There is a particular kind of room in San Francisco that exists only when a former research lab executive sits down with a senior partner at a tier-one venture firm and the meeting is, in fact, the deal. The room is small. It is usually upstairs. The partner is not pitching. The executive is not pitching. The two of them are, by the time the meeting happens, mostly walking through the shape of an agreement that both sides already know they are going to sign. The terms get rearranged. The board composition gets rearranged. The next round gets foreshadowed. The valuation, by the time the press release goes out, is a function of what both rooms wanted the headline to say.
The room in which Andreessen Horowitz committed to lead the seed round of Thinking Machines Lab was, by every account that has been printed since, a room very much like that one. Mira Murati had been a public person for two years already, by the time she walked out of OpenAI in the fall of 2024. She had been the company’s chief technology officer through the GPT-4 launch and the Dev Day announcements and the Sam Altman board firing and the five-day reinstatement that briefly made her the company’s interim CEO. She had, in a quiet but undeniable sense, accumulated more operating capital — the social, professional, and reputational kind — than almost any researcher of her cohort. When she left, the question was not whether she would raise. The question was how much, and from whom, and at what speed.
The answer, when it came, was $2B at a $12B valuation, in a seed round led by Andreessen Horowitz with participation from Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street, plus a $10M investment from the Government of Albania, her birth country, which appears in the round materials less as a financial signal than as a personal one. TechCrunch reported the round on July 15, 2025. CNBC ran the parallel story the same day. The round was, in raw-dollar terms, the largest seed in the history of venture capital. It was also a seed round in only the most charitable sense of the word. By the time the term sheet was signed, the company had been in existence for five months. It had announced a small handful of senior hires. It had not announced a product. It had not announced what kind of product it intended to announce. It had, in some specific way that the venture industry seemed to have decided was sufficient, announced Mira Murati.
"The thing about a seed round priced at twelve billion dollars is that it is not, in any meaningful sense, a seed round. It is a strategic bet on a person, made at the speed and on the terms that a strategic bet on a person gets made. Calling it 'seed' is a polite fiction. The polite fiction is the point."
By the time the rumors of the second round started circulating, in the fall of 2025, the polite fiction had become harder to maintain. Bloomberg reported on November 13, 2025 that Thinking Machines Lab was in talks for a new round at a $50B valuation — roughly four times the seed price, in roughly four months. The lab had, by that point, released a single research blog post. It had not released a model. It had not released an API. It had not released, by any reading that a customer or a developer could parse, anything that an outside observer could try.
What the lab had done, in those months, was hire.
The brain trust
The first thing to understand about Thinking Machines Lab is that it is, in some structural sense, an alumni-class company. The pattern is familiar from the post-Google generation of AI startups — the Mistral cohort, the Inflection cohort, the Adept cohort. You leave a frontier lab in a senior research role. You take with you a small group of people who have been working with you for years. You raise on the basis of the group’s collective bench depth, not on the basis of any particular product thesis. The bet, for the venture investors, is that a group of senior researchers who have repeatedly shipped frontier-grade work will continue to ship frontier-grade work in a new building.
The roster Murati assembled was, by even the most cynical reading, the strongest of the post-OpenAI exoduses. John Schulman, the co-author of the PPO algorithm that underpins reinforcement learning from human feedback, joined as chief scientist. Barret Zoph, the former head of post-training at OpenAI, joined as a co-founder. Lilian Weng, who had run the OpenAI safety team and whose technical blog is widely treated as a canonical reference in the post-training community, joined the technical staff. Bob McGrew, the former vice president of research at OpenAI, joined as an advisor. So did Alec Radford, the lead author of the original GPT papers. The roster was, in the most literal sense, a substantial fraction of the research talent that had built GPT-3 and GPT-4.
The pattern was not subtle. Murati had not, by all available reporting, set out to assemble the most loaded research roster in private AI. She had set out, in the version of the story she has been willing to tell publicly, to build a company that took post-training and human-AI collaboration seriously as research problems, and the roster was the function of who she could persuade to come with her. The persuasion, by all accounts, did not require a pitch. The people she called were, by and large, people who had worked with her for the better part of a decade. The company existed before it was named, in some sense, because the network existed before the company.
The optics, however, were the optics. By the time the seed round closed, the lab was being described in the trade press as “the strongest research bench in private AI,” and the description was, on the available evidence, defensible. The question that lingered, and that the second round amplified rather than answered, was whether a research bench, by itself, was sufficient to justify a fifty-billion-dollar private valuation.
The thesis you have to read between the lines for
Thinking Machines Lab does not, in any of its public communications, articulate a product thesis. The company’s website, as of the spring of 2026, consists of a short statement of values, a small number of research posts, and a careers page. The values statement is unremarkable in the way that frontier-lab values statements are unremarkable — it emphasizes safety, collaboration, openness, the desire to build AI that is useful to people. The research posts are technical and narrowly scoped. The careers page lists open roles for post-training researchers, reinforcement learning engineers, infrastructure engineers, and a small handful of operations roles. There is no product page. There is no announced roadmap. There is no pricing.
What the lab has communicated, in a series of carefully placed interviews and a single Murati blog post, is a thesis that has to be read between the lines. The thesis, as best one can reconstruct it from the public surface, has three legs.
The first leg is that the next generation of frontier models will be defined less by raw scale than by the quality of their post-training pipelines — the work that happens after a base model is trained, where preference data and RLHF and constitutional methods turn a next-token predictor into something a person can actually talk to. This is, in some sense, the work that Schulman and Zoph and Weng spent the previous half-decade doing at OpenAI. The bet, in the version of the thesis that can be reconstructed from the hires, is that the post-training stack is the thing that separates good models from useful ones, and that the team Murati assembled is uniquely positioned to push it further than anyone else.
The second leg is that the relationship between humans and AI systems — the interaction layer, the collaboration layer, the layer where an agent meets a person who is trying to get something done — is itself a research problem, not a product problem, and that the company intends to treat it as one. This is the part of the thesis that is hardest to read from the outside. It is also, on the available evidence, the part that Murati has personally invested the most attention in. Her single public blog post, from the summer of 2025, framed the lab’s research agenda around what she called “human-AI collaboration as a first-class problem” — language that the AI safety community has been using for years, but that has not, until recently, been the explicit founding thesis of a frontier-lab-class company.
The third leg is the leg that the venture investors actually care about, and that the lab does not say out loud. The third leg is that closed-source frontier models will continue to be the most valuable commercial assets in AI for the foreseeable future, that the companies that own them will accrue economic returns at a scale that no open-source effort can match, and that the window to build a closed-source frontier lab on the scale of OpenAI or Anthropic is, in the spring of 2026, very nearly shut. The investors in the seed round and the rumored round are not buying a research agenda. They are buying a ticket to the closed-frontier-lab category, at the last moment when a ticket can be bought.
The criticism
The criticism of Thinking Machines Lab is, in some sense, the same criticism that has dogged every frontier lab since the seed-round-as-strategic-bet model became standard. The criticism has two parts. The first part is that a fifty-billion-dollar private valuation on a company with no product is, by any classical reading of finance, a wildly speculative bet that is being underwritten by the strength of a few personal reputations and not by any defensible analysis of cash flows. The second part is that the lab’s reluctance to ship anything in its first twelve months is itself a signal — a signal that the company is not confident that what it can ship would justify the valuation, or that the team is more interested in building a research culture than a business, or both.
The first part of the criticism is, in raw financial terms, hard to argue with. A fifty-billion-dollar valuation implies, on a standard venture multiple, an expected revenue run-rate of several billion dollars within three to five years. There is no current evidence that Thinking Machines Lab has the kind of customer pipeline that would support that. It does not, by its own statements, have a customer pipeline at all. The valuation is being justified, in the private conversations that produce these rounds, by an analogy to the trajectories of OpenAI and Anthropic — both of which were also worth substantial multiples of their then-revenue, at comparable points in their lives, and both of which subsequently grew into the valuations. The argument, in the simplest possible form, is that the next OpenAI is worth what OpenAI is worth, less a discount for the risk that it doesn’t get there. The discount, in the case of Thinking Machines Lab, is currently being marked at something like one-fifth.
The second part of the criticism is harder to dismiss. The lab has been, by frontier-lab standards, conspicuously quiet. Anthropic, at the comparable point in its life, had shipped Claude. OpenAI, at the comparable point in its life, had shipped GPT-3 and the GPT-3 API. Mistral, at the comparable point in its life, had shipped open weights. Thinking Machines Lab has shipped, by the spring of 2026, a single research blog post and a careers page. The defense, in the version of the story that the lab’s allies in the venture press will offer, is that the team is doing serious work and that serious work takes time. The critique, in the version that the lab’s skeptics will offer, is that serious work that takes time should be funded at serious-work-that-takes-time prices, and that a fifty-billion-dollar valuation is not a serious-work-that-takes-time price. It is a price that demands that the work be both serious and fast, and the lab has, so far, only committed to the first half of the trade.
The most cutting version of the critique came, predictably, from outside the venture press, in a series of posts on technical blogs and on the social network that has replaced Twitter for serious AI conversation. The version of the critique that stuck used a word that has, over the last year, become a kind of shorthand for the broader anxiety about the gap between AI valuations and AI products. The word is vaporware. The accusation is that Thinking Machines Lab is, until further notice, vaporware-pricing — a company whose valuation is a function of its talent roster and its pedigree and the venture market’s willingness to credit those things at the price of a real business, in advance of any evidence that the business exists. The accusation is not unique to Thinking Machines Lab. It is, in some specific way, the accusation that defines the 2026 frontier-lab landscape. But Thinking Machines Lab is, in the early months of 2026, the cleanest object lesson the accusation has.
"The phrase 'vaporware-pricing' has become, in twelve short months, the most-used shorthand in the AI venture community. It applies, on a strict reading, to a large fraction of the labs that raised in 2025. It applies, on the strictest reading anyone is willing to publish, to almost all of them."
What Murati does in public
Murati, throughout this, has been doing what she has always done in public, which is very little. She does not give frequent stage talks. She does not maintain an active presence on the social networks where her competitors fight out their narratives. She has done, by the count of the AI press, a small handful of interviews — most of them on background, a few of them on the record, none of them detailed enough to give an outside observer a clear read on what the lab is building. The single public blog post from the summer of 2025 is, on any reasonable accounting, the longest piece of writing she has put her name to since leaving OpenAI.
This is, in the context of her cohort, anomalous. The other founders of comparable-vintage frontier labs — Mustafa Suleyman at Inflection, Aidan Gomez and his co-founders at Cohere, the Mistral team in Paris — have all, to different degrees, accepted that the founder-as-public-figure role is part of the job. They give keynotes. They run podcasts. They sit for long-form profiles. Murati does not. She has, in some specific way, decided that the public-figure role is optional in the way that she is choosing to play it.
Her allies will tell you that the choice is principled — that she has decided, after years of being the public face of OpenAI’s product surface, to use the founder phase of her career to be a researcher and a manager and not a public figure. Her critics will tell you that the choice is strategic — that the absence of public commitments is itself an asset, because it makes it harder to falsify the thesis that the lab is doing serious work, and because it preserves optionality on what the lab eventually ships. Both readings, in some sense, are right. The two readings are not, on close inspection, very different from each other.
What the choice does, in either reading, is concentrate the company’s narrative entirely in two surfaces — the roster, and the round. The roster is the public proof that the company is serious. The round is the public proof that the market agrees. Everything else is, until further notice, deferred. The deferral is the strategy.
What the venture investors are actually buying
The fifty-billion-dollar valuation is, in some specific way, easier to understand if you look at it from the venture side of the table and not the operator side. The Andreessen Horowitz partner who led the seed round did not, in the meeting where the deal got done, evaluate Thinking Machines Lab as a business in the way that a public-market analyst would evaluate it. The partner evaluated it as one of a small handful of seats at a category-defining table. The category, on a clean reading, is the next generation of closed-source frontier labs. The seats are limited. The price of admission is the price of admission. Once the price is paid, the cost of being wrong is bounded by the size of the check. The cost of being right is, in the version of the story that the venture industry tells itself, very nearly unbounded.
This is the same calculation that produced the OpenAI investment and the Anthropic investment and the Inflection investment and, in the period before they pivoted, the Mistral investment. It is the same calculation that produced the recent rounds for Sara Hooker’s Adaption Labs, the talent-led seed that funded a smaller-models thesis on the same logic at a smaller price. It is, by the spring of 2026, the dominant logic of the frontier-lab venture market. The thing being purchased is a seat. The seat is priced by the bidding behavior of the other people who want a seat. The product, if any, is downstream.
The argument the venture industry would make for itself, on this, is that the seat-at-the-table calculus is the only one that makes sense in a market where the upside distribution is fundamentally winner-take-most. If the frontier labs converge into a small number of category-defining outcomes, then the only meaningful investing question is which of the seats you can afford to buy. The price of being wrong is the price of the check. The price of being right is a percentage of a trillion-dollar outcome. The math, on that calculus, supports buying every seat you can.
The argument against the calculus is that it has, in the history of venture capital, repeatedly produced bubbles. Every time the seat-at-the-table logic has been allowed to dominate a category, the category has, sooner or later, encountered the reality that not all of the seats survive, and the seats that didn’t survive turn out, in retrospect, to have been priced as if they would. Thinking Machines Lab might, on its first product release, prove that the seat-at-the-table logic was correct, in this case, in this lab, with this team. Or it might not. The point of the valuation, from the venture side, is that the answer to that question is, by design, deferred.
What it means for the rest of the field
The Murati round, more than almost any other event in the 2025–2026 AI funding cycle, has functioned as a marker for the rest of the field. The marker says, in a way that the field has read clearly: the price of a serious team in a serious category, in 2026, is whatever the most aggressive seat-at-the-table buyer is willing to pay. The price is now, on the available comparables, somewhere between $10B and $50B for a research-stage company with no shipped product, depending on the strength of the roster. The price is no longer, in any meaningful sense, anchored to revenue or to customers or to a product roadmap. The price is anchored to a thesis about category outcomes and a roster.
This has had two specific effects on the rest of the AI venture market, both of which are visible in the funding-round data of the first half of 2026. The first effect is that the seed-as-strategic-bet round has become the dominant deal structure for any founder with a credible frontier-lab pedigree. The second effect is that the comparable-stage companies without that pedigree have, in the same period, found it dramatically harder to raise on comparable terms. The market has, in other words, bifurcated. Pedigreed teams can raise at vaporware-pricing. Non-pedigreed teams cannot raise at any price.
This bifurcation is, in some specific way, the most important structural fact about AI funding in 2026, and it is downstream of the Murati round. The round did not invent the dynamic. But it priced it, publicly, in a way that made the dynamic harder to deny. After the Bloomberg story, every senior researcher at every frontier lab in the United States had to decide whether to stay or to leave on the terms that Murati had set. Some of them stayed. A surprising number of them, by the spring of 2026, were on the public record as having left.
The criticism, from inside
The most interesting criticism of the Murati round, in the end, has not come from the technical blogs or the AI Twitter equivalents. It has come from inside the frontier-lab community itself, in conversations that have only intermittently surfaced in the press. The criticism, as best one can reconstruct it from the off-the-record reporting that has trickled out, is not that the round is too expensive. It is that the round is, in some specific way, too easy.
The argument inside the labs goes something like this. The founding generation of AI companies — the OpenAI of 2015, the Anthropic of 2021, the early Mistral team — had to convince capital to commit to a fundamentally unproven thesis at a price that, by the standards of the time, was modest. The discipline that came from having to do that work — having to articulate a real product thesis, having to actually ship things, having to build customer relationships in advance of a competitive landscape — was, in the version of the story that the senior people at those labs will tell, the discipline that made the companies possible. The capital came on terms that demanded discipline. The discipline produced the companies.
The Murati round, on this telling, has flipped the dynamic. The capital is coming on terms that do not demand discipline. The capital is committing in advance of the discipline. The capital is, in some specific way, betting that the discipline will materialize on its own, because the team is good enough that it will. The internal critique is that this bet has not, historically, been a reliable one. Good teams that are funded too easily, the argument goes, become teams that ship slowly and argue internally and lose the urgency that produced the work that earned the funding in the first place. The Murati round, on the most-anxious version of this critique, is not a bet on the team. It is a bet on the team in a configuration that has historically been bad for teams.
This is, on its face, an unfalsifiable critique. Thinking Machines Lab might, in some specific way, prove that the easy-capital configuration produces good work as reliably as the disciplined-capital configuration did. The early evidence will arrive when the lab ships its first product. The lab has not, as of the publication of this profile, announced when that will be. The deferral is, again, the strategy.
"The thing that worries me about the frontier-lab funding environment in 2026 is not the prices. The prices are what they are. The thing that worries me is that the prices have come uncoupled from the work, and I don't know, historically, whether the work survives the uncoupling."
What happens next
There are, on a clean reading of the available evidence, three possible futures for Thinking Machines Lab, and the next twelve months will determine which one materializes. The first is that the lab ships a frontier-class model on the timeline that its valuation implicitly demands, vindicates the seat-at-the-table logic, and grows into its price. The second is that the lab ships something — a model, a product, a research breakthrough — that is interesting but not category-defining, and the valuation, in the next round, gets quietly marked down through a structure that the venture industry has refined over a decade of dealing with cases like this. The third is that the lab does not ship anything that meaningfully changes the market’s read on its position, and the gap between the valuation and the work becomes, over time, the central story.
The most likely outcome, on the available evidence, is the second one. Frontier-lab cycles are long. Post-training pipelines are difficult to build. Human-AI collaboration research is, by its nature, slow and incremental and resistant to clean product packaging. The lab is, on every available signal, doing the work it said it would do. The work is, in some specific sense, exactly the kind of work that produces useful but not category-defining results. The bet on the lab is the bet that the work will, against the base rate of comparable bets, in fact be category-defining. The base rate, historically, is not high.
What is certain is that the round will not be the last of its kind. The Murati template — strong pedigree, fast seed, no product, talent-led roster, vaporware-priced — is now the template for the next generation of frontier-lab raises. The next round of the cycle is already, in the venture press, being foreshadowed. There are more former OpenAI executives in the network than there are open seats at the closed-frontier-lab table. There are more open seats at the table than the table can support. The bidding will continue. The prices, by every signal that the venture market is sending, will go up before they come down.
What the Murati round will be remembered for, in the end, is not the dollar figure or the roster or the rumored second round. It will be remembered for the specific way it priced a thesis that, until 2025, the AI venture community had been unwilling to price explicitly. The thesis is that the next generation of frontier AI will be built by closed labs led by senior researchers with strong pedigrees and large checks. The price of admission to that future, as of the spring of 2026, is fifty billion dollars. The product is, until further notice, deferred. The bet is on the deferral.
Carter Vance writes long-form profiles on frontier AI for Frontier Bylines. Reporting for this piece drew on coverage from TechCrunch, CNBC, Bloomberg, and Fortune’s coverage of adjacent talent-led seeds. Thinking Machines Lab declined to comment.