$40B or Nothing: Inside Thinking Machines' Second Try After the Round That Collapsed

Eight months after a $50 billion fundraising effort collapsed amid a talent exodus that became Silicon Valley’s favorite soap opera, Mira Murati’s Thinking Machines Lab is back at the table. According to reporting from TechCrunch on September 3 and The Information, the startup is in talks to raise $1 billion at a valuation of at least $40 billion, with existing seed investor Accel in negotiations to lead and Nvidia — an investor since the seed round — discussing a contribution of roughly $2.5 billion.

The number is lower than the $50-60 billion the company was chasing last November. But it is more than three times the $12 billion valuation at which Thinking Machines closed the largest seed round in AI history in July 2025. And this time, the pitch is built on a product and revenue rather than pedigree alone.

From record seed to soap opera

Thinking Machines Lab was founded in late 2024 by Murati, OpenAI’s former chief technology officer, shortly after she left the ChatGPT maker in September 2024. She was quickly joined by a roster of OpenAI veterans, including Barret Zoph as co-founder and CTO, Luke Metz as co-founder, and roughly two dozen researchers and engineers from frontier labs.

The pedigree attracted capital immediately. In July 2025, the company closed a $2 billion seed round at a $12 billion valuation — the largest seed round in AI history — led by Andreessen Horowitz with Nvidia, GV, Lightspeed, and Conviction Partners participating. At the time, skeptics noted that investors were largely paying for the résumés, not shipped products.

By November 2025, appetite had grown bolder. Reports emerged of talks for a round at $50-60 billion — a four- to five-fold jump in four months — justified by exactly one shipped product, the fine-tuning platform Tinker. Those talks never closed. What happened next reads like a television script: in January 2026, co-founder and CTO Barret Zoph announced his departure mid-all-hands. Less than an hour later, OpenAI announced his return, along with fellow co-founder Luke Metz and researcher Sam Schoenholz. The New York Times reported that roughly nine more of the company’s ~100 employees decamped to OpenAI or received offers in the following weeks. The valuation talks fell apart, and the company’s private mark reset to its $12 billion seed price. In July, founding researcher Lilian Weng also returned to OpenAI. Zoph, notably, has since moved on again — TechCrunch reported on August 27 that he is now at Google.

What changed: Inkling and revenue

What’s different in September 2026 is a product story, not a balance sheet. In July, the company shipped Inkling, an open-weight model that earns revenue through usage-based compute fees when customers adapt it to their own data on the Tinker platform. Annualized revenue now tops $100 million, according to a source with knowledge of the company’s finances cited by Tech Funding News.

That revenue figure cuts both ways. It proves the open-weight-plus-paid-compute strategy works — the company gives away model weights and charges for the infrastructure underneath them, a model pioneered at scale by Hugging Face and embraced by Meta. But $100 million against a $40 billion valuation implies a multiple above 400 times annualized revenue, with little precedent even in this AI cycle. For comparison, investors have grumbled about paying 30-50x for AI infrastructure names with billion-dollar revenue lines.

The compute story is central to the pitch. Nvidia struck a multiyear partnership with Thinking Machines in March, committing at least one gigawatt of its upcoming Vera Rubin systems and making an undisclosed equity investment of its own. Nvidia’s potential $2.5 billion participation in the new round — first reported by The Information — points to where much of the fresh capital would go: buying and operating the compute that Inkling’s business model depends on. In an AI economy where frontier labs are signing $45 billion compute deals (as Anthropic did with Nscale on August 26), being compute-constrained is now a competitive vulnerability, and being compute-backed is part of the investment case.

Accel’s conviction bet

Accel’s willingness to lead the round suggests its conviction hasn’t dimmed. The firm has committed roughly $8.5 billion across new funds this year, and Thinking Machines remains one of its highest-profile AI deals. Leading a $40 billion re-rating eight months after a collapsed round is a statement: Accel is betting that the January exodus was a one-time shock rather than a structural flaw, and that Inkling’s early revenue trajectory validates the open-weight infrastructure thesis.

There are reasons for caution. The round is not closed, and terms floated in talks have a way of moving. The valuation is a comedown from the $50 billion target — a signal that investors pushed back on the original ask. And the talent question hasn’t fully gone away: Murati remains the anchor, but the company has lost three co-founders and a founding researcher in eight months, and Google just hired its former CTO.

But the structural argument is real. Open-weight models are having a moment in 2026 — Meta’s Muse models, Alibaba’s Qwen family, and the K2 Horizon fleet released by MBZUAI on September 3 have all pushed weights open as a distribution strategy. Thinking Machines is one of the few labs monetizing open weights directly through its own compute platform. If Inkling’s revenue keeps compounding and the Vera Rubin compute lands on schedule, the $40 billion ask may look conservative. If growth stalls, the company will have raised a fortune to become a fine-tuning utility — a good business, but not a $40 billion one.

Whether $40 billion is a price investors will pay twice — after watching the first attempt collapse under a leadership exodus — is the question this round still has to answer.