A handful of startups want to move AI data centers into orbit, where sunlight is plentiful and zoning boards don't exist. The idea is current, serious, and still brutally expensive.
Last November, a 60-kilogram satellite called Starcloud-1 rode a SpaceX Falcon 9 into orbit with an Nvidia H100 GPU on board. Once there, Starcloud said it trained a small language model on Shakespeare and ran Gemma, Google's open model family, from space. That is a real milestone. It is also a long way from replacing the data center clusters being built in Virginia, Texas and the Middle East.
Redmond, Washington-based Starcloud is the company to watch because it is acting less like a science project and more like a company trying to grab spectrum, capital and customer attention before the giants arrive. SpaceNews reported in March that Starcloud raised a $170 million Series A led by Benchmark and EQT Ventures at a $1.1 billion valuation, making it the fastest Y Combinator company to reach unicorn status after completing the accelerator. The company has also filed with the FCC for a constellation of up to 88,000 satellites.
You don't ask regulators for 88,000 satellites on a whim.
The pitch is easy to understand because the pain on Earth is real. AI data centers need huge amounts of power, land and cooling, and the best sites near substations are getting harder to secure. Northern Virginia's data center corridor has already forced hard conversations about grid capacity, and large projects now live or die by access to electricity as much as by access to Nvidia chips. Put the compute in orbit, the argument goes, and you get near-constant solar exposure without a county planning meeting standing in the way.
That argument has pulled in bigger names. Google is running Project Suncatcher with Planet Labs, and Google's own research blog says it plans two test satellites carrying TPUs and free-space optical links for launch by early 2027. The company has modeled an 81-satellite cluster flying in formation at about 650 kilometers. SpaceX has filed plans tied to a much larger orbital computing push, with reports describing ambitions that run as high as a million satellites and 100 gigawatts of capacity. Jeff Bezos has talked about gigawatt-scale orbital facilities for Blue Origin. This is no longer a fringe slide deck passed around space conferences.
Still, the smaller companies give the story its edge. Axiom Space has been pushing orbital data-center work around its future commercial space station. Lonestar Data Holdings tested storage hardware on an Intuitive Machines lunar lander in February 2025 and later signed a $120 million deal with Sidus Space to build six storage satellites. OrbitsEdge, based in Cocoa Beach, Florida, sells radiation-hardened computing hardware for satellites that need to process data before sending it back to Earth. Cowboy Space, founded in 2024 by Robinhood co-founder Baiju Bhatt, is aiming at the same broad problem from a newer starting line.
Here's the problem. The economics are still ugly.
Google's researchers have said space-based data centers only start to compete with ground-based facilities if launch costs fall toward roughly $200 per kilogram. Falcon 9 rideshare pricing is still far above that, and even optimistic analyses tend to put the cost breakthrough in the 2030s, not next year. That gap isn't an accounting detail. It decides whether orbital AI is an infrastructure business or an expensive demonstration with a good press image.
Cooling is just as unforgiving. A data center on Earth can use air, water and industrial cooling systems. In vacuum, there is no surrounding medium to carry heat away from a hot chip, so heat has to be radiated out. Google's work points to radiator surfaces measured in the millions of square feet for a large orbital system. Then comes radiation, which damages electronics, and maintenance, which is simple in a warehouse and absurdly difficult at 650 kilometers above Earth.
Frankly, this is where the hype needs a handbrake. Sam Altman has called orbital data centers ridiculous for current AI computing needs, and he has a point. If a GPU fails in Abilene, Texas, someone can replace it. If it fails in orbit, you have a logistics problem attached to a physics problem attached to a balance sheet.
That doesn't make the idea silly. It makes it early. Starcloud's H100 satellite, Google's 2027 test mission and the FCC filings from Starcloud and SpaceX show that serious companies are now treating orbital compute as something to test, price and regulate. The right question isn't whether your next chatbot query will be processed in space. It won't. The question is whether the companies building AI infrastructure today are willing to reserve a place in orbit before the cost curve catches up.
For now, the answer is yes. The money is moving before the proof is complete, which is exactly how infrastructure races usually begin.
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
