SpaceX and NVIDIA Are Building a Million-Satellite AI Cloud

SpaceX picked NVIDIA to design the compute payload for Starmind, its orbital data center program, and the spec sheet is absurd. Each AI1 satellite carries Rubin GPUs and Vera CPUs, spans 70 meters tip to tip (wider than a 747), and pulls ~160kW average with a 250kW peak on battery assist — up 67% from the 150kW figure floated earlier. That's a rack-scale data center that flies.

The constellation target is up to one million satellites. SpaceX has told investors that once Starship hits regular cadence, it could add 100 gigawatts of AI compute per year — about a million tonnes of hardware annually at 100kW per tonne. Note the per year; the viral version of this claim quietly drops it.

Timing check: two AI1 prototypes are slated for early 2027, with volume production from a plant called Gigasat later that year. Nothing is in orbit yet. NVIDIA's Space-1 Vera Rubin module — 25x an H100 for space inference — is still listed as available at a future date, and NVIDIA's own space-computing launch named Starcloud, Axiom and Planet Labs as partners, not SpaceX. The hard part is heat: liquid radiator loops with redundant pumps and micrometeoroid shielding, because vacuum is a terrible heat sink.

Why it matters: if orbital compute pencils out, the ceiling on AI scaling stops being grid power and starts being launch cadence.