SpaceX has almost finished writing V1.0 of an in-house AI training stack in C

2026-05-29

SpaceX's in-house AI training stack makes heavy use of pipeline parallelism by exact-mapping to 220k GB300s with 800G NICs, getting as close to bare metal as possible. The potential speed improvement is over an order of magnitude. SpaceX's next goal is to write the inference stack in C for simultaneous high-speed RL across a large block of GB300s.