A handful of small startups are betting that space, not Virginia or Texas, is where the next AI data center gets built, and they're moving faster than Google or SpaceX.
Last November, a 60-kilogram satellite called Starcloud-1 rode a SpaceX Falcon 9 into orbit carrying a single Nvidia H100 GPU. Once in space, it trained a small language model on the complete works of Shakespeare and ran a version of Google's Gemini model, the first time either had happened outside a terrestrial data center. The company behind it, Redmond, Washington-based Starcloud, went through Y Combinator's summer 2024 batch. In March it raised a $170 million Series A led by Benchmark and EQT, pushing its valuation to $1.1 billion and making it, by some accounts, the fastest company to reach unicorn status coming out of a YC demo day. Commercial deals with AWS, Google Cloud, and Crusoe came with the round, according to a report from TechCrunch. Starcloud has since filed with the FCC for a constellation of up to 88,000 satellites.
That filing alone tells you how seriously the idea is being taken. You don't ask regulators for 88,000 satellites on a whim.
The pitch is straightforward. Land near substations is scarce, permitting takes years, and the grid itself is buckling under AI demand. Grid operator PJM has had to manage congestion around Northern Virginia's data center corridor, and Blackstone shelved a Virginia data center project this year over power constraints. In orbit, a satellite gets close to constant sunlight and no zoning board. Move the compute up, the thinking goes, and the power problem mostly disappears.
Starcloud isn't alone, and it isn't even the biggest name chasing this. Google is running its own effort, called Project Suncatcher, in partnership with Planet Labs, with two test satellites carrying TPUs and free-space optical links planned for launch by early 2027, according to Google's own research blog. The company has modeled an 81-satellite cluster flying in formation at roughly 650 kilometers up. SpaceX filed with the FCC in January for authorization covering up to a million satellites and 100 gigawatts of orbital computing capacity. Jeff Bezos has sketched gigawatt-scale orbital facilities for Blue Origin. Smaller players are carving out narrower bets: Axiom Space deployed two orbital computing nodes in January and is targeting a fully connected node aboard the International Space Station by 2027. Lonestar Data Holdings tested data storage hardware on an Intuitive Machines lunar lander back in February 2025 and has since signed a $120 million deal with Sidus Space to build six storage satellites. OrbitsEdge, out of Cocoa Beach, Florida, sells radiation-hardened hardware to satellite operators who'd rather process data in orbit than beam it down raw. Cowboy Space, founded in 2024 by Robinhood co-founder Baiju Bhatt, is aiming for its first satellite launch this year.
Here's the problem nobody chasing this bet has solved: launch costs. Google's own researchers have said orbital computing only becomes cost-competitive with ground-based data centers once launch prices fall to around $200 per kilogram. A SpaceX Falcon 9 currently runs closer to $2,700 per kilogram. That's not a rounding error, it's an order of magnitude. Cooling is the other wall. There's no air or water in vacuum to carry heat away from a chip, so Google's engineers have said a single orbital data center would need something like 2.15 million square feet of radiator surface to dump its heat, all while radiation degrades hardware that can't be repaired once it's up there.
Analysts at MoffettNathanson have pegged commercial viability at scale as more of a 2030s story, and only if Starship reaches a mass commercial launch cadence that actually gets costs down near that $200 threshold.
Not everyone in the industry is convinced it's even a matter of time. Sam Altman called orbital data centers
Also read: Micron Breaks Ground on a $9.3 Billion Bet to Crack SK Hynix's Grip on AI Memory • SK Hynix is racing to list on Nasdaq just as the AI chip trade it built its fortune on wobbles • Hackers Just Showed How Fragile the AI Software Supply Chain Really Is
