It does. But very slowly. Which is the point being tried to be made to you but you cannot grasp.
How cooling in vacuum actually works
On Earth, data centers dump heat by conduction into water and convection into air, then cool those with chillers and cooling towers. In space you lose conduction to ambient and convection entirely, so every watt of waste heat must leave as photons off a radiator surface pointed into cold, dark sky.
A typical proposal is: chips → cold plates → pumped liquid loop → big external radiators; the loop just moves heat from the electronics to the radiating surface. The vacuum is the insulator; the radiator is the thing that “sees” 3 K background and glows in infrared to get rid of heat.
Why this is such a headache
Radiation is weak compared with convection, so you need a crazy amount of radiator area per megawatt. One recent engineering estimate puts it at roughly 2,500–3,300 square meters of radiator for just 1 MW of data‑center‑class heat load, depending on temperature. Real AI training centers can be 50–100 MW, which pushes you into the hundreds of thousands of square meters of radiator surfaces, far beyond anything yet flown; the ISS has only on the order of a thousand square meters of radiators.
Even advocates acknowledge that, in practice, heat dissipation will strongly cap how much compute you can pack into each orbital platform and will add a ton of mass and complexity. That is why a lot of physicists and space engineers are skeptical when Musk waves it away as “space cooling is free.”
Although Alphabet seems very interested in them and Sundar Pichai seems like a rational actor. They’ve thought about it and see some merit in the idea.
I agree with you though, at first blush there seem to be lots of downsides.
Well, there’s the cooling issue which is, uh, difficult, but there’s the actual power issue which is even more ridiculous.
“The International Space Station generates enough electricity to power roughly 100 advanced AI chips, according to [one estimate] . An orbital data center may need to produce electricity to power many thousands of chips inside thousands of satellites.
Enter huge solar arrays. At an event in March, Musk showed a rendering of a SpaceX “AI Sat Mini.” Nothing was small about the device. It featured solar arrays that made the satellite far longer than the company’s Starship rocket, which is around 400 feet tall.
“Solar arrays of multikilometers in scale are what’s needed,” [Rocket Lab] Chief Executive Peter Beck told investors earlier this year.”
Data Centers in Space: A Pipe Dream, or AI’s Next Big Thing?”
The Wall Street Journal
I read one estimate that to have a data center comparable to a small data center on Earth, you would need a solar array 20 kiilometers by 20 kiilometers. Of course the radiating coolers wouldn’t have to be that big. Maybe only half the size?
Obviously no one is going to put a 20km x 20km solar panel in space, so they would do it with lots of little ones - but then those require a multitude of lifts, assembly, and monitoring. And each “data center” being smaller would be limited as to what it could, although you could link them together, I suppose, But in a world where we’re trying to shrink the distance between transistors on logic chips to nanometers to increase speed, having satellites 10 miles apart would create a new kind of latency problem.
Not that all this couldn’t be solved with really swell, gee whiz technology, and perhaps someday it will. Not real soon, I predict.
There was a crushing interview on NPR on the meanings behind the Trump visit to China. The second China shock is coming. If we do not compete fiercely in AI, the shock will be even worse.
We must raise taxes. Part of this is because the corporate tax rate is 21% in relative terms; the math has corporations not investing their own money. They can declare profits instead.
We need a massive reinvestment in our country. We need the economies of scale that would come from it. We need to displace China on the world stage.
The current me me me just give me the screw others profits will fail the country.
For some stuff latency isn’t too important. Training LLMs for example, or processing satellite imagery. There is is a bandwidth bottleneck between space and Earth, so you could do all the image processing in space, and just send the final product back to Earth.
Yes, but even more Elon loves “Projecting Projects” and then, after gaining lots of attention and perhaps some capital, loves the searching for solutions that often simply do not exist:
human colonization of Mars
tunneling processes that somehow obliterate big rocks
efficient governance that is heedless of politics and comprehensible policies
You might be giving him too much credit. Remember in the movie Apollo 13 when they are all shivering in LEM? That’s because they powered down all the electronics but the radiators continued to radiate, so it got very cold without the waste heat.
The issue is Musk’s claim that cooling in space is free. That’s true only if you neglect the cost of getting the radiator into orbit. Gemini tells me the radiators on the ISS can reject about ~131 kW of heat and weigh about ~22,532 lb. Or 172 lb per kW.
If a space data center is 1 MW, and it costs $1,200/lb to get to orbit, that means it would cost about $200 million just to get the radiators into space. That doesn’t include any of the hardware costs, or the cost to get the rest of the data center into space. So launch costs need to come down a lot for this to pencil out.
As I mentioned above, sober people are working on this, so it might come about someday but there are some significant challenges. With the SpaceX IPO coming up, it might be prudent to take Musk’s claims with a boulder or two of salt.
Separately, SpaceX IPO is going to be fun. There is a reason why Musk decided to rent his DC capacity to anthropic, i.e., xAI doesn’t have much use for those GPU’s, whereas OpenAI, and Anthropic cannot get enough of them.
For reference, Twitter is folded into xAI which is folded into SpaceX. So SpaceX is the parent co.
Let me ask a simple question. Why does the earth cool at night? And before you say, duh, there’s no sun pointing at it, you would be correct - partially.
Why do moons (and planets) glow in infrared when occulted by other heavenly bodies (not in clear line of solar radiation)?
The answer to the above is germane to the topic.
And the answer when applied to a single grace hopper server rack module (99% power consumption, continuously) requires a rejection of all heat (40Kw)
A Carbon composite radiator which rejects heat from both faces (in the perpetual shade of the solar panels which power the GH module) must be ~32M^2 when operating at the maximum dT between the Server rack (at 65C). If we desire the keep the system a bit cooler or to add engineering design factor for loss over the life of the unit, we simply factor up the radiator square area (double sided, of course)
The solar panels to power this unit would require ~100M^2 (assuming 30% efficiency solar panel array) which would DEFINITELY be sufficient enough to shield the module AND the radiator 100% of the time as long as the gyros continue to keep the orientation to the sun nominal.
To put this in context, a 1 MW data center today has a capex cost of around $10M. You’re right that there’s no amount of savings on energy that could justify the expense of getting just the cooling system into orbit at today’s prices.
I would guess, though, that Musk isn’t using today’s prices. He’s imagining Starship prices. Their target is to get to $10 to $50 per pound, or about two orders of magnitude less than current prices.
If you assume super-cheap access to space, then the numbers get a little better. That’s a pretty strong assumption, and it might entirely fail to materialize - Musk has often been wrong in projecting how much stuff will cost. But I think that’s the reason why he (and others) talk about Space AI Centers as a possibility. Because they’re counting on cheap space access pretty soon.
You seem to know what you’re talking about. What about quantum computing? Lots of talk about dissipating heat, but what if processing power ran cooler?
