Half a Round, All a Strategy: Nvidia's $2.5B Bet on Thinking Machines

Late in the first week of September 2026, one number is dominating AI finance chatter: $2.5 billion. According to The Information, Nvidia is in discussions to invest roughly that amount in Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati — and to supply about half of a larger round in which the lab is trying to raise between $5 billion and $6 billion at a valuation of at least $40 billion. If the talks conclude on those terms, a company that did not publicly exist two years ago would become one of the most valuable private AI labs on the planet, bankrolled in significant part by the chipmaker whose GPUs it intends to consume by the gigawatt.

What we know about the deal

The reported structure is straightforward but unusual in its concentration. Rather than spreading the round across a wide syndicate, Thinking Machines is negotiating to raise $5-6 billion, with Nvidia expected to contribute around $2.5 billion — a single investor writing a check for about half the total. Existing backers are circling as well: TechCrunch reported on September 3 that Accel, which led the company’s record-setting $2 billion seed round in 2025, is in talks to lead a $1 billion tranche at a roughly $40 billion valuation. The Information’s reporting cites an investment firm that holds a board seat at Thinking Machines Lab, giving the figures a degree of grounding beyond pure market rumor.

The valuation target matters because of what it reverses. Earlier in 2026, Thinking Machines reportedly explored a raise at $50-60 billion — talks that collapsed, according to Value Add VC’s reconstruction, before terms could be agreed. Co-founders Barret Zoph and Luke Metz left to rejoin OpenAI in January. The $40 billion figure now on the table is simultaneously a step down from the lab’s most aggressive ambitions and a more-than-tripling of the $12 billion valuation it secured with its 2025 seed round, which drew capital from Nvidia, AMD, Cisco, and Jane Street.

Nvidia’s involvement would also deepen a relationship that is already contractual. In March 2026, the two companies announced a multi-year strategic partnership under which Thinking Machines committed to use at least a gigawatt of Nvidia-powered compute beginning in early 2027, alongside what Nvidia’s blog called a “significant investment” in the lab. The $2.5 billion now under discussion should be read against that backdrop: Nvidia backing American open-weights developers is strategic logic, not charity. Every frontier lab Nvidia funds is a durable customer for its accelerators, and every open-weights release shifts the ecosystem’s center of gravity toward infrastructure buyers rather than closed API monopolies.

The Inkling factor

What has changed since the failed $50 billion talks is that Thinking Machines finally has a product. On July 15, 2026, the lab released Inkling, its first in-house foundation model — an open-weights release that immediately reset expectations for what a US startup would give away. Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active per token, a one-million-token context window, and native multimodality. Sebastian Raschka’s architecture notes highlighted its “controllable thinking effort” — a mechanism letting developers dial reasoning depth up or down per query — as one of the design surprises. The Artificial Analysis Intelligence Index scored it at 41, three points clear of the previous leading US open-weights model, per coverage at release.

The Wall Street Journal framed the release as a deliberate “bid to loosen AI giants’ grip”: where OpenAI keeps its frontier behind an API, Murati’s lab followed the DeepSeek playbook of shipping weights that anyone can download, fine-tune, and self-host. The Register was blunter, headline-writing that the former OpenAI CTO “does what Altman won’t.” The strategic bet — articulated in the lab’s own July essay “A Safe Path to Open Weights” — is that customization, not one-size-fits-all APIs, is where enterprise value accrues, and that open weights are the credible vehicle for it.

That bet is precisely what makes the company investable at $40 billion despite the turbulence. An open-weights lab with a demonstrated frontier-adjacent model, a gigawatt-scale compute contract, and a founder who shipped ChatGPT, GPT-4, and DALL-E is a scarce asset. The same concentration that makes the round unusual — one chipmaker writing half of it — is also what makes it viable, because Nvidia’s return case does not depend solely on equity appreciation; it depends on the compute demand the lab’s roadmap generates.

Why it matters

Three implications stand out. First, the round confirms that the OpenAI-style “neolab” funding model — raise billions on pedigree and roadmap before shipping — survived its first stress test at Thinking Machines, but only barely: it took a real model release to revive the valuation story after the $50-60 billion talks died and co-founders walked. Second, it marks another leg in Nvidia’s transformation from component supplier to kingmaker-investor across the AI stack, hot on the heels of its reported $12.93 billion acquisition of Hugging Face. Third, it hardens the open-weights faction of the US AI race: with Inkling shipped, Meta’s Muse Spark 1.3 in market, and Chinese labs shipping aggressively, the question for buyers is no longer whether open weights can reach frontier quality, but which balance sheet will fund the next generation of them.

Nothing is signed. Talks at this scale routinely slip — this company has already lived through one collapsed mega-round. But if Nvidia’s $2.5 billion lands, the deal will say less about one startup’s valuation than about how AI’s compute economics now work: the entity that sells the shovels is increasingly the one funding the diggers, and it is choosing to fund the ones who give the gold away.