Starcloud Launched the First Data-Center GPU Into Orbit Last November. Its Next Satellite, Due to Fly This Month, Just Slipped to 2027.
Eleven months ago, a Seattle startup put a single Nvidia H100 into low Earth orbit and used it to run an AI model — proof that a data-center-class GPU can compute in space. The follow-up satellite, meant to prove this actually scales, was supposed to launch this month. It didn't, and the reason comes down to one unglamorous fact: there's no air up there to cool a chip with.
Size the radiator yourself
A simplified physics model, not a spacecraft design: assumes one flat, single-sided panel at a uniform steady temperature (emissivity 0.9), ignores structural mass, solar/Earth heat hitting the panel itself, and multi-surface geometry — real hardware is more complex. Drag "Heat to reject" (the GPUs' waste heat) and "Radiator temperature" and watch the panel area, computed from the Stefan–Boltzmann law, grow or shrink against a basketball court for scale.
The plain version
Picture the AI boom's biggest headache: building power plants, transformers, and water-cooled buildings fast enough to keep up with GPU demand. A handful of companies — a Seattle startup called Starcloud, Nvidia, SpaceX, and a startup literally named Orbital — think they've found a shortcut. Skip the power grid entirely. Put the data center in orbit, where the sun never sets and there's no utility company to beg for a hookup.
Last November, Starcloud proved the first piece works. It launched a 60-kilogram satellite, about the size of a mini-fridge, carrying a single Nvidia H100 — the same chip that powers plenty of today's AI data centers — and used it to run an AI model in orbit. That's the easy part.
The hard part is cooling. On Earth, a data center cools its chips with fans and chilled water — cheap, effective, boring. In the vacuum of space, none of that exists: there's no air to blow and no atmosphere to carry heat away by touch. The only way to get rid of waste heat is to radiate it outward as infrared light, the same way a hot stove warms your face from across a room without touching it. That effect is far weaker, per square meter, than moving air or water. So the bigger and more powerful an orbital data center gets, the bigger its radiator panels have to get — potentially bigger than the solar wings powering it in the first place.
That's reportedly why Starcloud's next, more powerful satellite — originally due to launch this month, carrying a newer Nvidia chip and a much bigger radiator — has slipped into 2027. The company's funding and valuation keep climbing into the billions anyway. Orbital data centers aren't science fiction anymore. They're just stuck on the same unglamorous problem every data center on Earth already solved with central air conditioning: where does the heat go?
The expert version
The case for orbital compute rests on one genuine physical advantage: solar collection. A ground-based photovoltaic panel loses a large share of incoming sunlight to atmospheric scattering and absorption and sits idle through night, cloud cover and low sun angles, which pushes real-world capacity factors to roughly 20–25%. A panel in the right orbit — a sun-synchronous dawn–dusk low-Earth orbit, or geostationary — sees unfiltered sunlight well over 99% of the time. Caltech researchers working on space-based solar power, including Ali Hajimiri and Sergio Pellegrino, put the resulting collection advantage at roughly 5–8x more energy per unit panel area than the same panel on Earth's surface, before any transmission loss, since this is about powering hardware on the satellite itself, not beaming power down.
The counterweight is heat rejection. Ground data centers dump waste heat via forced convection — moving air or water, which transfers heat far more effectively per unit area, at the temperature differentials chips tolerate, than radiation does. In vacuum, convection and conduction don't exist; the only channel left is thermal radiation, governed by the Stefan–Boltzmann law, P = εσAT⁴. Because radiated flux scales with the fourth power of absolute temperature, running a radiator hotter shrinks the area it needs sharply — but GPU silicon has a hard junction-temperature ceiling that caps how hot the loop feeding it can realistically run. One published engineering preprint sizes a roughly 2kW thermal load to a 2m×3m, 5kg radiator operating near 80°C. Hold that same physics and scale to the hundreds of kilowatts a useful GPU cluster draws, and required radiator area scales roughly linearly with heat load — which is why Nvidia CEO Jensen Huang has publicly called cooling, not power generation or compute, the actual bottleneck, estimating it could take years to solve at useful scale.
Confirmed: Starcloud flew a single H100 to low Earth orbit on November 2, 2025, and ran inference and training workloads using Google's Gemma model — independently reported by multiple outlets. Reported, less certain: Starcloud-2, originally slated to launch this month carrying a Blackwell-class Nvidia GPU and roughly 100x Starcloud-1's power generation plus the largest deployable radiator yet flown on a private satellite, has reportedly slipped into 2027 per coverage since September; the company's valuation across funding rounds is cited anywhere from roughly $1.1 billion to $2.3 billion depending on which round and date a given source is describing — treat the exact figure as unsettled.
Why it matters for tech + supply chain: If orbital compute ever works at scale, it trades Earth's power-grid, land and water fights for a launch-cadence and radiator-manufacturing problem — but cooling physics, not funding, is what's actually setting the timeline.
Why it matters for tech + supply chain: Heat rejection, bound by the Stefan–Boltzmann law rather than capital availability, is the binding constraint on orbital data-center scale — which is why the launch slip tracks a physics problem, not a funding one.