OpenAI Reclaims Two Thinking Machines Co‑Founders, Exposing a Deeper Talent and Governance Crisis

Key Takeaways

  • Two Thinking Machines Lab co-founders have returned to OpenAI amid dueling accounts of alleged misconduct
  • Nearly half the startup’s founding team has departed within a year, putting its valuation and fundraising strategy at risk
  • The exits highlight the structural disadvantages newer AI labs face when competing with established players for talent

The timing was striking. Less than an hour after Mira Murati announced that Thinking Machines Lab had parted ways with co‑founder Barret Zoph, OpenAI revealed that not only Zoph but also co-founder Luke Metz and early team member Sam Schoenholz were returning to their former employer. The choreography wasn’t subtle, and it immediately set off a wave of speculation across the AI industry about what had gone wrong inside one of the sector’s most closely watched young labs.

The dueling narratives that followed didn’t make the picture any clearer. Thinking Machines sources alleged that Zoph was fired for sharing confidential information with competitors, a claim circulated by multiple reporters. OpenAI, meanwhile, told employees the company had been recruiting the trio for weeks and dismissed the misconduct concerns outright. For investors trying to assess whether Thinking Machines has an internal governance problem or is being strategically undercut by an industry giant, the contradictions are almost impossible to reconcile.

What followed was an even more worrisome pattern for a startup barely a year old. The departures of Zoph, Metz, and Schoenholz add to a growing list of early exits that includes co-founder Andrew Tulloch, plus researchers Lia Guy and Ian O’Connell. Guy is reportedly heading to OpenAI as well. In practical terms, close to half the founding group has now walked away, and most have gone back to the very incumbents the company set out to challenge.

Funding pressures only amplify the visibility of this momentum shift. Thinking Machines had been negotiating a new round following a $2 billion seed raise, with aspirations of significantly increasing its valuation. That ambition is now on shaky ground. Early investors had bet heavily on Murati’s ability to recruit—and retain—elite researchers from OpenAI. Watching several of those individuals reverse course so quickly undercuts a core pillar of that thesis.

Startups in this sector already walk a narrow line. The competition for senior AI researchers has become brutal, with Meta, Google DeepMind, and OpenAI all offering compensation packages that simply don’t resemble anything seen in traditional software. Seven-figure cash offers are increasingly common. Stock options come with accelerated vesting schedules that convert to liquidity in months, not years. And with multiple AI giants hinting at potential IPOs, the liquidity gap between established players and private labs has grown even wider.

Against that backdrop, it’s worth asking how a startup—even one backed by billions—can realistically compete. Thinking Machines can offer equity upside, but equity upside in a long-horizon private company doesn’t carry the same gravitational pull as shares that may soon be liquid on public markets. The math starts to tilt hard in one direction.

Then there’s compute. While startups scramble for access to scarce Nvidia GPUs, companies like Google enjoy the advantage of internally developed TPU hardware and a vast global infrastructure footprint. For researchers whose work depends on massive training runs or specialized architectures, those capabilities matter as much as cash. Maybe more.

OpenAI, for its part, appears to have gained more than just three returning employees. The company recently lost its VP of research, making experienced talent especially valuable. And Zoph’s new placement in the applications division—not core research—hints at a broader organizational shift at OpenAI toward productization. Turning a former competitor’s CTO into a product-focused leader gives OpenAI a useful blend of institutional knowledge and external perspective. It also raises a small but intriguing question: how much insight about Thinking Machines’ approach to customization and fine-tuning now walks back through OpenAI’s doors?

That said, the most defining issue for Thinking Machines may not be the optics, the funding timeline, or even the talent drain. It’s the uncertainty. Two competing stories about why a CTO was fired or left—misconduct versus pre-planned recruitment—create a fog that’s difficult for any early-stage company to survive. Investors can tolerate risk, but they struggle with ambiguity of this kind, where either explanation points to materially different vulnerabilities.

And yet, younger AI labs face a version of this dynamic constantly. When the largest players decide to pull talent back into orbit, the force is hard to resist. It’s not just the compensation or compute; it’s the familiarity, the stability, the momentum. Once one key researcher leaves, others begin to question the long-term trajectory. Eventually, the story becomes self-reinforcing.

In the end, Thinking Machines now finds itself in a position no startup wants: navigating a highly public crisis while simultaneously trying to raise new capital and hold onto the talent that remains. Soumith Chintala’s rapid appointment as CTO helps restore operational continuity, but it also underscores how little planning existed for succession in the first place.

Whether this moment becomes a turning point or a footnote depends largely on the company’s ability to reestablish internal clarity—and external confidence—before more researchers follow the same path out the door.