On Aug 4 earnings, SpaceX CEO Elon Musk announced the company will build AI infrastructure exclusively on Nvidia's Vera Rubin architecture, calling it the best AI computer. SpaceX plans to deploy Vera Rubin NVL72 rack-scale systems (codenamed Kyber) at 2 gigawatts of compute capacity by year-end 2026, scaling to 10GW by end of 2027. This endorsement came alongside a joint Nvidia partnership to design Starmind AI1, the first satellite carrying Rubin GPUs and Vera CPUs for production AI workloads in orbit, expected to launch next year.
SpaceX is now the primary AI compute landlord for Google ($920M/month for ~110,000 GPUs) and Anthropic ($1.25B/month at Colossus 1 in Memphis). Q2 2026 AI revenue jumped 213% to $2.56B; 95% derives from GPU rental. The company booked $15.8B in AI capex in Q2 alone, $23.6B in H1 2026. Musk characterized Vera Rubin's cable-less NVL72 design as a radical simplification versus standard racks, citing advantages for orbital deployment: no complex cabling, better reliability.
The orbital dimension is unprecedented. SpaceX filed with the FCC for a constellation of up to one million satellites to support Starmind, arguing that on-orbit compute offers constant solar power and avoids terrestrial zoning bottlenecks. Same Vera Rubin architecture on Earth and in space creates standardization: workloads can flow toward wherever power and cooling are cheapest. Musk expects $100B in annualized AI revenue by year-end 2026 and claims the orbital AI buildout could reach $1T annually by 2030.
For Nvidia investors, this is a landmark vote of confidence: SpaceX controls one of the world's largest captive GPU fleets and just committed to Vera Rubin as sole hardware. Analysts estimate 2–3 GW of SpaceX deployments alone implies ~$100B in incremental Nvidia revenue. For compute architects planning frontier workloads, Starmind signals a new tier of infrastructure: production AI running above the atmosphere next year, with SpaceX's capital ($28B H1 capex) and Musk's timeline-defying ambition backing the bet.