extremely impractical?
[dead]
Google taking a lesson from the marketing school of Musk? Public spectacle to boost stock price?
Is this the piece of junk that can only run for 15min at a time before it overheats? They probably could have done some more analysis before launching it into space.. Like put the radiator in a vacuum chamber maybe? Or do some thermal modelling? Or did they know it would do this and launch it anyway? Would love to know why.
Could you provide a source for that claim?
Why yes, I sure can!
> The chips can operate for about 15 minutes in space before needing to be shut down so they can cool off, said Travis Beals, Google’s senior director of product management for Project Suncatcher.
https://www.nytimes.com/2026/09/24/technology/google-suncatc...
Thanks!
The point is not to run longer than 15 min. They knew this before even launching it.
Part of development is prototyping and testing. This is part of a process to learn and inform the design of later versions.
We would have very little science or technology if we did not do such tests.
Did they actually have to launch the thing to do that testing? I think a much simpler and more direct explanation is that it is simply a deceptive marketing stunt. "Hey look we're solving the power problem, don't worry we'll be able to build all these megawatts of compute we promised" when in reality it will never happen and they had no intention of even trying. It's a short term con to try to calm fears in the market that the promises exceed their capabilities.
That is a pretty cynical take. I think it is likely that a lunch would be appropriate for a legitimate development process. They did testing and vacuum and testing with high energy particles on Earth. The cost isn't great to validate those models with the real thing.
I think most people are overestimating the cost and difficulty of compute in space, and underestimating the advantages
> poem confirmed success:
What a load of horseshit. A model running on a ESP32 running off a questionable powerbank supply with dodgy coding and worse storage reliability can do that.
The only thing anyone cares about is whether the (economic)math works on put AI in space.
>The only thing anyone cares about is whether the (economic)math works on put AI in space.
This is a greater fool market. The only thing anyone cares about is whether they think someone else will believe the math will work out.