SpaceX’s homegrown C-based AI framework nears completion for Grok v5

Elon Musk revealed that SpaceX's internal team is nearing completion of version 1.0 of a custom massive-scale AI training framework written in C.

The new training stack is precisely tailored to a cluster of 220,000 Nvidia GB300 accelerators connected by 800G networking. To wring out every drop of performance, the framework is designed extremely close to the bare metal and heavily leverages pipeline parallelism. In large-scale training tasks, the pure-C architecture could potentially run over an order of magnitude faster than Google's mainstream advanced AI framework, JAX. The in-house training stack will run on SpaceX's Colossus supercomputing cluster and will be used for the full training and iteration of the next-generation Grok v5 model.