SpaceXAI (formerly xAI) is bleeding talent worse than previously known. This week, three senior researchers — pre-training lead Zhuang Juntang, search post-training lead Tianyi Zhang, and voice lead David Haxton — all quit within days. Now new disclosures show that since SpaceX completed its acquisition in February, more than 50 Grok model developers have been laid off, fired, or left voluntarily. Last year the R&D team numbered over 200; today the pre-training team is down to a handful.
Multiple former employees point directly at Elon Musk's management style. During his hyper-engaged phase in AI, the team was required to attend meetings at the Palo Alto office seven days a week while facing deadlines widely considered unrealistic. To hit those targets, engineers had to skip critical steps that affected model performance. Musk claimed in April that the "model factory" was running, and promised a new base model every two weeks — but the planned 1-trillion-parameter Grok 4.4, due in early May, has yet to materialize.
The departing talent is being snapped up by rivals. Meta has hired at least 11 people; Mira Murati's Thinking Machines has taken at least 7; former code lead Beibin Li joined Chen Tianqiao's MiroMind. Meanwhile, SpaceXAI has leased all the compute from one of its large data centers to Anthropic.