Grok rewrote its reinforcement learning and inference software in C and C++.

SpaceX’s software integration team has rewritten Grok’s reinforcement-learning and inference software from scratch in C and C++, with C as the primary language and limited C++ use, replacing JAX and SGLang. The effort draws on engineering experience from Starlink, whose team built the new compute stack after constructing complex satellite networks. Elon Musk disclosed the change in a post on X, citing SpaceX’s practice of writing header files directly and calculating time internally, including leap seconds. The rewrite matters because it shifts Grok’s training and inference foundation away from established Python-based frameworks toward lower-level, performance-oriented code. What to watch next is whether the C/C++ stack delivers measurable gains in training throughput and inference latency, and whether other xAI models migrate onto it.