SpaceX Targets Late 2027 AI Satellite Launches

Engadget reported that SpaceX's project is called Starmind and that its initial AI1 satellite is intended to use NVIDIA Rubin GPUs and Vera CPUs. The publication also reported that the satellites would send AI outputs to Earth through high-speed laser links connected to SpaceX's Starlink network.

According to Engadget, each satellite is expected to contain roughly 72 NVIDIA chips, an arrangement the publication compared with a single NVIDIA server rack. It also reported a projected average of up to 175 kW of computing power per satellite and a longer-term goal of deploying as many as one million satellites in low Earth orbit. Those figures are project claims reported by Engadget, rather than demonstrated operating results.

The concept places accelerator infrastructure in an environment with solar energy but demanding constraints around deployment and scaling. Engadget noted skepticism around the feasibility of deploying NVIDIA's most advanced hardware in space, citing technical hurdles, costs, and the challenge of scaling the constellation.

For ML infrastructure teams, the proposal is notable because it treats the spacecraft as a distributed compute node rather than solely a communications relay. Comparable distributed-compute architectures generally require careful design of workload placement, data transfer, fault tolerance, and inference latency. Whether orbital compute can compete with terrestrial data centers will depend on verified accelerator performance, laser-link throughput, power availability, heat rejection, and the cost per delivered inference or training operation.

The late-2027 target remains an announced timetable. Public reporting cited here does not establish that SpaceX has launched an AI1 prototype or published performance data from an operational orbital compute system.