The Great Walkout: How Elon Musk Lost Every xAI Co-Founder and 50 of His Engineers

Correction (October 9, 2026): An earlier version of this article said more than 80 researchers and engineers had left; The Information reported more than 50. It also said Meta recruited xAI staff to Yann LeCun's lab (LeCun left Meta at the end of 2025), that Mira Murati raised $2 billion in early 2026 (Thinking Machines' $2 billion seed round was in July 2025), that xAI was a $50 billion lab (the SpaceX deal valued it at $250 billion) and that SpaceX's IPO could value the company above $350 billion (the merger valued the combined company at $1.25 trillion). We removed unsourced claims about safety testing, a 200,000-H100 figure for Colossus, and the recruiting pitch at Meta and Thinking Machines.

When SpaceX acquired xAI in February 2026, Elon Musk pitched the combination as the path to scaling AI — two of his most important companies merged into a single, vertically integrated force. Three months later, all 11 of xAI's original co-founders are gone. More than 50 researchers and engineers have left since the merger, through departures, firings and layoffs. The supercomputer in Memphis is still humming, Grok is still shipping updates, and Musk is still posting about AI dominance — but the people who actually built the thing have quietly walked out the door.

The Merger That Broke the Culture

xAI was founded in 2023 by a cohort of researchers who mostly came from DeepMind, Google, Microsoft, and OpenAI. The founding culture reflected that lineage: deliberate, research-oriented, comfortable with long cycles between breakthrough and product. SpaceX is the opposite. It is famous — and genuinely admired in some circles — for moving faster than any aerospace organization in history, often by discarding procedures that most engineers treat as load-bearing. When the two companies merged, those two cultures did not harmonize. They collided.

Reporting on the departures points to that collision. One source told The Information that Musk set "unrealistic deadlines" for training models, which led to corners being cut on Grok; others left as SpaceX's tender offers and expected IPO gave them a chance to cash out. The pre-training team — which handles the foundational model training that everything else depends on — shrank to a handful of people after its lead, Juntang Zhuang, departed, a loss that has particularly concerned employees. Leaders in world models, coding, and Grok voice have also left. By late March 2026, Manuel Kroiss, who led pre-training, and Ross Nordeen — the last two remaining co-founders — had left as well, completing a co-founder exodus with few precedents at a comparably valued AI lab.

Where They're Going

The departing talent did not disappear quietly into retirement. At least 11 former xAI staff joined Meta. At least seven went to Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, which raised a $2 billion seed round at a $12 billion valuation in July 2025.

The pattern matters because it represents a direct transfer of institutional knowledge away from one of the most heavily capitalized AI projects in the world. These are not only junior engineers frustrated with bureaucracy. They include the people who led pre-training, coding, world models and voice — the teams that determine whether a model is ready to ship. Their departure does not make SpaceXAI incompetent, but it does mean the lab is rebuilding institutional knowledge at exactly the moment it is trying to accelerate.

What's Actually at Stake

The stakes here extend beyond hurt feelings and LinkedIn updates. SpaceXAI is still developing Grok, competing directly with OpenAI's GPT models and Google's Gemini. It operates the Colossus supercomputer in Memphis, one of the largest AI training clusters ever assembled. And SpaceX has been signaling for months that a public offering is being prepared; the merger itself valued the combined company at $1.25 trillion.

A talent exodus at this scale introduces real risk into each of those threads. Model quality depends on the people who understand what went wrong in previous training runs. Training infrastructure at Colossus scale requires engineers who built and debugged it from the start. And institutional investors doing IPO diligence tend to notice when a company's entire founding research team has left, most of it since the start of the year. Musk publicly acknowledged in March that xAI "was not built right" the first time around and is being "rebuilt from the foundations up," framing the SpaceX integration as a rebuild rather than a continuation — a characterization that may be honest but is not exactly reassuring to anyone evaluating the company's technical trajectory.

The Harder Question

The SpaceXAI situation is unusual in its scale, but it points to a tension that shows up across the AI industry: the best AI researchers tend to come from environments that value intellectual autonomy and long research horizons, while the companies best positioned to productize their work tend to operate with the kind of velocity that compresses both. OpenAI navigated this — imperfectly, with significant drama — for years before the culture tipped decisively toward product. Google DeepMind has been managing it since the 2023 merger that joined DeepMind's academic culture with Google Brain's engineering culture. The consensus is that it takes years to find an equilibrium, and that the equilibrium often involves losing people in the process.

Musk has never been interested in that kind of patience. The question is whether, at the frontier of AI capability, that impatience costs more than it buys.

Sources: TechCrunch (May 14) · TechCrunch (March 28) · The Information · The Next Web · CoinDesk