The valuation whiplash at Thinking Machines Lab is a masterclass in how fast the AI market changes its mind. Mira Murati's outfit, founded in early 2025 by the former OpenAI chief technology officer and a team of roughly two dozen OpenAI veterans, is back in the market negotiating to raise at least 1 billion at a pre-money valuation of about 40 billion, The Information reported on September 3. Existing backer Accel is in talks to lead the round, and Nvidia has discussed joining. Less than a year ago, the same company was chasing a 50-to-60-billion valuation. Those talks collapsed in January. This is the third price tag Thinking Machines has floated in just over a year, and the second time it has had to come back to investors with a smaller number and a straight face.

The Deal on the Table

The shape of the round matters as much as the size. The Information, citing a person familiar with the matter, says Thinking Machines is negotiating to secure at least 1 billion at a pre-money valuation of roughly 40 billion. Accel, an existing backer from the seed round, is in talks to lead. Nvidia, which has been in the cap table since day one, has held discussions about joining as well. Neither company commented publicly when TechCrunch asked. That silence is standard while a round is still being negotiated, but it also says the company learned something from January: do not announce a valuation until the money is signed.

Where would the capital go? Standard frontier-lab answers: training models, renting compute, and hiring. The difference is that Thinking Machines is less than two years old and, by the standards of the labs it competes with, 1 billion is pocket change. OpenAI closed the largest private funding round in history earlier this year at roughly 122 billion. Anthropic pulled in a 30 billion Series G. Thinking Machines is trying to build a frontier-adjacent lab on a round that is 1 percent of OpenAI's last check.

The Valuation Arc: 12 Billion, 50 Billion, 12 Billion, 40 Billion

Follow the marks and you can chart the whole market. In July 2025, Thinking Machines closed the largest seed round in AI history: 2 billion at a 12 billion post-money valuation, led by Andreessen Horowitz with Nvidia, GV, Lightspeed, and Conviction Partners. Investors were not buying a product. They were buying pedigree: Murati, who steered ChatGPT through its explosive years, plus the researchers who followed her out of OpenAI.

By November, appetite had gotten bolder. The company was in talks at a 50-to-60-billion valuation, a four-to-five-fold jump in four months, on the strength of one shipped product, Tinker, its model customization platform. That round never closed. When the dust settled in January, the private mark reset all the way back to the 12 billion seed price.

Now comes the third number: 40 billion at the low end, a fourfold increase over the 10 billion pre-money of the seed, but a clear discount from the 50-plus billion the company wanted last year. Markets do not usually hand out do-overs like this. The fact that Accel is willing to lead tells you how badly the top tier of venture capital wants exposure to another independent AI lab.

What Killed the 50 Billion Round: A CTO Walked Out Mid-All-Hands

The January collapse was not a macro event. It was a people event with a specific, almost cinematic sequence. Barret Zoph, co-founder and chief technology officer, left the company in the middle of an all-hands meeting. Less than an hour later, OpenAI announced his return, along with fellow Thinking Machines co-founder Luke Metz and researcher Sam Schoenholz. Lilian Weng, another founding researcher, left for OpenAI in July.

Read that again. A company trying to raise at 50 billion had its CTO poached back by the old employer in the middle of the internal meeting where employees were supposed to hear about the future. Whatever valuation conversations were alive died in that hour. When the people who built the credibility of the round walk out the door, the number walks out with them. The lesson for every founder in this cycle: pedigree rounds are only as strong as the pedigree that stays in the building.

The Business Under the Hood: Free Weights, a Paid Toll Booth

What has changed since January is a product, not a balance sheet. In July, Thinking Machines shipped Inkling, its first family of open-weight models, released under Apache 2.0. Inkling is a 975-billion-parameter mixture-of-experts model with 41 billion active parameters, natively multimodal across text, images, and audio. Inkling-Small followed at the end of July, 276 billion parameters with 12 billion active. Both are free to download.

The revenue engine is Tinker, the fine-tuning service launched in October 2025. Customers pay usage-based compute fees to adapt Thinking Machines models to their own proprietary data. Bridgewater Associates has been named as a client. Annualized revenue now tops 100 million, according to a source with knowledge of the company's finances; some reports put the run rate at a few hundred million. The model in one line: give away the brain, charge for the customization. Open weights are the marketing. Tinker is the toll booth, and that strategy depends on having a lot of compute to rent, which is exactly why Nvidia matters so much to this company.

The 400x Revenue Question

Here is where the number gets uncomfortable. A 40 billion valuation against a revenue run rate north of 100 million is a multiple above 400 times revenue. There is little recent precedent for that in this cycle, and there is a reason: sane markets do not usually price anything at 400 times sales.

Do the comparison math. Anthropic is reportedly at a 380 billion valuation with around 30 billion in annualized revenue, roughly 13 times revenue. OpenAI's 852 billion mark against its reported revenue is a high multiple, but it is in a different galaxy from 400. Enterprise software companies, even great ones, trade at 10 to 20 times revenue. A 400x price means investors are not buying the income statement. They are buying distribution, optionality, and the hope that fine-tuning open models becomes a default enterprise behavior. It also means the round is priced for a future where Thinking Machines grows revenue into the billions. That is a bet on Tinker becoming the AWS of model customization. It may be right. It is not cheap.

Nvidia Is the Real Story in Every AI Deal

Nvidia's potential participation is worth more than the check. In March, the chipmaker announced a multiyear partnership with Thinking Machines that includes deploying at least one gigawatt of Nvidia's upcoming Vera Rubin systems, plus an undisclosed investment. Training Inkling already ran on Nvidia GB300 NVL72 systems. The company is, in a very real sense, built on Nvidia's roadmap.

The same week this report broke, Nvidia confirmed its 12.93 billion acquisition of Hugging Face, the distribution hub for open-weight models. Put those two moves together and the strategy is unmistakable: Nvidia wants to own every layer of the open-model economy, from the chips that train the weights to the platform that distributes them. Thinking Machines gives away the models. Nvidia sells the shovels, the land, and the tolls on the road to the gold rush.

What This Means: Open Weights Are an Infrastructure Strategy Now

Strip away the drama and the real signal is strategic. Thinking Machines is betting that the future of enterprise AI is not one monolithic model from a single vendor, but customized open models tuned on proprietary data. Murati has called it collaborative general intelligence. The positioning matters more than the label: if you cannot outspend OpenAI and Anthropic on frontier scale, you differentiate on control, customization, and trust.

That bet now has a price tag: 40 billion. It is a discount to what the founders wanted, but it is still a massive number for a company whose public products are less than three months old and whose founding team has thinned out. The market is effectively saying it will believe the product, not the pedigree. After January, that is exactly the right discipline.

What Comes Next

The questions this round has to answer are simple and brutal. Does it close at 40 billion, or does another shoe drop first? Does Nvidia convert its discussions into an actual check, and does that deepen the dependency or just the balance sheet? When does the next model ship, and does it prove the research bench is still deep after the departures?

And the biggest one: can a 400-times-revenue company ever grow into a valuation that high, or is this the moment the open-weights story gets stress-tested the way the closed labs were stress-tested on compute costs? Thinking Machines is the test case for whether an open-weight lab can be a real business at frontier-adjacent scale. If it works, every enterprise AI strategy in the world gets a new template. If it does not, the next round will carry a fourth valuation, and the number will be smaller still.

Either way, Mira Murati just gave the market a rare gift: a clean, public experiment in what an independent AI lab is actually worth when the hype is stripped out of the price.

— Allan Ali, Sylt.ing