SpaceX wants to run AI in orbit, and it is bringing in Nvidia‘s most powerful chips to do it. The company announced that it is partnering with Nvidia to design the compute payload for Starmind AI1, the first of a planned fleet of satellites built to act as a data center in space. Each satellite will carry Nvidia’s next generation Rubin GPUs and Vera CPUs, hardware normally found in ground based AI data centers, now headed for orbit.
In SpaceX’s words, each of the Starmind satellites will include “Nvidia Rubin GPUs and Vera CPUs for datacenter class space compute.” In plain terms, these are not simple communications satellites. They are meant to be full AI computers that happen to fly, running the same kind of heavy workloads that today fill warehouse sized buildings on the ground, only powered by sunlight and cooled by the vacuum of space.
To fit that hardware, SpaceX has had to feed it far more power. The company raised the satellite’s peak power about 67 percent to roughly 250 kilowatts, up from 150, and its average power about 33 percent to 160 kilowatts, up from 120. That jump is not arbitrary. It is enough to run a full Nvidia Rubin NVL72 rack, a single system that packs 72 Rubin GPUs, the same dense building block Nvidia sells for data centers on the ground.
The ambition behind it is enormous. This is the compute layer for Starmind, Elon Musk’s plan for a constellation that could eventually reach a million satellites, each a node in what would effectively be a distributed AI supercomputer circling the Earth. Nvidia says the space version of its Vera Rubin module can deliver up to 25 times the AI performance of an H100, the chip that powered much of the current AI boom, and Musk has said the same NVL72 rack design will run both on the ground and in space.
The logic for moving AI off the planet is power and heat. Training and running big models eats enormous amounts of electricity and throws off huge amounts of heat, and on the ground that means straining power grids and building ever larger cooling systems. In orbit a satellite gets nearly constant sunlight and can radiate heat straight into space, at least in theory. SpaceX plans to start prototype testing of the Starmind AI1 satellite in early 2027 and to begin mass production later that year if the work stays on track.
None of this is proven yet. Putting data center grade electronics in space raises hard problems of radiation, cooling, repair, and cost, and a million satellite fleet is a staggering thing to build and launch, even for the company that already flies more rockets than anyone. But the pieces are lining up. SpaceX supplies the satellites, the launches, and the power, Nvidia supplies the brains, and together they are betting that the next wave of AI infrastructure will not sit in the desert but orbit overhead. For the bigger picture of where this fits, see our guide to how the space industry works.