Thinking Machines exodus to OpenAI and Meta

This turnover shows that Thinking Machines is not just competing on models or product ideas, it is competing for a tiny set of people who can actually build a frontier lab. The company launched in February 2025 with a founder heavy bench from OpenAI, but by October 2025 Andrew Tulloch had left for Meta, and by early 2026 Barret Zoph and Luke Metz had rejoined OpenAI. That turns talent retention into a core operating risk, because these are the researchers and infrastructure leaders who shape model training, post training, and technical direction.1 2 3 4 5

• Thinking Machines began with six named founders, Mira Murati, Barret Zoph, John Schulman, Lilian Weng, Myle Ott, and Devendra Chaplot. When a lab this small loses even one founder, it is not like losing one engineer, it is losing part of the original map for what to build and how to build it.1
• The exits split across the two richest buyers of AI talent. Meta pulled Andrew Tulloch in October 2025, while OpenAI pulled back Barret Zoph and Luke Metz by early 2026. That matters because each company offers not just pay, but giant compute budgets, existing teams, and immediate access to frontier scale experiments.2 3 4 5
• This is the same structural problem hitting other elite lab spinouts. These startups can raise billions before real revenue, but they still sit downstream of a labor market controlled by a few hyperscaled labs. If the incumbents want specific people back, they can often outbid a startup on cash, GPUs, and prestige all at once.6 7

Going forward, the winners in frontier AI will be the labs that turn early star power into institutions that survive individual departures. For Thinking Machines, that means replacing founder magic with repeatable systems for recruiting, training, and shipping, before OpenAI and Meta can keep converting isolated exits into a permanent drain on its leadership bench.2 6 7