Orbital Data Centers: Compute Goes to Space for Real

In November, a 60-kilogram satellite called Starcloud-1 carried an Nvidia H100 into orbit and trained a language model in space — the first time that's ever happened. Since then, SpaceX has filed FCC plans for up to a million orbital data center satellites, a five-month-old startup called Orbital asked the FCC for permission to fly 100,000 more, and Starcloud has raised $250 million on a roadmap to 88,000 computing satellites. The pitch decks have become filings. I read all of them so you don't have to.

It's Real Now: The Proof Points

Let's start with what has actually happened, because the fact base here changed dramatically in the last year.

Starcloud-1, launched via SpaceX in November 2025, is the first satellite to fly an Nvidia H100 GPU — by the company's own estimate, roughly 100 times more powerful than any compute previously sent to space. In December it became the first spacecraft to run a version of Google's Gemma in orbit and the first to train an LLM in space — Andrej Karpathy's nanoGPT, with Karpathy himself posting "It begins." CEO Philip Johnston's claim to CNBC: "Anything you can do in a terrestrial data center, I'm expecting to be able to be done in space."

Since then the field has gotten crowded fast. Starcloud is prepping Starcloud-2, an 8-kilowatt spacecraft launching about eight months after the first, with signed contracts to provide cloud and edge services that — per Johnston — already pay for the spacecraft. Orbital, a Los Angeles startup founded by Spin scooter founder Euwyn Poon, filed with the FCC in June 2026 for up to 100,000 data center satellites targeting 10 gigawatts of orbital compute: 100-kilowatt-class spacecraft at 500–850 km, with solar arrays and radiators spanning around 100 meters and dry mass of 1.5–2.5 tonnes. SpaceX itself has outlined 150-kilowatt-class orbital data centers with filings for up to a million of them. Blue Origin is in the hunt too, and Axiom Space is attaching orbital data center nodes to its commercial station plans.

When the company that owns the launch market files regulatory paperwork for a million computing satellites, "data centers in space" stops being a joke and becomes a market.

The Physics Bill Comes Due

Here's where the renderings and reality part ways, because the hard problems are all still ahead.

Getting rid of heat is brutal in space. On Earth, data centers dump waste heat into water and air via convection — cheap and effectively unlimited. In vacuum you have exactly one option: radiation, which requires enormous radiator area for AI-scale heat loads. Orbital's filing is honest about this — those ~100-meter solar arrays and radiators on a 1.5–2.5 tonne dry mass bus are the whole engineering challenge. Every square meter has to be launched, deployed robotically, and survive 15 years of thermal cycling and micrometeoroids.

Everything waits on Starship. Johnston's own math: the 88,000-satellite constellation is "very dependent on Starship flying frequently," with first commercial customer payloads penciled in for 2029 or 2030, and the cost math only closes once Starship gets below roughly $500 per kilogram. Orbital's CEO put it more bluntly: "The complexity is all launch. The rest of it is first principles physics and manufacturing" — solar panels, radiators, electronics, plus radiation shielding he waves away as "solvable." Everyone in this market is standing in the same line, waiting on the same rocket that Artemis is also waiting on.

The demand side starts niche. The near-term money isn't in competing with AWS — it's in Earth observation processing and edge compute for other spacecraft, where Johnston says providers can charge 100 to 1,000 times terrestrial GPU-hour rates. That's a real business. It is not "the AI power crisis, solved."

Reading the Roadmaps Honestly

The numbers being filed strain credulity, and it's worth saying so plainly. 88,000 satellites at 200 kilowatts each. 100,000 more from Orbital. A million from SpaceX. Even at Starship's most optimistic cadence, that's a mass-to-orbit bill larger than everything humanity has ever launched, combined, by a wide margin. These filings are regulatory land grabs — spectrum and orbit rights are use-it-or-lose-it, so everyone files absurdly large to reserve the address space. The same pattern played out with broadband constellations a decade ago.

Also worth noting: Johnston himself says model training won't happen in space "anytime soon" because it requires docking large structures together, and he estimates training will be under 1% of AI workloads within five years anyway. The actual target market is inference workloads that don't need sub-50-millisecond latency — code generation tasks, back-office agents, customer service. That's a far more sober pitch than the "unlimited solar power for AI" headline, and notably, it's the opposite of what Starcloud-1 just demonstrated. The demo did the sexy thing; the business plan is the boring thing.

There's a real idea buried under the hype, though. Terrestrial data centers are hitting genuine walls — grid interconnect queues, water for cooling, land permits. Orbital solar is continuous, un-metered, and competes with nobody's electrical grid. If launch costs genuinely break below $500/kg, the energy-cost argument stops being nonsense. It's a big if, but it's no longer a physically impossible one.

The Politics Nobody Priced In

One more wrinkle the renderings ignore: regulators are now the gatekeepers. Orbital's FCC filing was explicitly framed as "a first regulatory step" while the satellite design is still being finalized, and CEO Euwyn Poon — who built the Spin scooter network before selling it to Ford — is talking publicly about space traffic management for compute constellations the way cities once had to sort out scooter fleets. Between FCC orbital debris rules, international spectrum coordination, and the growing list of players (Starcloud, Orbital, SpaceX, Blue Origin, Cowboy Space) all wanting adjacent orbits at 500–850 km, the coordination problem is getting ahead of the hardware. Meanwhile, nobody in Washington has even started writing the framework for "what happens when a data center deorbits." The companies filing 100,000-satellite paperwork are betting regulators stay asleep for a few more years. Based on how the last constellation boom went, that's a bad bet.

What I Think

I went into this expecting to write a debunking, and I'm only delivering half of one. The thermal engineering remains the unsolved core problem — radiators at gigawatt scale, deployed and maintained robotically, are hand-waved in every deck I read. The constellation numbers are regulatory fiction. And the entire industry is a single point of failure called Starship, which — as anyone tracking Artemis knows — has its own schedule to worry about.

But the H100 in orbit is a fact. The first space-trained LLM is a fact. Starcloud-2 flying with paying edge-compute contracts is a near-term fact. The honest forecast: orbital data centers won't replace terrestrial ones — they'll annex the top of the market, the same way GPUs annexed training from CPUs. Workloads that need low latency stay on Earth forever. Earth-observation processing, spacecraft edge compute, and eventually latency-tolerant bulk inference move up. And your next frontier model still trains in a warehouse in Abilene, because even the people building orbital data centers say so.

Watch the radiator mass ratios on Starcloud-2 and the first Orbital demonstrator next year. That one number tells you whether this is an industry or an art project.