Musk Reveals Grok's New Training and Inference Stack Replacing JAX and SGLang

On November 10, 2026, Elon Musk posted on X revealing progress on Grok's next-generation SI training and inference infrastructure. According to the post, SpaceX's SI team rewrote Grok's reinforcement learning and inference software from scratch, replacing JAX and SGLang with a proprietary system. The stack is written mainly in C, with a small amount of C++. Musk said the work was carried out by the engineers who built Starlink's complex satellite network, and that Starlink's engineering experience would support Grok's iteration. He added that SpaceX writes its own header files and computes time independently, including leap seconds, a capability he said most people would not grasp but that professionals in the field would recognize. In his framing, this low-level engineering accumulation is the core foundation of Grok's new AI stack. Only this disclosure has appeared so far, with no other sources confirming it, and no details yet on rollout timing or performance numbers.