xAI drops JAX GPU for an in-house training framework

xAI放弃JAX GPU转向自研训练框架

SemiAnalysis says xAI dropped JAX GPU and moved to a C training framework written with Grok Build; the snippet claims xAI’s JAX stack had MFU below 10%, but the post does not disclose reproducible benchmark conditions.

Why it matters: HKR-H/K/R all pass: xAI changing its training stack is a strong hook, MFU <10% is a concrete claim, and infra cost will spark debate. Single-source tweet format and no reproducible setup keep it at 80, not P1.