Let me start with a number that matters more than any pitch deck projection: $13.5 billion.
That is the secondary market valuation of Thinking Machines Lab as of August 2026, according to Forge Global trades. The same company is now reportedly seeking $1 billion in primary capital at a $40 billion valuation โ a 3x premium over what informed institutional buyers are willing to pay for its shares on the open market. [[41]]
This is not a growth-stage discount. This is a signal that the bid-ask spread between founder expectations and market reality has become a canyon.
I have audited enough term sheets to know that when secondary markets price a company at one-third of its targeted primary round valuation, the round does not close on those terms โ or it closes with heavily structured preferred shares that dilute the narrative as much as the equity.
Let me walk through the ledger.
The Context: A Lab Built on Talent Premium
Thinking Machines Lab was founded in February 2025 by Mira Murati, former CTO of OpenAI, alongside roughly two dozen researchers poached from OpenAI and DeepMind. [[21]] The founding thesis was straightforward: assemble enough top-tier talent, back it with record-breaking capital, and build a legitimate competitor to the frontier-model oligopoly.
The execution has been real but incomplete. In July 2025, the company closed a $2 billion seed round โ the largest in Silicon Valley history โ led by Andreessen Horowitz, with participation from Nvidia, AMD, Accel, Cisco, and Jane Street. [[25]] The post-money valuation landed at $12 billion.
By October 2025, they shipped Tinker, an API for fine-tuning open-weight models. [[49]] By July 2026, they released Inkling, a 975-billion-parameter Mixture-of-Experts model under the Apache 2.0 license. [[31]] Inkling is not the strongest model on the market โ Thinking Machines openly admits this โ but it is the leading open-weights model from a U.S. lab, scoring 41 on the Artificial Analysis Intelligence Index. [[32]]
Revenue now stands at over $100 million annualized run rate, according to a source cited by TechCrunch. [[44]] Some reports place it closer to a few hundred million. [[45]]
That is real output. Real product. Real revenue.
But here is where the math breaks down.
The Core: Valuation Is a Function of Timing, Not Talent
$40 billion on $100 million to $300 million in revenue implies a price-to-sales multiple of 133x to 400x.
Let me be precise. At the high end of reported revenue โ say $300 million annualized โ a $40 billion valuation represents 133x revenue. At $100 million, it is 400x.
A company with two products, one of which is an open-weights model that generates zero direct licensing revenue, is being priced like a monopoly pharmaceutical company during a pandemic.
The open-weights decision matters enormously here. Inkling is released under Apache 2.0. Anyone can download it, fine-tune it, ship it in a product, and never pay Thinking Machines a cent. [[36]] The company's monetization pathway runs through Tinker โ its fine-tuning API โ and enterprise support contracts. That is a services-led revenue model, not a platform-moat model.
Service businesses do not command 133x revenue multiples. They command 5x to 15x. Platform businesses command 20x to 40x. Frontier-model monopolies might command 100x. Thinking Machines is none of those.
Let me reference my own audit history here. In 2021, I evaluated a similar situation โ an NFT project with a 50x floor-price-to-revenue ratio driven entirely by team pedigree and narrative momentum. I wrote a liquidation script before the floor dropped 80%. The mechanics are the same: when valuation decouples from the unit economics of the underlying business, the correction is not a question of if, but of when.
Now add the capital structure. The company has raised $2.01 billion across five rounds. [[43]] At a $40 billion valuation, the $2 billion seed investors are sitting on a 3.3x paper return. But secondary markets value the company at $13.5 billion โ roughly a 1.1x return. The primary market is offering sellers a 3x premium over what buyers will actually pay in cash. That discrepancy does not survive a down round.
The Contrarian: Why the $40 Billion Round Might Still Close โ And Why That Is Worse
Here is the counter-argument that every VC is making to justify this price.
Mira Murati was OpenAI's CTO. John Schulman, a co-founder of OpenAI, joined. The team has 208 employees drawn from the top AI labs in the world. [[46]] Nvidia has committed at least one gigawatt of compute capacity through a multi-year partnership. [[3]] The company has locked down a $471.7 million contract with Boost Run for GPU infrastructure. [[47]]
The bet is not on what Thinking Machines is today. The bet is on what it could be in 2028 โ a frontier-model leader that rivals OpenAI and Anthropic.
That thesis has surface-level plausibility. But it ignores three structural realities that I have observed across multiple AI investment cycles.
First, the talent premium is depreciating. In November 2025, Thinking Machines was in talks at a $50 billion valuation. Those talks collapsed by January 2026. [[3]] Simultaneously, three co-founders โ Andrew Tulloch and two others โ returned to OpenAI. [[50]] A leadership exodus at a company that has existed for less than 18 months is not a recovery story. It is a retention failure priced at $40 billion.
Second, the open-weights strategy caps the upside. DeepSeek proved that open-weight models can achieve frontier-level performance at a fraction of the training cost. But DeepSeek also proved that open-weight does not generate monopoly rents. No one is paying DeepSeek 133x revenue. The market for fine-tuning APIs is competitive โ Tinker already supports Qwen, Nemotron, DeepSeek-V3.1, and Kimi K2.6 alongside Inkling. [[49]] Commoditizing your own moat is a strange strategy for a company seeking a $40 billion valuation.
Third, the competitive landscape is not static. OpenAI and Anthropic continue to release proprietary models that significantly outperform Inkling on key benchmarks. [[35]] Chinese labs like DeepSeek and Moonshot AI have produced models that Thinking Machines itself used as architectural references for Inkling. [[49]] The gap between "best open U.S. model" and "best overall model" is widening, not narrowing. Being the best in a sub-category is not a $40 billion thesis.
Let me state the contrarian angle clearly. If the $40 billion round closes, it will be because existing investors โ particularly a16z and Nvidia โ are marking up their own books to signal market strength, not because the fundamental math supports the price. The same mechanics drove the $50 billion talks in late 2025. Those talks collapsed. The secondary market never agreed with that price then, and it does not agree with $40 billion now.
A round that closes at a valuation three times the secondary market price is capital structure arbitrage, not value discovery.
The Takeaway: This Is a Derivative on the Frontier, Not the Frontier Itself
I am not bearish on AI. I am not bearish on Mira Murati. I am skeptical of a valuation that requires the company to execute flawlessly for three years in an environment where three co-founders have already left.
FutureSearch projects Thinking Machines' median annualized revenue at $175 million by the end of 2027 โ roughly where it is today. [[41]] The same analysis gives the company a 13% chance of cracking the top 10 on Artificial Analysis benchmarks before 2028.
The margin of safety in this valuation is zero. The upside scenario is priced in. The downside scenario โ slower revenue growth, continued talent attrition, commoditized open-weight competition โ is not even discounted.
Liquidity is a vanishing act, not a guarantee. Secondary markets are telling you the real price. Primary markets are telling you the hoped-for price. In my experience, when those two numbers diverge by 3x, the secondary market wins.
Floor prices are just opinions with timestamps. The $40 billion valuation is an opinion. The $13.5 billion secondary print is data. I know which one I trade off.
The market doesn't care about your cost basis. It only cares about the next transaction.