A data centre doesn’t just need electricity. It needs cooling, land, fibre connections, somewhere to physically sit. SpaceX and Nvidia are trying to answer all of those constraints at once with an idea that would have sounded like science fiction just a few years ago: moving part of the compute behind artificial intelligence directly into orbit, where the sun barely sets for most of the time and the vacuum of space works as a free cooling system.
The project is called Starmind, announced in early August as an exclusive partnership between the two companies. Nvidia will supply the compute payload for the first-generation satellites, named Starmind AI1, built on the Vera Rubin NVL72 platform, which pairs the new Rubin GPUs with Vera CPUs designed to handle the heavy workloads of training and running AI models. Each AI1 satellite, according to early specifications, is expected to deliver around 120 kilowatts of sustained compute power, peaking at roughly 150 kilowatts.
The satellite’s planned dimensions give a sense of the project’s scale: thirty metres tall once deployed, a seventy-five-metre wingspan, designed to operate in a sun-synchronous orbit, the kind that keeps it almost continuously exposed to sunlight, so the panels stay powered without the nightly gaps that ground-based infrastructure has to work around. Cooling happens through radiation, using the vacuum of space instead of the water or air systems that conventional data centres rely on, systems already drawing criticism in many parts of the world over water use.
SpaceX’s timeline puts the first two prototypes in orbit as early as the first months of 2027, followed by the start of industrial-scale production before the end of that same year at the company’s Gigasat facility in Bastrop, Texas. But the number that best captures the scale of the ambition is a different one: SpaceX has already filed with the US Federal Communications Commission for a constellation that, at full build-out, could reach as many as one million AI compute satellites, spread across altitudes of 500 to 2,000 kilometres.
A figure of that size raises more than one question, starting with orbital traffic. Low Earth orbit is already crowded with tens of thousands of telecommunications satellites, and adding compute infrastructure at this scale opens up questions of orbital space management that go well beyond technical feasibility. The economics aren’t settled either, starting with the real cost of launch and maintenance compared with ground-based data centres, nor are the questions around the security of data processed in an environment that is, by definition, far harder to inspect or repair than a warehouse on the ground.
For now, Starmind remains a project on paper, albeit one backed by two of the most highly valued companies in their respective fields. If it works, it could redefine where the artificial intelligence we already use every day actually gets processed, literally shifting the centre of gravity of global computing a few hundred kilometres upward.
