Thinking Machines Lab Seeks $1B at $40B Valuation

Thinking Machines Lab is reportedly in talks to raise at least $1B at a pre-money valuation of at least $40B, a proposed financing that would push one of the youngest frontier AI labs deeper into the economics of chips, power, talent, and enterprise distribution. The Information reported the talks on September 3, 2026, citing people familiar with the process.

Existing investor Accel is reportedly discussing leading the round, while NVIDIA has discussed participating. No financing has closed and no final terms have been announced. The current story is a negotiation, not a completed $1B raise or an achieved $40B valuation.

The proposed deal matters because Thinking Machines is now asking investors to price a business that has moved beyond founder pedigree and technical promise. The company has released an open-weight model, built a paid customization platform, and committed to infrastructure at a scale normally associated with much larger labs. What remains unresolved is whether those assets can produce durable enterprise economics fast enough to support frontier-scale spending.

What Happened

The reported terms call for at least $1B of new capital at a valuation of at least $40B before the investment. The Information also reported that Thinking Machines generates at least a few hundred million dollars in annualized revenue, based on an unnamed source. That figure describes a current run rate, commonly calculated from a recent month, and should not be treated as audited annual revenue.

The proposed valuation would represent a substantial increase from the company's last completed financing. Thinking Machines raised $2B in 2025 at a reported $12B post-money valuation, one of the largest Seed rounds in technology history. Andreessen Horowitz led that round, with NVIDIA, Accel, ServiceNow, Cisco, AMD, Jane Street, and others participating.

The company also explored a much larger financing in late 2025 at a valuation above $50B, according to subsequent reporting, but that process did not produce a publicly announced close. The new discussions therefore arrive with a lower reported valuation target and more product evidence than investors had during the 2025 round.

From Promise to Product

Mira Murati launched Thinking Machines in February 2025 after serving as CTO of OpenAI. The San Francisco company describes its mission as building AI that extends human will and judgment. Its strategy combines frontier model training, tools for customization, interfaces for more continuous human-AI collaboration, and published research.

The first major model proof point arrived on July 15, 2026 with Inkling, an open-weight Mixture-of-Experts model trained from scratch. Thinking Machines says Inkling has 975B total parameters, 41B active parameters, a context window up to 1M tokens, and native reasoning over text, images, and audio. The company explicitly says Inkling is not the strongest model available; it presents the model as a broad, customizable base rather than a benchmark champion.

That positioning matters commercially. Open weights allow a customer to download and modify the model, while Tinker gives organizations a platform for fine-tuning models around their own data and goals. Thinking Machines is effectively betting that enterprises will pay for the tools, training, and infrastructure that help them turn a general model into an owned capability shaped by private knowledge.

The Infrastructure Behind Open Weights

The company's product philosophy favors distributed ownership, but frontier model development still depends on concentrated infrastructure. In March 2026, Thinking Machines and NVIDIA announced a multiyear partnership to deploy at least 1 gigawatt of next-generation Vera Rubin systems for model training and customizable AI platforms. Deployment is targeted to begin in 2027, and NVIDIA disclosed a significant investment without naming an amount.

One gigawatt is more than a technical footnote. It signals a commitment to operate at a level where compute access becomes a strategic relationship, capital planning becomes inseparable from the product roadmap, and infrastructure utilization matters as much as model quality. A model can be released with portable weights, but the ability to create its successor remains tied to expensive systems that only a small number of companies can secure.

NVIDIA's position is especially consequential. It is a prior investor, a strategic infrastructure partner, and a possible participant in the proposed round. That relationship gives Thinking Machines access to the hardware roadmap behind its training plans while giving NVIDIA exposure to another frontier lab built around open and customizable models.

Why the Proposed Valuation Matters

A $40B pre-money valuation would price Thinking Machines at more than 3 times its reported 2025 post-money valuation. That comparison is striking, but it should not be mistaken for a completed markup. The financing is still under discussion, and valuation only becomes an operating fact if investors agree to fund the company on those terms.

The investment case has nevertheless changed. In 2025, capital largely underwrote Mira Murati's reputation, the assembled technical team, and the possibility that another frontier competitor could emerge. Investors considering the new round can now evaluate Inkling, Tinker, the NVIDIA infrastructure commitment, company-reported model results, and a reported revenue run rate.

The harder question is how those pieces fit into a repeatable business. Open-weight models reduce dependence on metered access to a closed API, which is part of their appeal. They also allow users to run models elsewhere, leaving the originating company to earn revenue from customization, training, hosting, enterprise support, and the cadence of future releases. Thinking Machines must prove that this relationship can be valuable enough to finance the infrastructure behind it.

What This Signals for Enterprise AI

Thinking Machines is testing a different answer to a question facing every AI buyer: should an organization rent intelligence from a central provider or own more of the model and shape it around the work? Closed-model providers offer convenience and leading general performance. Open-weight providers offer control, portability, and a clearer path to adapting a model around proprietary knowledge.

For enterprises, ownership is only useful if the model can be trained, evaluated, governed, and maintained without creating another fragile internal science project. That puts pressure on the tooling layer. The valuable product may be less the downloadable model file than the system that helps a company keep teaching, testing, and operating it as the underlying business changes.

The proposed Thinking Machines round places a very large price on that transition. If the financing closes, the capital would give the company more room to train models and build the ecosystem around them. The longer record will be written in customer retention, workload economics, infrastructure utilization, and whether organizations keep paying to make the models their own.

The Market Still Has Work to Do

The Information's report supplies an important commercial signal, but the disclosure gaps remain substantial. The company has not publicly confirmed the new financing, named committed investors, disclosed final terms, or provided audited revenue. Customer concentration, contract duration, gross margins, infrastructure obligations, and the exact use of proceeds are also unknown.

Those gaps do not erase the progress from a pre-product Seed round to public models and reported revenue. They define the next stage of the story. Thinking Machines has made open weights portable; now it has to show that the business around customization can travel far enough to keep paying for the concentrated industrial stack that makes each new model possible.