SpaceX and NVIDIA are putting Vera Rubin chips in orbit

SpaceX is partnering with NVIDIA on Starmind, an orbital compute venture whose first satellite — AI1 — carries a payload built around the Vera Rubin NVL72, NVIDIA's next-gen rack-scale system succeeding Blackwell. The pitch: capture solar energy in space, run inference up there where power is free and cooling is radiative, and serve the results to customers on the ground.

The logic isn't crazy. Datacenter buildouts are now gated by grid interconnects, not chips — multi-year waits for power in Virginia and Texas are routine. In orbit you get near-continuous sunlight in the right sun-synchronous track and no water for cooling. What you don't get is cheap mass to orbit, easy heat rejection in vacuum, or serviceable hardware when a GPU dies. Starship is the only reason the mass math is even worth attempting.

The real bottleneck is downlink. An NVL72 chews through data far faster than laser or RF links can move it, so the workloads that make sense are ones where inputs and outputs are small relative to compute — batch inference, training on data already up there, satellite imagery processed in place instead of shipped down raw.

Why it matters: if orbital compute pencils out, the constraint on AI scaling shifts from megawatts to launch cadence — and SpaceX owns that.