Orbital bets on GPU satellite mesh

Space data centers sound like sci-fi, but the idea is getting specific fast. Orbital wants to put GPU-equipped satellites in low Earth orbit and sell AI inference from space. The pitch is simple enough to remember — sunlight is abundant up there, the power grid is not, and satellites can process data without waiting for a downlink. (spectrum.ieee.org) What changed this month is that the story stopped being just a pitch deck: Orbital has a 2027 test mission lined up, Cowboy Space has raised $275 million to build the launch hardware this market would need, and NASA’s Prithvi model has already shown that useful AI can run on orbiting platforms. ### What is Orbital actually building? (spacenews.com) Orbital is not talking about one giant server in space. It wants a mesh constellation of small satellites, each carrying a GPU server rack for inference workloads — the kind of AI work that turns a trained model into answers, classifications, or decisions. Orbital emerged from stealth in mid-April, is based in Los Angeles, and says its first prototype mission will launch in 2027 on a SpaceX Falcon 9. ### Why aim at inference, not training? Training frontier models in orbit is the flashy version of the idea, but inference is the lighter lift. Inference can run on smaller distributed nodes, and in some cases the customer is already in space — an Earth-observing satellite, for example, that wants to classify floods or clouds before sending anything home. (spectrum.ieee.org) That is the part Orbital seems to be targeting first: match a hard-to-power compute job with a place that already has constant solar input. ### Why does space look attractive for compute? Power is the whole argument. Orbital says each long-term satellite could deliver about 100 kilowatts, using solar panels and radiator panels each roughly the size of a tennis court. (spectrum.ieee.org) On Earth, AI infrastructure keeps running into grid limits, land constraints, and cooling costs. In orbit, you trade those for launch cost, radiation hardening, and brutal thermal engineering. Basically, you are swapping one infrastructure problem for another. ### Isn’t cooling easier in space? Not exactly. Space is cold, but there is no air to carry heat away. A hot chip cannot use fans or evaporative cooling the way a terrestrial data center can. (spectrum.ieee.org) It has to dump heat by radiation, which is why these concepts keep showing up with huge radiator surfaces. The catchy line is “free solar power.” The catch is that power and heat rejection become one packaging problem. ### Has anyone proved this works yet? A little — and that matters. IEEE Spectrum notes Starcloud already ran a similar orbital GPU test last year. Separate from the commercial race, NASA said on May 7 that its open-source Prithvi geospatial foundation model was uploaded and demonstrated on two in-orbit platforms: the Kanyini satellite and the IMAGIN-e payload aboard the ISS. (spectrum.ieee.org) That was a smaller, application-specific demo, but it showed real AI analysis can happen before data reaches the ground. ### Why is launch suddenly part of the AI story? Because rockets may be the real bottleneck. Cowboy Space — the company previously known as Aetherflux — said today it raised $275 million at a $2 billion valuation to build rockets with orbital data-center upper stages. (spectrum.ieee.org) Its point is blunt: even if orbital compute makes sense, there are not enough affordable heavy-lift options to deploy it at scale, especially if Starship and New Glenn stay constrained or delayed. ### So what changes if this works? Satellite design starts to look more like data-center design. Power budgets grow. Thermal surfaces become mission-critical. (spectrum.ieee.org) Inter-satellite links matter more because the “cloud” is now a moving mesh. And some payloads may stop behaving like dumb sensors that dump raw data to Earth — they become edge-compute clients living inside an orbital network. ### Bottom line? Orbital’s bet is not really that space is magically better for compute. It is that AI demand is now big enough to make weird infrastructure look rational. If that bet holds, the winning company will not just build a satellite — it will build the power system, the thermal system, the network, and probably the rocket too. (spacenews.com) (spectrum.ieee.org)