Mira Murati, the chief technology officer who shipped GPT-4 and Sora at OpenAI, left the company at its peak in September 2024 and raised roughly $2 billion by July 2025 for a new research lab with no public product, no published research, and a founding team almost entirely drawn from senior OpenAI alumni. The seed round, reported as among the largest in venture history, signals a fundamental shift in how frontier AI labs are being built: the bet is on people and institutional structure, not on who can spend the most on computing power.

The story of Murati's departure and what came next has become a case study in how the AI industry's most talented researchers are rethinking what a frontier lab should look like. After spending four years at OpenAI shipping major products including DALL-E, ChatGPT, GPT-4, GPT-4o with voice and vision capabilities, and Sora, Murati announced Thinking Machines Lab in February 2025 with almost no details about what the company would actually build.

Why Did Murati Leave OpenAI at Its Peak?

The November 2023 board crisis at OpenAI shaped Murati's thinking about institutional design. When Sam Altman was removed from his position on a Friday evening, Murati was named interim CEO that same night. By Sunday, most of the lab's senior staff had signed a letter demanding Altman's return, and by Tuesday morning, the board reversed its decision. Murati's handling of the crisis was widely credited with preventing a more chaotic outcome, but the underlying structural problems remained unresolved.

In the months that followed, a pattern emerged. The lab's research mission and its product surface had drifted apart. The safety team's relationship to the deployment team remained complicated. Senior departures accelerated: Ilya Sutskever left in May 2024, Jan Leike days later, and Murati herself in September. These exits read like a slow-motion exodus from what had been the world's most influential AI research organization.

Murati's core insight was that these problems were structural, not personal. OpenAI's unusual governance structure, which combined a nonprofit 501(c)(3) board with a for-profit subsidiary, created misaligned incentives and veto points that made it difficult to move quickly or maintain a coherent research direction. She believed that redesigning this architecture from scratch could solve the problem.

What Is Thinking Machines Lab Actually Building?

When Thinking Machines was announced in February 2025, the launch post revealed almost nothing about products. Instead, it focused on three research areas: alignment, multimodal models, and human-AI collaboration. This was a research stance, not a product roadmap. The company said publicly that its first major product release is targeted for 2026, and that its early years will be research-led rather than launch-led.

The founding team tells the real story. Murati brought together John Schulman, Barret Zoph, and a long roster of senior OpenAI alumni. This was the kind of hire list that startups normally spend a decade assembling, and it arrived on day one. The $2 billion seed round at a $12 billion post-money valuation was effectively a bet on the people, not on a specific product or technology.

The market context matters here. The foundation-model layer, which powers systems like GPT-4 and GPT-4o, is rapidly commoditizing. Anthropic, Google DeepMind, Meta's open-weight teams, and Chinese frontier labs are converging on similar capabilities at similar costs. The economic value is migrating up the stack to applied surfaces like Perplexity, Cursor, and Notion, and to vertical applications built on top of these models. A new lab entering at the model layer has to either commit to enormous computing costs to match the frontier or find a research focus that bends the cost curve.

Thinking Machines' public stance leans toward the latter. The areas Murati has called out, alignment, multimodal models, and human-AI collaboration, are not the same as racing to build GPT-5. They imply a research culture that wants to ship narrower, better-aligned models rather than a single monolithic frontier capability. Whether that turns into a product the market will pay for, and whether that product can compete with OpenAI and Anthropic, which are spending tens of billions a year, is the question the next eighteen months will answer.

How to Understand the Institutional Lessons from Murati's Move

The Murati story has become an organizational case study among other technical founders. Three key lessons are being pulled out by the venture capital and startup communities:

• People Over Models: A research lab is fundamentally a company about people, not about models or computing power. Murati's founding was effectively a hiring exercise dressed up as a fundraise; the $2 billion was for the cohort, not for GPUs. The cap table matters less than the founding team you start with.
• Institutional Architecture Compounds: OpenAI's structural mismatch between its nonprofit board and its for-profit subsidiary was the proximate cause of the November 2023 crisis. Murati's reported governance design at Thinking Machines is meant to avoid that mismatch, with a clearer chain of decision-making, fewer veto points, and the founder protected from the kind of board action that almost ended Altman's tenure.
• Safety and Capability Are Not Opposites: Murati shipped DALL-E, ChatGPT, GPT-4, and Sora before she ever publicly argued the safety side of the trade. That sequence matters. Founders who try to argue the safety side without first having built the capability side find it hard to be taken seriously. Founders who build first, then argue, get a different audience.

Founders raising in adjacent categories are now asking what one venture capitalist called "the Murati question": who's coming with you? The insight is that the founding cohort is the actual product in the early years of a research lab. The governance design is a quiet structural lesson that most consumer-AI founders are now studying.

What Does This Mean for the Future of Frontier AI?

Thinking Machines is hiring aggressively in San Francisco and remotely, and the funding gives Murati a multi-year runway to pursue research without immediate product pressure. The lab's first major product release is targeted for 2026, but the real test will come in the years after that.

The founding decision has already redrawn the talent map of frontier AI research. Whether Thinking Machines becomes a generational research lab at the scale of Anthropic or larger is not yet knowable. But the bet Murati is making, that institutional design and talent density matter more than raw computing power, is now being tested at a scale that will influence how the next generation of AI labs are built.

" }