Picture this: you drop a couple of million dollars on the hottest GPU money can buy, strap it to a 60-kilogram satellite, and yeet it into low Earth orbit. No, this is not the plot of a billionaire's fever dream. This is StarCloud, and the NVIDIA H100 that just left the atmosphere is apparently doing great up there.
Here is the story of how an H-series GPU became the most extra piece of hardware in existence, and why the startup behind it now wants 88,000 more.
The Backstory: A GPU Goes to Space
Back in November 2025, StarCloud hitched a ride on a SpaceX launch and sent its first satellite, Starcloud-1, into orbit. The 60-kilogram spacecraft carried something the space industry had never seen: a compute system built around an NVIDIA H100 GPU.
"It's about 100 times more powerful than any GPU that has been in space before," CEO and cofounder Philip Johnston said at a SpaceNews on-orbit computing event this spring. And then the flexes started coming.
- First to train an AI model in space.
- First to run high-powered inference on synthetic aperture radar (SAR) data in orbit.
- First to run a version of Google's Gemini model on a satellite.
Translation: the "cloud" in cloud computing is now literally above your head. When someone complains their GPU runs hot, StarCloud can say "cute, mine needs a radiator rated for the vacuum of space." 🛰️
The Money Moves: From One H100 to 88,000 Satellites
StarCloud is not here for one mildly impressive science experiment. In March, the Redmond, Washington startup raised $170 million in a Series A led by Benchmark and EQT Ventures at a $1.1 billion valuation. That makes it one of the fastest startups to hit unicorn status after Y Combinator, with the whole glow-up happening about 17 months after graduating.
The plan on the other side of that funding is genuinely unhinged in the best way. StarCloud has filed with the FCC for a constellation of 88,000 satellites. Each one is a three-ton spacecraft packing 200 kilowatts of compute. Add it all up and you get roughly 20 gigawatts of new AI compute floating in orbit, powered by the sun, no power grid required.
And the big players are already knocking. Crusoe, the company literally building OpenAI's data centers, signed an agreement for 10 gigawatts of power from StarCloud starting in the early 2030s. NVIDIA itself is working with StarCloud on a ruggedized "Space-1" variant of its Vera Rubin platform, tuned for mass, thermal, and radiation constraints.
The Reality Check: Space Is Hard, GPUs Are Dramatic
Before you pack your bags for the orbital data center, know the catch list. The whole 88,000-satellite timeline depends on SpaceX's Starship flying frequently, which Johnston expects means real commercial payloads around 2029 or 2030. The next spacecraft, Starcloud-2, is an 8-kilowatt bird launching in about eight months, and it already has paying cloud and edge contracts.
Radiation is the other fun one. StarCloud has run three rounds of proton testing at a cyclotron facility in Knoxville and a heavy-ion round at Brookhaven National Lab, all to figure out how much shielding and software babysitting the chips need so they do not randomly forget their weights when a solar flare rolls through.
There is also the radiator situation. Space is a vacuum, which means no air to dump heat into, so StarCloud designed a deployable radiator that it claims is about 10 times lower mass per watt than the ISS radiator and around 100 times cheaper per watt.
And the workloads? This is not a place for your 20-millisecond chat app. Johnston says orbital compute targets inference jobs that can tolerate latency above 50 milliseconds: code generation, back-office business agents, customer service bots, and Earth observation data processing where the GPU-hour rate can run 100 to 1,000 times what you pay on Earth.
Training giant models in space? Not happening anytime soon. Training needs docking huge structures together, and as Johnston points out, training will be under 1 percent of all AI workloads within five years anyway. The money is in inference, and apparently the money is in space.
So yes, an H100 went to orbit, it trained a model, it ran Gemini, and now a startup is trying to build the largest constellation in human history on the back of that one flex. The H-series era ended with a plot twist nobody saw coming: the GPUs went up. 🚀
Comments