SpaceX Bets on NVIDIA Hardware for Starmind Orbital Data Center

SpaceX is planning to launch a massive orbital data center project dubbed Starmind, which Elon Musk intends to use to decouple xAI's growth from land, cooling, and power limitations on Earth, as Engadget first reported. According to the report, SpaceX has filed an FCC application to eventually deploy up to one million satellites in Low Earth Orbit.

For the initial phase, known as the AI1 satellites, SpaceX will utilize NVIDIA's Vera Rubin NVL72 rackscale system, featuring Rubin GPUs and Vera CPUs. Musk stated during a Q2 earnings call that this architecture is the most effective and cost-efficient option available. Each AI1 satellite is designed to function as a single NVIDIA server rack, housing approximately 72 chips and producing an average of 175kW of computing power. To sustain this, SpaceX plans to use a solar panel array in a sun-synchronous orbit to generate 210kW of power, while employing 1,700-square-foot liquid radiators to vent heat into the vacuum of space.

While Musk expressed optimism for a 2027 launch, Engadget notes that critics cite technical and scaling hurdles. A primary concern involves the use of advanced chips with smaller transistors, which Dr. Benjamin Lee of the University of Pennsylvania's Department of Electrical and Systems Engineering warned are more susceptible to radiation-induced "bit flips." While companies like Starcloud and Google have tested orbital compute, Dr. Lee told Engadget that the constant need to detect and correct these data corruption errors could render the orbital data centers inefficient.

Long-term, Starmind may rely on Terafab, a $119 billion joint chip manufacturing venture between SpaceX and Tesla, though NVIDIA will provide the hardware for the immediate timeline.