Thinking Machines at $40B: bet on what actually shipped

Accel is reportedly in talks to lead a $1 billion round for Thinking Machines at a valuation of at least $40 billion, per The Information, reported by TechCrunch on September 3. Neither company commented, so treat the numbers as reporting, not fact. The interesting part for anyone building on top of a lab is not the valuation. It is which of the lab's outputs you would still have if the lab stopped existing.

What actually happened

Thinking Machines was founded early last year by former OpenAI CTO Mira Murati. Its prior round — a $2 billion seed led by Andreessen Horowitz, with Nvidia, GV, Lightspeed and Conviction Partners participating — valued it at $12 billion. The company reportedly sought roughly $50 billion this time and is now in talks around $40 billion.

Revenue run rate is over $100 million annually. At $40 billion that is a multiple around 400x. The company has also lost co-founders: Lilian Weng and Luke Metz have gone back to OpenAI.

What it has shipped is two things. Tinker, a hosted platform for adapting models to your data. And Inkling, an open-weights model released in July.

Why vendor durability matters more than the valuation

You are not an investor. You do not care whether $40 billion is right. You care whether the thing you wired into production on Tuesday is still running next year, and a 400x multiple on a company that has already had founder departures is a fair reason to ask.

So sort your dependencies by what survives the vendor, not by how good the demo was. Open weights survive. If you fine-tuned Inkling and you hold the artifact, a shutdown, an acquisition, or a 5x price increase is an inconvenience — you move the weights to another host and keep running. Hosted platforms do not survive. Tinker is genuinely useful and it is also a service: your adapters, your job history and your integration live inside someone else's roadmap.

That is not an argument against using hosted tooling. We use plenty of it. It is an argument for knowing which column each dependency sits in, writing it down, and making sure nothing in the "does not survive" column is load-bearing for revenue. For a small shop the practical version is three questions per vendor: can I export the artifact, is there a second provider that runs it, and how long would the swap take. If the answer to the third one is "we'd have to rebuild," you have a single point of failure with a funding round attached.

Sources: TechCrunch.