Thinking Machines Lab

IWhat it does

Thinking Machines Lab is a private artificial intelligence company in San Francisco. It was launched in February 2025 by Mira Murati, who had been chief technology officer of OpenAI⁶. It is not listed on any stock exchange. Its shares cannot be bought on the open market, and no public filing shows its accounts. Everything below about its money comes from press reports and a few company posts, so read the numbers as estimates.

The company says it wants to make AI systems more widely understood, customisable and generally capable, and that it publishes research, papers and code¹. In practice it has shipped two kinds of thing so far.

The first is Tinker, launched on 1 October 2025 and made generally available on 12 December 2025². Tinker is a developer service for fine-tuning. Fine-tuning means taking an existing AI model and training it further on your own data or tasks so it does one job better. With Tinker the customer writes the training loop, the reward functions (the scoring rules used in reinforcement learning) and decides when to save checkpoints, while Thinking Machines runs the graphics-chip infrastructure³. The page lists customers such as Bridgewater, Glean, Exa, Mercor and Krea, plus university researchers³.

The second is its own models. On 15 July 2026 it released Inkling, an open-weights model, and on 30 July 2026 a smaller one, Inkling-Small². Open-weights means the trained model files can be downloaded and run by anyone, though the training data and code may not be shared. The company’s release post describes Inkling as a mixture-of-experts model with 975 billion total parameters (a rough measure of model size), of which 41 billion are active at a time. It says the design largely follows DeepSeek-V3 and that post-training began with a fine-tuning step on synthetic data generated by open-weights models including Moonshot’s Kimi K2.5²². Inkling is also offered through a serverless inference service, still in beta³.

What the company has not shown, as of early November 2026, is a closed frontier model sold to the public the way ChatGPT or Claude is. Its public pitch is human and AI collaboration and models people can shape themselves¹¹⁰. I cannot judge from outside how good its unreleased research is.

A plain way to see the company: it has two businesses of very different size. One is a software service with usage-based pricing, which can be run with few staff and has a clear customer list³. The other is a bet on training large models, which costs far more and has no sales yet. The valuation talks are almost entirely about the second. A reader trying to judge the company should keep the two apart, because the first could be a decent business on its own and still not justify a price set for the second.

The papers and blog posts matter more here than at a normal software firm. The news page shows research posts on reinforcement learning for text-to-SQL, expert judgment in financial tasks and forecasting world events with a partner, Mantic, as well as safety and interactivity research grants². This is how the company recruits and builds a name, and it is also a hint at where it sees paying work: finance, search, data and code³.

IIHow it makes money

The company has not published revenue. The best public clue is usage pricing on the Tinker page: customers pay per million tokens (a token is a small chunk of text) for training, sampling and prefill, plus $0.10 per gigabyte per month for stored checkpoints, with rates varying by model³. Serverless inference for Inkling is billed per million tokens of input and output³. PYMNTS, citing The Information, describes the model as selling tools that let businesses customise AI with their own data while giving open-weight models away¹⁷.

Revenue estimates conflict. The Information, via reporter Stephanie Palazzolo, put annualised revenue in the hundreds of millions of dollars. TechCrunch, using another source, put it above $100 million¹⁶. Annualised means a recent month or quarter multiplied up to a year, so it overstates a full year if growth is lumpy. RuntimeWire points out these figures are self-reported by the company to reporters and unaudited¹⁴. I use $0.3 billion as a midpoint for the base case. That is my estimate, not a reported number.

On the other side of the ledger, costs are very large. Training and serving big models needs rented or owned graphics chips. The Nvidia agreement covers at least one gigawatt of systems, and the Financial Times, as relayed by The Next Web, said the chip arrangement alone is worth tens of billions of dollars¹¹. Jensen Huang has said a gigawatt of data-centre capacity can cost up to $50 billion¹¹. No source says how much of that Thinking Machines itself must pay, and how it is financed is not public. That gap is the single biggest unknown in valuing this company.

Two ideas help in reading the revenue range. Fine-tuning demand tends to be lumpy: a handful of large customers running big training jobs can lift a quarter’s usage and then pause. And if revenue is mostly compute resold with a margin, gross profit is much lower than for ordinary software. The company has not said which it is. I assume a modest gross margin in the base case, and that is part of why I use a terminal multiple of 10 times sales, not the higher multiples of the software names above.

Revenue and funding figures, with source type

IIIThe market and the competition

Thinking Machines competes in two markets at once. In the developer tools market, Tinker competes with fine-tuning services from the large model makers and cloud providers, and with open-source training software that customers run themselves. The Tinker page supports models from Moonshot, Z.ai, Qwen, OpenAI (gpt-oss), DeepSeek and Nvidia as well as Thinking Machines’ own³. That helps customers, and it also shows Tinker is a layer on other people’s models, not a lock on one.

In the model market, the company says it wants to build frontier models, meaning the most capable systems¹. The Daily Upside reads the Nvidia deal as a sign it intends to compete with Anthropic, Google, OpenAI, Meta and Alibaba, though the company has not said so in those terms¹². TechCrunch’s seed coverage also named OpenAI, Anthropic and Google DeepMind as rivals⁹. These competitors have far more revenue and compute. I do not have a verified, current figure for their revenue from a source I opened, so I make no numeric comparison here.

Open-weights models are a double-edged choice. They win attention and trust with researchers, and Tinker customers can build on them. But giving the model away makes it hard to charge for the model itself, so money has to come from services around it: training, hosting and support¹⁷.

How it could be disrupted from below: cheap open models are improving quickly, and tools that make fine-tuning easier are getting more common. If a customer can fine-tune a good open model on its own servers in an afternoon, a hosted service has to offer reliability, scale and support to earn its fee. Tinker’s pitch includes automatic retry if a graphics chip fails, security options such as customer-managed encryption keys, and audit logs³. These are real features for large buyers. They are also things cloud providers already sell.

How it could be disrupted from above: a large lab could bundle fine-tuning with its own frontier model at low or no extra price. That would turn Tinker from a product into a feature. I have no source saying this has happened, so I treat it as a risk, not a fact.

IVThe moat and how it could be disrupted

I rate the moat as thin. The company has real strengths: a founder with a strong network, a research team with a record at OpenAI, PyTorch and others¹, and now a large compute commitment with Nvidia¹⁰. But none of these yet produces the sort of protection that keeps a customer from leaving.

Switching costs for Tinker look low to moderate. A customer’s training recipes and reward functions are theirs, and the service runs on open-weight models, so moving to another trainer is mostly an engineering job³. There is no network effect I can see. Data advantage is limited: the company states it does not train on customer content³, which is good for trust and means it does not accumulate customer data.

People are the main asset and also the main risk. Of the six cofounders named in early coverage, four have left by early November 2026: Andrew Tulloch went to Meta, Barret Zoph and Luke Metz returned to OpenAI in January 2026, and Lilian Weng left in July 2026¹⁶. Murati and John Schulman remain, and Soumith Chintala, co-creator of PyTorch, now leads technical work¹⁶. Axios, cited by The Neuron, said hires outnumbered departures and headcount grew from about 30 to about 120¹³. Both can be true.

Compute is a form of scale, but it is rentable by anyone with money, and the Vera Rubin systems are not due until early 2027¹⁶. A deal signed before the hardware exists is a promise, not an advantage already in place¹¹. If better open models appear from larger labs, or if a rival offers cheaper fine-tuning, the case for paying Thinking Machines weakens.

What an owner could actually take out is an open question. Nobody who holds shares in a private company of this kind can sell them freely. There is no dividend and none is plausible while the company is spending on compute. Returns would come only from a later sale of shares, an acquisition or a listing. A strong founder vote¹⁶ means an outside holder cannot force any of those. Secondary sales, where early holders sell to new buyers, may be allowed only with company approval, and I have no source on that. Pooled funds such as the one in the Form D carry their own fees and terms⁴.

VThe numbers we can see

There are almost no hard numbers. The company files nothing publicly beyond one small SEC notice. That notice, a Form D filed on 27 June 2025, belongs to a special-purpose fund set up by Sydecar that sold $595,030 to 14 investors on a first sale date of 11 June 2025⁴. It tells us that small investors pooled money into the company through a fund. It does not show the company’s own accounts, and I did not find any Form D from the operating company itself in the EDGAR search I ran⁵.

Valuation history, all reported: Bloomberg said in June 2025 that the company was raising close to $2 billion at $10 billion before the money went in⁸. TechCrunch reported the round closed in July at $12 billion, led by Andreessen Horowitz, with Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street⁹. Bloomberg reported in November 2025 early talks at about $50 billion, with some figures up to $60 billion⁷¹⁷. Those talks did not close¹⁴. In September 2026, The Information reported talks for $5 billion to $6 billion at a pre-money value of at least $40 billion, with Accel in talks to lead and Nvidia giving roughly half¹⁶. An earlier version of the story had said just over $1 billion¹⁴. Sources disagree on size, and none shows a signed deal as of 3 November 2026.

Simple arithmetic on the pending round: $40 billion divided by revenue of $100 million to a few hundred million is roughly 80 to 400 times annualised sales, as The Next Web notes¹⁶. For comparison, on 2 to 3 November 2026 Palantir traded at 57 times trailing sales¹⁸, Snowflake at 18.6 times¹⁹ and Nvidia at 16.0 times²⁰. Those are enterprise value to trailing twelve-month revenue, from StockAnalysis.

Compared with others in the same line, two points stand out. The company has lost senior people faster than most young labs, and some of those moves were to rivals who are also its competition for hires¹³¹⁶. And its own post-launch momentum has been product-led: Tinker in October 2025, general availability in December, a gigawatt deal in March, Inkling in July, and one-year Tinker notes in October 2026². That is a steady record of shipping, which is a point in its favour that I weigh against the departures.

The numbers I would most like to see are cash on hand, the split of revenue between customers, gross margin on Tinker, and the cost schedule of the Nvidia commitment. None is public. Until at least some arrive, any valuation of this company is mostly a view on what a frontier AI lab is worth, not on what this firm earns.

Public comparables used in my model (enterprise value to trailing revenue, as of 2 November 2026 close)

VIWho runs it and who owns it

Mira Murati is chief executive. She was OpenAI’s chief technology officer before leaving in September 2024¹¹. John Schulman, an OpenAI co-founder, is chief scientist⁶. Press reports say Murati turned down an acquisition approach from Meta in 2025⁹¹¹.

On control, The Next Web reports that Murati holds voting control over the board on major corporate decisions, whatever the other directors do¹⁶. If that is right, outside investors, including at $40 billion, would hold little say. I could not verify it from a company filing.

Known investors from the seed round are Andreessen Horowitz (lead), Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street⁹. Conviction Partners was also named⁸. Albania’s public broadcaster RTSH reported in May 2025 that the Albanian government had decided to invest $10 million, using an amendment to the 2025 state budget²³. I do not have a verified cap table, so I cannot say what share of the company founders, staff or investors own. Expected dilution from the pending round is in my model, not in any source.

VIIValuation: is the last price supported?

Equity value in US$ billions. Not a share price: a private company has none. Latest mark: primary round, post-money (seed, reported by TechCrunch); a $40B+ pre-money round is under negotiation but not closed.

What the last valuation assumes

The $12B seed price assumed a team and a plan. The $40B+ pre-money in talks, about $45B to $46B after the money goes in¹⁶, assumes the company reaches something like my bull case: a frontier-grade model and several billion dollars of revenue within four years. On $100M to a few hundred million of annualised sales today, it is paying for the future almost entirely.

Bear, base and bull

• Bear. Annualised revenue starts at $0.15B (low end of the press range, my estimate) and grows 80%, 50%, 30%, 20% to about $0.6B by 2030. Terminal enterprise value is 6x sales, below Snowflake’s 18.6x¹⁹. Discount rate 25%. Dilution 40% from needed raises. Equity about $0.9B, shown as 1.
• Base. Starts at $0.3B (midpoint of press range, my estimate¹⁶). Growth of 150%, 100%, 70%, 50% gives about $3.8B revenue in 2030. Terminal 10x sales, a discount to Snowflake 18.6x and Nvidia 16.0x¹⁹²⁰ because growth is slowing and margins are unproven. Discount rate 20%, my choice for a venture-stage company. Dilution 35%. Equity about $12.0B.
• Bull. Starts at $0.5B. Growth of 250%, 150%, 100%, 70% gives about $14.9B in 2030, meaning a successful frontier model. Terminal 15x sales, near Nvidia’s 16.0x²⁰ and well under Palantir’s 57x¹⁸. Discount rate 18%, dilution 35%. Equity about $75B.

Comparables used

Where I come down

The last closed price, $12B, is near my base of about $12B. The talks at $40B or more are well above it: roughly three to four times my base and just over half of my bull. To justify $40B pre-money I would have to believe revenue reaches about $10B or more by 2030 and that the company stays independent and keeps its researchers. The market may still be right. Frontier AI has repeatedly produced revenue growth that models like mine did not allow for, and a lead investor with Nvidia on board can value compute access and strategic position, not just revenue. Nvidia also profits when its customer succeeds, which can make its investment less price-sensitive. I give the bull case only a modest chance. Margin of safety is my base less 25 percent, about $9B, and the talks are far above that. My model is sensitive: a one-turn change in the terminal multiple moves base value by about $1.2B, and a different starting revenue moves it in proportion.

What I cannot judge

I cannot see the company’s real revenue, costs, cash, cap table or the terms of the Nvidia investment. I cannot judge model quality beyond what the company and press say. I cannot say who pays for the gigawatt of compute. Treat my range of about $1B to $75B as wide on purpose. It is research, not investment advice.

VIIIThe IPO

No IPO plan reported. I found no report of an IPO filing or plans as of 3 November 2026. The only SEC filing I found is a small Form D for a pooled fund⁴⁵. The company is raising private capital instead¹⁶. Murati is reported to hold voting control¹⁶, which is typical of founders who stay private for a long time. No exchange, ticker, banks or timing are reported.

What to watch

• Whether the $5B to $6B round closes, at what valuation, and who leads¹⁶
• Any published revenue figure or audited accounts
• Further cofounder or senior researcher exits
• First deployment of Vera Rubin systems, targeted early 2027¹⁰¹³
• A closed frontier model, or Inkling adoption data
• Any Form D or other SEC filing from the operating company

IXWhat could go wrong

What most likely destroys this business: spending on compute outrunning revenue. If the company must pay for the gigawatt commitment from its own funds, that is tens of billions of dollars against revenue that is at most a few hundred million¹¹¹⁶. If it cannot keep raising, the commitments or the research would have to shrink.

What has to go right together. First, Tinker and Inkling revenue keeps growing quickly. Second, the company ships a model good enough to charge for at the frontier. Third, it keeps its key researchers. Fourth, it raises the capital for compute on acceptable terms. Fifth, no larger rival undercuts it. If I give each a rough 60 to 70 percent chance and treat them as loosely linked, the combined chance is nearer 10 to 20 percent. That is my estimate, not a measured figure. It is also why a venture investor can still do well: payoffs are lopsided, and one success can cover many failures.

Other risks: the funding talks might not close or might close lower, as the $50 billion talks did not¹⁴. Open-weight releases can be copied. Key staff keep leaving¹⁶. The next best alternative for the same dollar is a listed company with revenue you can read, such as Nvidia at 16 times sales²⁰, or a larger AI lab if one becomes available, at least when you want filed accounts. I cannot say any of these is better. They are simply easier to check.