Thinking Machines Reportedly Seeks $5B at $40B as Inference Sits on a Waitlist
Business Insider reports Accel plans to lead the round, which could close within days. Thinking Machines sells fine-tuning through Tinker on three usage meters and raised two of them about 50% in July.
TL;DR
- Business Insider reports Accel plans to lead a $5 billion raise for Thinking Machines Lab at a $40 billion pre-money valuation.
- We think investors are pricing a self-serve fine-tuning meter, since Thinking Machines has never published revenue and production inference is waitlisted.
- The next tests are the close, which could come within days, and whether Tinker's production inference opens beyond its email waitlist.
Thinking Machines Lab is reportedly raising $5 billion, and what it sells today is a metered fine-tuning service.
Business Insider reported that Accel plans to lead the round at a $40 billion pre-money valuation and that it could close within days, according to Tech Funding News. Accel is the only named investor so far. The company declined to comment.
Thinking Machines has never published its revenue. It sells Tinker, a self-serve service that charges by the token to train and run models, and its production inference still sits behind a waitlist. We think Accel is pricing that meter, along with the odds that cheap experiments on it turn into steady production spend.
Tinker bills on three meters and gates production behind a waitlist
Tinker's pricing page bills per million tokens on three meters: prefill, sample and train. Storage costs $0.10 per gigabyte a month. Inkling 2 Small is listed at $1.00 to sample and $1.50 to train, against $4.68 and $5.61 for the original Inkling. The company launched Inkling 2 Small on Tinker the day the raise was reported.
The page lists no price for production serverless inference. It says "If you're interested in production use, email us" and asks buyers for their expected volume, latency needs and use case.
Tinker's pricing over its first year, according to Beam's pricing history:
- October 2025: launched free
- November 2025: moved to usage pricing
- December 2025: reached general availability
- July 17, 2026: raised prefill and sample rates about 50% and train about 10%, citing compute costs
New users get $150 in credits. The July increase reflects the same pressure behind Nebius's October 1 price increase.
TechCrunch reported in September, citing an anonymous source, that the run rate was above $100 million. At the time, Accel was in talks to lead a $1 billion round at the same valuation. That puts $40 billion at a multiple of at most 400 times the reported run rate. Nvidia is reportedly in separate talks to invest about $2.5 billion, per AI Weekly's relay of the exclusive.
A GPU seller says recurring fine-tunes belong off Tinker
Beam, which rents GPUs, wrote in July that renting hardware is roughly four to five times cheaper than Tinker for small models. Beam advises moving recurring fine-tunes off Tinker once they reach production: "Starting on Tinker and migrating recurring jobs later is a real strategy." If buyers follow that advice, Tinker loses accounts at the point they become steady revenue.
Tech Funding News notes that the valuation is more than four times BRC's $8.63 billion forecast for the 2030 custom-LLM-training market. On that view, investors are paying for the models Thinking Machines builds as much as for its fees.
Beam's case covers recurring fine-tunes, the part of Tinker with a public price. A production inference contract priced on stated volume is likely harder to move to a rented GPU, but that's the product still behind the waitlist. The round closing is the near-term signal. The longer one is whether production inference opens beyond the email form.
Production inference pricing decides whether the usage sticks
Any AI vendor selling on usage meters faces the same test. Cheap tokens and starter credits fill the top of the funnel, and the revenue depends on how much of that usage reaches production at prices that survive an increase like July's.
Tinker's experiment pricing is public and has already gone up once. Its production pricing is still a form that asks how much you plan to use.
Starting on Tinker and migrating recurring jobs later is a real strategy.
Business Insider reports the round could close within days. The second test is whether Tinker's production inference opens beyond its email waitlist.
