Electricity generation is the constraining factor, but the sun does not turn off in space. xAI data centers in space drives cost to zero, even with inferior models.
Solar in space produces 30% more power, and doesn't turn off at night, meaning you don't need batteries. That means power costs, say, 25% of what it currently does measured against terrestrial solar and batteries.
The 75% electricity discount needs to pay for launch vehicles, specially designed satellites, and the inability to service the hardware or resell it when it's EOL for the data center.
It's a gamble. Maybe it'll turn out to be a slight edge, maybe it'll turn out to fail, but it's not a sure thing and it certainly isn't going to hugely decrease the cost.
Especially since they're competing against Google and their custom designed hardware that's far more power efficient for AI. It's not clear that NVIDIA running at a 75% dollar discount beats Google's best TPU in compute per dollar.
There are far too many variables still unknown to all parties. Anyone trying to say with certainty "X will lose", whether X is terrestrial or space based DCs, is lying and probably trying to sell you something.
Time is a factor here. You need to launch each of these into space. Now, in the time it took to send those satellites up into space, how many datacenters and how many solar panels and how many GPUs could you have set up and built and had operational? If latency isn't a factor, why bother building them in the US? Build them in Iceland where they have cheap geothermal! The amount of rockets you'd need to launch to fit the same number of GPUs as a datacenter is definitely quite a few. SpaceX has done 165 flights in 2025. How many datacenters or solar farms or battery plants can you build in the time it'd take to launch 1 datacenter worth of satellites?
I'm entirely ignoring the problem of cooling in space (which is a huge problem!), and of how much it'd cost to launch the satellites.
For the price of launching a datacenter into space, you could probably build one in NA, one in Europe, and one in Asia and solve the "sun sets" problem that way with the side benefit of having excess capacity you can turn on by paying for local non-solar electricity.
You also have easy upgradeability and expansion, easier cooling and the value of the land and hardware as an asset. None of which are available in space.
You are going to be utterly shocked when you realize that solar panels work on the ground, too. You can buy so many batteries, and so many geographically separated locations for your panels, for the price of launching a datacenter into space.
> would you mind quoting the relative cost of batteries vs. solar panels for a 150kW solar-powered satellite
OK.
At a good location (~25% capacity factor), you need about 600 kW of panels to average 150 kW. Utility-scale solar runs roughly $0.50–$1.00/W installed, so call it ~$450K–$600K. Overnight storage (say ~16 hours) requires ~2,400 kWh. Adding a buffer for cloudy days, say 4,000–7,000 kWh total. At roughly $200–$350/kWh (utility-scale Li-ion), that's ~$1M–$2M.
In a favorable orbit, capacity factor is ~90–100% (GEO or sun-synchronous), so you need roughly 160–170 kW of panels. Space-qualified solar panels historically cost $100–$300/W. Even optimistically at $50–$100/W with newer manufacturing, that's 167 kW * $100/W = ~$17M optimistically, or 167 kW * $200/W = ~$33M realistically. You also need space-rated power management, thermal systems, and radiation-hardened electronics.
Even ignoring launch costs entirely, space solar is roughly 10–20x more expensive than ground solar + batteries, driven almost entirely by the enormous cost premium of space-qualified solar panels. Ground-based solar is extraordinarily cheap now (~$0.50–1/W), while space-grade panels remain orders of magnitude more expensive per watt.
The ground option wins overwhelmingly. The space option would only start to make sense if space-grade panel costs dropped to near terrestrial levels, which would require a revolution in space manufacturing.
150kW solar kit seems to cost around $150k[1]. With the cost of launch with Falcon Heavy, this would pay for about 100kg of payload[2]. Each Starlink satellite weighs ~300kg[3] so I suspect a 150kW "datacenter" satellite would weight much more. Where are the savings supposed to come from? Seems like you could overprovision terrestrial solar panels by 3-4x and still obviously come out ahead. And that's all before considering the R&D costs of building AI datacenter hardware that can survive the orbital radiation environment.
Never forget cooling. People imagine a square box with a ginormous sea of solar panels attached, and forget the atrocious and horrible cooling required to vent all the heat that makes your home sewage line look like bottled water by comparison.
Electricity generation is the constraining factor, but the sun does not turn off in space. xAI data centers in space drives cost to zero, even with inferior models.
I see no other future than SpaceXai winning.