Data Centers in Orbit: Bold Vision or Expensive Distraction?
Nvidia, Google, and SpaceX are all racing toward orbital compute. The physics say it might work eventually. The economics say not yet.
By Debra Brewster | Partner / CEO, Axiom AI Group USA, LLC
It sounds like something out of a science fiction serial: server racks orbiting the earth, cooled by the vacuum of space and powered by unlimited sunlight, training the next generation of AI models without ever touching a congested terrestrial power grid. It is also, as of this year, no longer purely hypothetical. Nvidia unveiled a dedicated space computing platform at its March conference, Google is testing satellite-based TPU clusters under its Project Suncatcher initiative, and a startup called Starcloud has already trained a large language model aboard an orbiting satellite carrying a single Nvidia H100 chip.
The appeal is not hard to understand given everything else happening in terrestrial data center development. Solar power in orbit is available continuously and at a fraction of the cost of grid electricity in some estimates, running as low as half a cent per kilowatt-hour. There is no water use for cooling, and no interconnection queue to wait years for. For an industry currently strangled by four-to-seven-year grid connection timelines and a multi-year transformer backlog, the pitch of simply building around the entire terrestrial bottleneck has obvious appeal to investors and executives alike.
The physics, though, come with real caveats that even the most enthusiastic proponents acknowledge. Cooling in space is not free; it is a mass-and-electricity tradeoff rather than an outright advantage.
Engineering estimates published this year suggest a single high-performance GPU rack would require roughly 80 square meters of radiator surface area to shed its heat through radiation alone, and that radiator degrades by as much as 40 percent over five years of exposure. The result is an effective power usage efficiency of around 1.3 once radiator mass and area are accounted for, which is a real improvement over ground-based air cooling, but not the free lunch the headline numbers imply.
The bigger challenge is connectivity, not cooling. Nvidia links individual GPUs within a training cluster at roughly 7.2 terabits per second. The best current optical satellite links deliver perhaps 100 gigabits per second, with next-generation systems promising maybe 400. That is one to two orders of magnitude short of what a single GPU needs, before accounting for the tens of thousands of GPUs that a frontier AI training run actually requires working in close coordination. Skeptics within the space industry argue this gap makes large-scale AI training in orbit implausible for years to come, even as smaller-scale inference and satellite-imagery processing tasks are already technically feasible today.
Then there is the cost of getting hardware into orbit in the first place. Even with SpaceX’s Starship promising launch costs as low as $10 to $13 per kilogram at full reusability and high flight cadence, a Google-authored feasibility study concluded that orbital data centers likely do not reach cost parity with ground-based facilities until launch costs fall to around $200 per kilogram and Starship is flying roughly 180 times a year, a scenario the study itself places around 2035. SpaceX’s current Starship track record stands at seven successes out of twelve flights, progress, but a long way from that kind of operational cadence.
The honest assessment, shared even by executives at companies racing to build this technology, is that orbital data centers are technically real and commercially premature. They are best understood today as a genuine long-term hedge against a terrestrial power crisis that shows no sign of easing, worth tracking closely and worth the current wave of billion-dollar experimentation, but not yet a practical alternative for the gigawatt-scale AI training campuses being built on the ground right now. As one AWS executive put it bluntly earlier this year, the industry is still “pretty far” from making this work at meaningful scale.