So many ideas involving AI just seems to be built off of sci-fi (not in a good way), including this one. Like sci-fi, there are little practical considerations made.
I asked Google for more information about AI datacenter in space. This was the first sentence, 'AI data centers are being developed in space to handle the massive energy demands of AI, using solar power and the vacuum of space for cooling.'
> After laughing at "the vacuum of space for cooling" I closed the page because there was nothing serious there. Basic high school physics student would be laughing at that sentence.
As someone with a similar background to the writer of this post (I did avionics work for NASA before moving into more “traditional” software engineering), this post does a great job at summing up my thoughts on why space-based data centers won’t work. The SEU issues were my first though followed by the thermal concerns, and both are addressed here fantastically.
On the SEU issue I’ll add in that even in LEO you can still get SEUs - the ISS is in LEO and gets SEUs on occasion. There’s also the South Atlantic Anomaly where spacecraft in LEO see a higher number of SEUs.
Single event upsets are already commonplace at sea level well below data center scale.
The section of the article that talks about them isn’t great. At least for FPGAs, the state of the art is to run 2-3 copies of the logic, and detect output discrepancies before they can create side effects.
I guess you could build a GPU that way, but it’d have 1/3 the parallelism as a normal one for the same die size and power budget. The article says it’d be a 2-3 order of magnitude loss.
> On the SEU issue I’ll add in that even in LEO you can still get SEUs
As a sibling post noted, SEUs are possible all the way down to sea level. The recent Airbus mass intervention was essentially a fix for a badly handled SEU in a corner case.
What about on the Moon? My understanding is that heat is the killer. There you could sink pipes into the surface and use that as a heat sink. There are “peaks of eternal light” near the poles where you could get 24/7 solar power.
Latency becomes high but you send large batches of work.
Probably not at all economical compared to anywhere on Earth but the physics work better than orbit where you need giant heat sinks.
It's not a viable heat sink because it's a thermal insulator that doesn't support transport of heat. The thermal conductivity of lunar regolith is lower than rock-wool insulation,
(Imagine, for entertainment purposes, what would happen if you wrapped a running server rack in a giant ball of rock-wool insulation, 50 meters in radius).
Only way to dissipate large amounts of heat on the moon is with sky-facing radiators.
Regardless of how terrible an idea it is, I wouldn't mind some billionaires funding R&D that advances the state of the art in thermal management in space.
The one thing that space has going for itself is space. You could have way bigger datacenters than on Earth and just leave them there, assuming Starship makes it cheap enough to get them there. I think it would maybe make sense if 2 things:
- We are sure we will need a lot of gpus for the next 30-40 years.
- We can make the solar panels + cooling + GPUs have a great life expectancy, so that we can just leave them up there and accumulate them.
Latency wise it seems okay for llm training to put them higher than Starlink to make them last longer and avoid decelerating because of the atmosphere. And for inference, well, if the infra can be amortized over decades than it might make the inference price cheap enough to endure additional latencies.
Concerning communication, SpaceX I think already has inter-starlinks laser comms, at least a prototype.
Starship is on a fast track to failure. It is not a cheaper way to get to orbit and will never get there at the current pace. And even if it were, it would not make getting to orbit so cheap that it would somehow make it economically viable to put a datacenter there.
You still have to build the GPUs, etc for the datacenter whether it’s on Earth or in orbit. But to put it in space you also need massive new cooling solution, radiation shielding, orbital boosting, data transmission bandwidth, and you have to launch all of that.
And then, there are zero benefits to putting a datacenter in space over building it on Earth. So why would you want to add all that extra expense?
I don't agree with the logic that "something is hard/can't be done right now" is equivalent to "this is a terrible idea and won't work."
There are dozens of companies solving each problem outlined here; if we never attempt the 'hard' thing we will never progress. The author could have easily taken a tone of 'these are all the things that are hard that we will need to solve first' but actively chose to take the 'catastrophically bad idea' angle.
From a more positive angle, I'm a big fan of Northwood Space and they're tackling the 'Communications' problem outlined in this article pretty well.
What reason is there to build datacenters in space, though? Literally, what limitation do we face in building datacenters on Earth would building them in space improve?
It's not that it's hard, it's that it's stupid - it's based on a misunderstanding of the physics involved which completely negates any of the benefits.
It's the opposite of engineering, where you understand a problem space and then try to determine the optimal solution given the constraints. This starts with an assumption that the solution is correct, and then tries to engineer fixes to gaps in the solution, without ever reevaluating the solution choice.
Unless thermodynamics suddenly changes, how exactly is the cooling problem being solved? Yeeting hot chunks of matter out the back? On a planetary body you have an entire massive system of matter to reject your heat into. In space, you have nothing.
There are things which are difficult and have unsolved problems, and there are things that just fundamentally make no sense.
Nobody is proposing data centers at the South Pole. This isn’t because it’s difficult. It is difficult, but that’s not the reason it’s not being looked at. Nobody’s doing it because it’s pointless. It’s a massive hassle for very little gain. It’s never going to be worth the cost no matter what problems get solved.
Data centers in space are like that. It’s not that it’s difficult. It’s that the downsides are fundamentally much worse than the advantages, because the advantages aren’t very significant. Ok, you get somewhat more consistent solar power and you can reach a wider ground area by radio or laser. And in exchange for that, you get to deal with cooling in a near perfect insulator, a significantly increased radiation environment, and difficult-to-impossible maintenance. Those challenges can be overcome, sure, but why?
This whole thing makes no sense. Maybe there’s something we just aren’t seeing, or maybe this is what happens when people are able to accumulate far too much money and nobody is willing to tell them they’re being stupid.
Always remember the magic words: dual use technology. The people pushing these aren't saying to you that they want to build data centers in space because conventional data centers are at huge risk of getting bombed by foreign nations or eventually getting smashed by angry mobs. But you can bet they're saying that to the people with the dual-use technology money bag. Or even better, let them draw that conclusion themselves, to make them think it was their idea - that also has the advantage of deniability when it turns out data centers in space was a terrible solution to the problem.
The only vaguely valid dual use technology I can see coming out of this is improving space-rated processing enough that deep space probes sent out to Uranus or whatever can run with more processing power than a Ti-82 and thus can actually do some data processing rather than clogging up the deep space network for three weeks on an uplink with less power than a lightbulb
None of these problems seem intractable, just really hard and probably not being solved soon, but one has to start somewhere... so at least the billionaires will fund some scientists and engineers who will do that work?
Building datacenters in the arctic also has the added benefit that sysadmins would have to take polar bear safety lessons, which would be pretty funny.
Risky/untried things aren't dumb because they're hard, they're dumb when they're more expensive/harder than cheaper/easier alternatives that already exist that do the same thing.
I agree with most of this post and think the problems are harder than the proponents are making them seem.
But, 1) literally the smartest people and AI in the world will be working on this and 2) man I want to see us get to a type 2 civilisation bad.
The layout of this blog post is also very interesting, it presents a bunch of very hard items to solve and funny enough the last has been solved recently with starlink. So we can approach this problem, it requires great engineering but it’s possible. Maybe it’s as complicated as CERNs LHC but we have one of those.
Next up then is the strong why? When you’re in space, if you set the cost of electricity to zero, the equation gets massively skewed.
Thermal is the biggest challenge but if you have unlimited electricity, lots of stuff becomes possible. Fluorinert cooling, piezoelectric pumps and dual/multi stage cooling loops with step ups. We can put liquid cooling with piezos on phones now, so that technology is moving in the right direction.
For a thought experiment, if launch costs were $0/kg, would this be possible? If the answers yes, then at some point above $0/kg it becomes uneconomical, the challenge is then to beat that number.
If you think about it, all the existing data centers are in space already. They're just attached to a big ball of rock, water, and air that acts as a support system for them, simplifying cooling and radiation protection.
If humans are going to expand beyond the Earth, we'll certainly need to get much better at building and maintaining things in space, but we don't need to put data centers in space just to support people stuck on the ground.
Google’s paper [1] does talk about radiation hardening and thermal management. Maybe their ideas are naive and it’s a bad paper? I’m not an expert so I couldn’t tell from a brief skim.
It does sound to me like other concepts that Google has explored and shelved, like building data centers out of shipping container sized units and building data centers underwater.
The only sentence in the whole "paper" about cooling is
> Cooling would be achieved through a thermal system of heat pipes and radiators while operating at nominal temperatures
Which is kind of similar to writing a paper about building a bridge over the Pacific and saying "The bridge would be strong enough by being built out of steel". Like you can say it, but that doesn't magically make it true.
Only legit thing I can see this being used for is redundant archival storage or just general research into hardening equipment to radiation or micrograv (eg for liquid cooling). But anything that generates significant amounts of heat seems like it'd be a huge problem.
Then again there's lots of space in space, perhaps it's possible to isolate racks/aisles into their own individual satellites, each with massive radiant heatshedding panels? It's an interesting problem space that would be very interesting to try to solve, but ultimately I agree with OP when we come back around to "But, why?" Research for the sake of research is a valid answer, but "For prod"? I don't see it.
One thing I haven't seen talked about at all: how quickly would space heat up?
I presume Earth's gravity largely keeps the exosphere it has around it. With some modest fractional % lost year by year. There is a colossal vast volume out there! But given that there's so little matter up in space, what if any temperature rise would we expect from say a constant 1TW of heat being added?
The sun’s radiation hitting earth is 44,000 terawatts. I think we’re fine with an “extra” terawatt. (It’s not even extra, because it would be derived from the sun’s existing energy.)
At sea-level there can be 1.225 kg/m^3 of particles. There's a lot of matter to absorb heat.
In the exosphere we have 1e-13 kg/m^3 of particles.
My point is that the exosphere while huge has an incredibly tiny thermal battery. I'm not convinced that, were we able to dump heat into it, that it really would be insignificant heating over time.
And there's little way for the articles here to cool down. There's no matter to transfer their energy to.
I guess the thing is, it doesn't matter. It seems like the exosphere is actually already >500 degrees: that after you leave the 80km menopause temperatures soar quickly, in what scant air is left. I was still using a model of thermal transfer. But the only cooling possible is passive radiative cooling, is to glow your energy away. Some of this will find other exospheric particles to hit & excite more, but they're already incredibly energetic up there, and there's just not many particles at all, so perhaps a lot of that radiation might escape the exosphere without collision. Again my mistake: thermal transfer is simply not that relevant (aside some shielding against these particles in vulnerable spots), it's all passive radiation being used to cool.
It would still be interesting to me to have some guestimates for what the current energy balance of the exosphere is. What is heating it, how much, and where/how-much is it able to dissipate its energy?
Except you don’t build a data center, you add a GPU to an individual starlink node. If you can do that a couple hundred or thousand times you’ve got a lot of compute in space. The next question is how would you redesign compute around your distributed power and cooling profiles?
The article doesn’t talk about the actual engineering challenges. (Such as scaling down the radiative cooling design, matching compute node to the maximum feasible power profile, etc)
I’m not arguing it’ll be easy or will ultimately work, but articles like this are unhelpful because they don’t address the fundamental insight being proposed.
Starlink satellites would be pointless for doing computation because they are spread across the Earth resulting in horrible latency. AI companies spend lots of money on super fast connects within a datacenter.
Starlink with GPU might have some advantage for running edge GPU. But most Starlink customers are close to ground station and it makes a lot more sense to have GPUs there. It is a lot easier to manage them than launching new satellites which could take years.
Shhh, I'm begging people, if brain-dead VCs want to waste their money on things that are obviously farcical (and not actively destructive), please let them and stop doing their due diligence for them. The alternative is that they turn their impossible amounts of capital towards societally-destructive acts like buying up all the real estate in the world and turning us back into land-slaves.
"[..] deploying a solar array with photovoltaic cells – something essentially equivalent to what I have on the roof of my house here in Ireland, just in space. It works, but it isn't somehow magically better than installing solar panels on the ground – you don't lose that much power through the atmosphere"
As an armchair layman, this claim intuitively doesn't feel very correct.
Of course AI is far from a trustworthy source, but just using it here to get a rough idea of what it thinks about the issue:
"Ground sites average only a few kWh/m²/day compared to ~32.7 kWh/m²/day of continuous, top-of-atmosphere sunlight." .. "continuous exposure (depending on orbit), no weather, and the ability to use high-efficiency cells — all make space solar far denser in delivered energy per m² of panel."
Orbital data centers are very hard but this isn't a good explanation of why. There really is more light in space since certain orbits are always in daylight. Radiators are no larger than the solar panels so if you can build multi square kilometer solar arrays you can probably also build massive radiators.
Datacenters in space is about circumventing nation states masked as ambitions to generate more power.
Follow the rationale:
1. Nation states ultimately control three key infrastructure pieces required to run data centers (a) land (protected by sovereign armed forces) (b) internet / internet infra (c) electricity. If crypto ever became a legitimate threat, nation states could simply seize any one of or all these three and basically negate any use of crypto.
2. So, if you have data centers that no longer rely on power derived from a nation state, land controller by a nation state or connectivity provided by the nation state's cabling infra, then you can always access your currency and assets.
91 comments
[ 5.0 ms ] story [ 76.1 ms ] thread> After laughing at "the vacuum of space for cooling" I closed the page because there was nothing serious there. Basic high school physics student would be laughing at that sentence.
On the SEU issue I’ll add in that even in LEO you can still get SEUs - the ISS is in LEO and gets SEUs on occasion. There’s also the South Atlantic Anomaly where spacecraft in LEO see a higher number of SEUs.
The section of the article that talks about them isn’t great. At least for FPGAs, the state of the art is to run 2-3 copies of the logic, and detect output discrepancies before they can create side effects.
I guess you could build a GPU that way, but it’d have 1/3 the parallelism as a normal one for the same die size and power budget. The article says it’d be a 2-3 order of magnitude loss.
It’s still a terrible idea, pf course.
As a sibling post noted, SEUs are possible all the way down to sea level. The recent Airbus mass intervention was essentially a fix for a badly handled SEU in a corner case.
Latency becomes high but you send large batches of work.
Probably not at all economical compared to anywhere on Earth but the physics work better than orbit where you need giant heat sinks.
https://pmc.ncbi.nlm.nih.gov/articles/PMC9646997/ ("Thermophysical properties of the regolith on the lunar far side revealed by the in situ temperature probing of the Chang’E-4 mission" (2022))
https://www.engineeringtoolbox.com/thermal-conductivity-d_42...
(Imagine, for entertainment purposes, what would happen if you wrapped a running server rack in a giant ball of rock-wool insulation, 50 meters in radius).
Only way to dissipate large amounts of heat on the moon is with sky-facing radiators.
Latency wise it seems okay for llm training to put them higher than Starlink to make them last longer and avoid decelerating because of the atmosphere. And for inference, well, if the infra can be amortized over decades than it might make the inference price cheap enough to endure additional latencies.
Concerning communication, SpaceX I think already has inter-starlinks laser comms, at least a prototype.
You still have to build the GPUs, etc for the datacenter whether it’s on Earth or in orbit. But to put it in space you also need massive new cooling solution, radiation shielding, orbital boosting, data transmission bandwidth, and you have to launch all of that.
And then, there are zero benefits to putting a datacenter in space over building it on Earth. So why would you want to add all that extra expense?
The obsolete stuff can be deorbited or recycled in space.
There are dozens of companies solving each problem outlined here; if we never attempt the 'hard' thing we will never progress. The author could have easily taken a tone of 'these are all the things that are hard that we will need to solve first' but actively chose to take the 'catastrophically bad idea' angle.
From a more positive angle, I'm a big fan of Northwood Space and they're tackling the 'Communications' problem outlined in this article pretty well.
It's the opposite of engineering, where you understand a problem space and then try to determine the optimal solution given the constraints. This starts with an assumption that the solution is correct, and then tries to engineer fixes to gaps in the solution, without ever reevaluating the solution choice.
Nobody is proposing data centers at the South Pole. This isn’t because it’s difficult. It is difficult, but that’s not the reason it’s not being looked at. Nobody’s doing it because it’s pointless. It’s a massive hassle for very little gain. It’s never going to be worth the cost no matter what problems get solved.
Data centers in space are like that. It’s not that it’s difficult. It’s that the downsides are fundamentally much worse than the advantages, because the advantages aren’t very significant. Ok, you get somewhat more consistent solar power and you can reach a wider ground area by radio or laser. And in exchange for that, you get to deal with cooling in a near perfect insulator, a significantly increased radiation environment, and difficult-to-impossible maintenance. Those challenges can be overcome, sure, but why?
This whole thing makes no sense. Maybe there’s something we just aren’t seeing, or maybe this is what happens when people are able to accumulate far too much money and nobody is willing to tell them they’re being stupid.
- More junk whizzing around Earth.
- Inaccessibility for maintenance.
- Power costs.
- Susceptibility to solar storms and cosmic rays.
Risky/untried things aren't dumb because they're hard, they're dumb when they're more expensive/harder than cheaper/easier alternatives that already exist that do the same thing.
But, 1) literally the smartest people and AI in the world will be working on this and 2) man I want to see us get to a type 2 civilisation bad.
The layout of this blog post is also very interesting, it presents a bunch of very hard items to solve and funny enough the last has been solved recently with starlink. So we can approach this problem, it requires great engineering but it’s possible. Maybe it’s as complicated as CERNs LHC but we have one of those.
Next up then is the strong why? When you’re in space, if you set the cost of electricity to zero, the equation gets massively skewed.
Thermal is the biggest challenge but if you have unlimited electricity, lots of stuff becomes possible. Fluorinert cooling, piezoelectric pumps and dual/multi stage cooling loops with step ups. We can put liquid cooling with piezos on phones now, so that technology is moving in the right direction.
For a thought experiment, if launch costs were $0/kg, would this be possible? If the answers yes, then at some point above $0/kg it becomes uneconomical, the challenge is then to beat that number.
Any active cooling solution you can think of actually makes the problem worse (unless it's "eject hot mass").
If humans are going to expand beyond the Earth, we'll certainly need to get much better at building and maintaining things in space, but we don't need to put data centers in space just to support people stuck on the ground.
It does sound to me like other concepts that Google has explored and shelved, like building data centers out of shipping container sized units and building data centers underwater.
[1] https://services.google.com/fh/files/misc/suncatcher_paper.p...
> Cooling would be achieved through a thermal system of heat pipes and radiators while operating at nominal temperatures
Which is kind of similar to writing a paper about building a bridge over the Pacific and saying "The bridge would be strong enough by being built out of steel". Like you can say it, but that doesn't magically make it true.
Then again there's lots of space in space, perhaps it's possible to isolate racks/aisles into their own individual satellites, each with massive radiant heatshedding panels? It's an interesting problem space that would be very interesting to try to solve, but ultimately I agree with OP when we come back around to "But, why?" Research for the sake of research is a valid answer, but "For prod"? I don't see it.
I presume Earth's gravity largely keeps the exosphere it has around it. With some modest fractional % lost year by year. There is a colossal vast volume out there! But given that there's so little matter up in space, what if any temperature rise would we expect from say a constant 1TW of heat being added?
https://www.nasa.gov/wp-content/uploads/2015/03/135642main_b...
In the exosphere we have 1e-13 kg/m^3 of particles.
My point is that the exosphere while huge has an incredibly tiny thermal battery. I'm not convinced that, were we able to dump heat into it, that it really would be insignificant heating over time.
And there's little way for the articles here to cool down. There's no matter to transfer their energy to.
I guess the thing is, it doesn't matter. It seems like the exosphere is actually already >500 degrees: that after you leave the 80km menopause temperatures soar quickly, in what scant air is left. I was still using a model of thermal transfer. But the only cooling possible is passive radiative cooling, is to glow your energy away. Some of this will find other exospheric particles to hit & excite more, but they're already incredibly energetic up there, and there's just not many particles at all, so perhaps a lot of that radiation might escape the exosphere without collision. Again my mistake: thermal transfer is simply not that relevant (aside some shielding against these particles in vulnerable spots), it's all passive radiation being used to cool.
It would still be interesting to me to have some guestimates for what the current energy balance of the exosphere is. What is heating it, how much, and where/how-much is it able to dissipate its energy?
I’m not arguing it’ll be easy or will ultimately work, but articles like this are unhelpful because they don’t address the fundamental insight being proposed.
Starlink satellites would be pointless for doing computation because they are spread across the Earth resulting in horrible latency. AI companies spend lots of money on super fast connects within a datacenter.
Starlink with GPU might have some advantage for running edge GPU. But most Starlink customers are close to ground station and it makes a lot more sense to have GPUs there. It is a lot easier to manage them than launching new satellites which could take years.
As an armchair layman, this claim intuitively doesn't feel very correct.
Of course AI is far from a trustworthy source, but just using it here to get a rough idea of what it thinks about the issue:
"Ground sites average only a few kWh/m²/day compared to ~32.7 kWh/m²/day of continuous, top-of-atmosphere sunlight." .. "continuous exposure (depending on orbit), no weather, and the ability to use high-efficiency cells — all make space solar far denser in delivered energy per m² of panel."
Follow the rationale:
1. Nation states ultimately control three key infrastructure pieces required to run data centers (a) land (protected by sovereign armed forces) (b) internet / internet infra (c) electricity. If crypto ever became a legitimate threat, nation states could simply seize any one of or all these three and basically negate any use of crypto.
2. So, if you have data centers that no longer rely on power derived from a nation state, land controller by a nation state or connectivity provided by the nation state's cabling infra, then you can always access your currency and assets.