Thinking Machines and NVIDIA strike multi-year, 1 GW Vera Rubin compute partnership
On March 10, Thinking Machines Lab and NVIDIA jointly announced a multi-year strategic partnership under which the lab will deploy at least one gigawatt of compute on NVIDIA's next-generation Vera Rubin platform, according to releases on both companies' websites and reporting by CNBC, TechCrunch and Data Center Dynamics. NVIDIA will also make an undisclosed strategic investment in the lab; neither party disclosed the dollar value of the compute commitment or the contract length, beyond confirming that Vera Rubin deployments begin in 2027 and that the agreement is structured as a long-term, supply-side relationship.
The deal builds on a relationship that began at Thinking Machines' July 2025 seed round, in which NVIDIA participated alongside lead investor Andreessen Horowitz at a $12 billion post-money valuation. Murati, formerly OpenAI's chief technology officer, has positioned the lab around custom-model research and an open-source release strategy, which she signalled at the seed announcement. NVIDIA, for its part, has been steadily extending compute-plus-equity arrangements with frontier labs — including separately disclosed agreements with OpenAI and xAI in 2025 — that bind GPU capacity to long-dated commercial agreements rather than spot purchases.
In the joint release, Murati said: "NVIDIA's technology is the foundation on which the entire field is built. This partnership accelerates our capacity to build AI that people can shape and make their own, as it shapes human potential in turn." NVIDIA founder and CEO Jensen Huang framed the deal in research terms in the same release: "AI is the most powerful knowledge discovery instrument in human history. Thinking Machines has brought together a world-class team to advance the frontier of AI. We are thrilled to partner with Thinking Machines to realize their exciting vision for the future of AI."
The scale is unusual even by 2026 standards. One gigawatt of dedicated AI compute is, by NVIDIA's own published rule of thumb cited by Trending Topics, equivalent to roughly $50 billion to $60 billion in total data-center spending — of which NVIDIA's silicon and rack systems account for around $35 billion — making this single agreement larger than most listed semiconductor companies' annual revenue. By comparison, Microsoft and OpenAI's previously disclosed Stargate phase-one site is targeting a similar gigawatt envelope; xAI's Memphis Colossus expansion sits in the same band. None of these figures imply that all of the capacity ships in 2027 — Vera Rubin volume ramps through 2028 — but the contractual GPU allocation has effectively been removed from the open market.
Reaction in the industry was mixed. Equity analysts cited by CNBC noted that the announcement, like NVIDIA's earlier OpenAI and xAI deals, blurs the distinction between supplier and shareholder in a way that pulls Vera Rubin allocation away from second-tier labs and inference customers. Open-source researchers on social media — including Soumith Chintala, who posted the announcement — focused instead on whether Thinking Machines' first product will continue to include the open-source component Murati signalled at the seed round. Sherwood News noted that Thinking Machines has yet to publicly demo a frontier model, making the gigawatt commitment a forward bet on team and trajectory.
For us at Enpo Sekai, the partnership is a clean restatement of a structural fact we have been pricing into our product roadmap since 2025: the cost of competing at the foundation-model layer has now crossed gigawatt-scale, which means a Tokyo-based studio our size has no path to that layer and no reason to want one. Our focus stays where it has been — characters, voice and persona on top of whichever frontier model is most appropriate per use case, local-first desktop builds where latency and privacy matter, character games as a vertical, and B2B engine licensing for studios and consumer-product teams that need a defensible character layer without owning the model.
We will be watching three things over the next twelve months: (1) whether Thinking Machines actually ships the open-source component Murati has now committed to twice, since that determines whether mid-tier studios get a real handle on a model trained at this scale; (2) how the gigawatt allocation maps to actual delivered hours in 2027 — if NVIDIA's Vera Rubin yield slips, the agreement is a press release before it is a fleet; (3) whether the equity-plus-supply pattern hardens into a full alignment template across the top labs, because that determines whether the inference market we sell into ends up as five vendors or fifty.
