Space isn't "cold" like Greenland is cold. Space is a vacuum. A vacuum is the best insulator in the world, because the most effective methods of heat transfer, convection and conduction, are off the table. It is easier to cool things in the desert than in space. We've been putting stuff in space for 70 years now; radiative heat transfer is a solved problem. We know how to gather solar energy and dissipate heat. And unfortunately, while space is great for the first thing, it absolutely fucking sucks for the second thing. The only thing space has going for it is, potentially, 24-hour solar energy. Obviously, that's only true if your orbit keeps you out of the Earth's shadow.
The TL;DR before I get going is the napkin math doesn't work.
An NVIDIA B200 DGX system holds 8 GPUs, draws about 11kW, weighs around 130 kg, and costs around $500k.
There are lots of different types of space solar panels. The latest ultra-lightweight thin-film panels deliver about 0.2 kW/m^2 and weigh about 0.1 kg/m^2
Space radiators vary as well, but your best-case scenario is around 0.35 kW/m^2 and 12 kg/m^2. Radiators are really heavy.
An AI datacenter, like the ones xAI and OpenAI use, has the equivalent around 60,000 DGX systems in it.
SO, how much shit needs to go into space to build our space datacenter?
GPUs: 130kg x 60,000 = 7,800,000 kg
Solar panel: 0.1 kg / m^2 x 1/(0.2 kW / m^2) x 11 kW x 60,000 = 330,000 kg
Radiators: 12 kg/m^2 x 1/(0.35 kW/m^2) x 11kW x 60,000 = 22,628,571 kg
So to put one datacenter in space with the capability of training frontier models, we need to put about 30,700,000 kg into orbit.
A fully expendable Falcon Heavy can put 63,800 kg into orbit, so it will take about 480 launches to get the space datacenter built. An expendable launch, which has the best price/kg, costs about $150m. So the cost of just launches and GPUs:
$150m x 480 + 0.5 m x 60000 = $102b
I have no idea how much the solar panels and radiators cost, but it's not negative money. $100b is probably the absolute floor of what that space datacenter costs. xAI's terrestrial datacenter cost, according to various estimates, around $7-$18b. AI datacenters are already losing money, and you want me to believe that a datacenter that costs 10x more than a conventional datacenter--which by the way can't be serviced without a $50-$100m rocket launch, and we haven't even considered the communications infrastructure--is somehow going to make money?
Let's do a sanity check, correlate this with Starlink costs. According to that infallible AI we need so much of that space datacenters somehow make sense, a Starlink V2 Mini satellite costs about $1m to put in orbit and is a complete system, compute + solar + radiator + radiation shielding, the works. The internet says LEO comms satellites draw around 1kW of power. So, roughly, 1 DGX ~ 10 Starlink V2 Minis. 60,000 x 10 x $1m = $600b. So yeah, looks like the hundred-billion range is about right for a space datacenter. As with all things, "do it in space" is about a 10x multiplier in costs.
It's not happening, not next year, and not ever. It's not happening because space fucking sucks. Space is the absolute worst place to do anything that is physically possible on Earth.
