In February 2025, Mira Murati unveiled Thinking Machines Lab with the kind of founding team that makes VCs salivate: herself as CEO, fresh off six years as OpenAI's CTO; Barret Zoph as CTO; John Schulman as chief scientist. By July, they'd raised $2 billion at a $12 billion valuation. By January 2026, the company was hemorrhaging talent back to OpenAI, two researchers had quit via Slack during an all-hands meeting, and investors were questioning whether their checks would clear.
The Dream Team
When Murati left OpenAI in September 2024, the AI industry assumed she'd land somewhere interesting. She'd been at the center of every major OpenAI product launch, from ChatGPT to GPT-4 to Sora. She'd briefly served as interim CEO during the November 2023 board crisis. She knew where the bodies were buried.
Thinking Machines Lab assembled a roster designed to compete directly with her former employer. Barret Zoph had led OpenAI's post-training team. John Schulman was one of the original architects of RLHF, the technique that made ChatGPT possible. The founding team also included Andrew Tulloch from Meta and researchers who'd worked on some of the most influential papers in the field.
The funding came fast. By February, they were targeting $1 billion at a $9 billion valuation. By July, the actual terms had improved: $2 billion at $12 billion, led by a16z with participation from Nvidia, Accel, and others. Murati negotiated board voting rights that outweighed all other directors combined—a power structure that would become relevant when things went wrong.
The pitch was straightforward: build AI models that could be fine-tuned for specific business outcomes. In October, they launched Tinker, an API for fine-tuning language models. The product seemed to work. Google Cloud counted them as a customer. The path forward looked clear.
The First Cracks
In August 2025, a strange story emerged: Mark Zuckerberg had offered Andrew Tulloch, one of Thinking Machines' co-founders, a $1.5 billion package over six years to return to Meta. Tulloch declined. The offer was so large it seemed like a prank—more than the total funding of most AI startups.
But two months later, Tulloch left for Meta anyway. The package he eventually accepted wasn't disclosed, but the defection was the first sign that all was not well at Thinking Machines.
Behind the scenes, tensions were building. According to later reporting, Thinking Machines leadership confronted CTO Barret Zoph in summer 2025 about an alleged relationship with another employee. The confrontation happened quietly. The employee in question left the company. For months, nothing public happened.
The January Collapse
On January 15, 2026, everything came apart at once. Thinking Machines announced it had terminated Barret Zoph for "unethical conduct." Within hours, OpenAI's COO Fidji Simo announced on X that Zoph, Luke Metz, and Sam Schoenholz—three key Thinking Machines researchers—were returning to OpenAI. Simo noted the hire "has been in the works for several weeks."
The sequence of events made Thinking Machines' "termination" look like face-saving. Zoph later disputed the characterization, saying the company only fired him after learning he was leaving, and at no time cited his performance or unethical conduct until after his departure was announced.
Wired reported that Zoph had shared confidential company information with competitors. The accusation—essentially that Zoph was spying for OpenAI while employed at Thinking Machines—raised questions about what exactly had been going on during those "several weeks" of discussions.
The exodus accelerated. Sources told reporters that at least two more Thinking Machines staffers were expected to join OpenAI soon. AI researchers quoted in the coverage expressed exhaustion with the industry's "constant drama."
The All-Hands Meeting
On January 19, Thinking Machines held an all-hands meeting to address the situation. It did not go well. During the meeting, two researchers quit via Slack in real-time, their resignation messages visible to the entire company.
The optics were devastating. Investors who had written $2 billion in checks five months earlier were now questioning whether the company could raise again. The founding team that had made Thinking Machines attractive was scattering. The product—Tinker—was live, but products don't matter much when your research organization is imploding.
One observer on Mastodon called it an "implosion" and predicted the company would become "the next big acquihire by a big tech." The messy public battle—with leaked allegations and counter-allegations—made the company look unstable, regardless of who was at fault.
The Meta Question
The New York Times later revealed that Meta had discussed buying Thinking Machines Lab, though talks never progressed to a formal offer. This wasn't entirely new—Zuckerberg had discussed an acquisition as early as June 2025—but the timing raised questions.
Had Murati rejected an acquisition that might have provided a stable home for her team? Was the talent drain partly driven by employees who saw more opportunity at established labs? The reporting suggested Thinking Machines "lacks a clear product or business strategy" and had been "struggling over the past couple of months to raise a new round of financing."
The most damning detail: Barret Zoph had been in talks with OpenAI since October 2025—three months before his firing. If true, it meant Thinking Machines' CTO was negotiating his return to a competitor while still employed, while the company was still celebrating its product launch and massive valuation.
The OpenAI Reunion
For OpenAI, the talent acquisition was a coup. Zoph was assigned to lead the company's enterprise push, part of a broader reorganization. The company had successfully poached key talent from its most promising competitor without paying acquisition prices—just standard compensation packages and the appeal of working at the industry leader.
The irony wasn't lost on observers: Murati had left OpenAI to build something independent, recruited top talent away from her former employer, and then watched that talent return. The gravitational pull of OpenAI—with its resources, distribution, and momentum—proved stronger than the appeal of a well-funded startup with an impressive founder.
What Went Wrong
Thinking Machines' collapse offers several lessons, none of them new:
Star power isn't enough. Murati was arguably the most credentialed AI startup founder of 2025. She had the pedigree, the connections, the funding. What she didn't have, apparently, was a team that wanted to stay.
Speed creates fragility. The company went from announcement to $12 billion valuation in five months. That pace doesn't allow time to build culture, resolve interpersonal conflicts, or ensure alignment on strategy. When problems emerged, there was no institutional resilience to absorb them.
Talent markets are brutal. In AI, the best researchers can name their price. Tulloch could walk away from $1.5 billion and still land at Meta. Zoph could negotiate his return to OpenAI while nominally employed elsewhere. The leverage belongs entirely to the talent, and they know it.
The majors have gravity. OpenAI, Google, Anthropic, and Meta can always outbid startups—not just on compensation, but on compute, data, and the ability to ship products at scale. A well-funded startup is still a startup.
Mira Murati still controls Thinking Machines Lab. The company still has hundreds of millions in the bank. Tinker is still live. The story isn't necessarily over.
But the narrative has shifted. In February 2025, Thinking Machines was the most anticipated AI startup since Anthropic. In January 2026, it's a cautionary tale about how quickly things can unravel when your team decides they'd rather be somewhere else.
The AI industry moves fast. Sometimes it moves faster than the companies trying to ride it.