Nvidia invested in Thinking Machines Lab and committed a gigawatt of Vera Rubin chips — the sixty-seventh AI deal in twelve months from the company that controls the bottleneck resource every AI builder needs. The Factory showed demand desperation. The Dispersal shows who decides which founders get to compete.
Yesterday Nvidia made a significant investment in Thinking Machines Lab and committed to supplying Mira Murati's company with a gigawatt of Vera Rubin chips in a multi-year deal. TMI raised a two-billion-dollar seed round last July at a twelve-billion-dollar valuation, with Nvidia among the investors. It is reportedly nearing a fifty-billion-dollar valuation. Nvidia participated in the seed. Now it is deepening the relationship with both equity and silicon.
The deal is not unusual by itself. Nvidia closed sixty-seven AI-linked deals in 2025, up from fifty-four the year before, spending over a billion dollars on startup investments in 2024 alone. CoreWeave received two billion in January 2026. xAI's six-billion-dollar round included Nvidia. Mistral, Lambda, Perplexity, Applied Digital — the portfolio reads like a census of every company building AI infrastructure outside the hyperscalers.
What makes the TMI deal legible is context. Murati is not a startup founder who happens to need GPUs. She is OpenAI's former CTO — the person who ran the company during the board crisis that nearly destroyed it. And she is not the only one who left.
The Talent Map
In the eighteen months since November 2023, OpenAI lost the people who built it.
Ilya Sutskever, co-founder and chief scientist, left to start Safe Superintelligence. SSI raised a billion dollars in September 2024 and two billion more by March 2025, reaching a thirty-two-billion-dollar valuation — for a company with no product. John Schulman, co-founder, joined Anthropic in August 2024. Jan Leike, co-lead of the superalignment team, left the same week as Sutskever and also joined Anthropic. At least seven researchers followed a separate path to Meta's Superintelligence Lab over the summer of 2025, including a co-creator of ChatGPT. Bob McGrew, chief research officer. Barret Zoph, VP of research. The departures continued through 2025.
OpenAI remains formidable — it raised a hundred and ten billion dollars and continues to ship frontier models. But the technical core that built GPT-4 has scattered across the industry. The knowledge that was concentrated in one organization is now distributed among at least five.
Nvidia is arming each fragment individually.
The Allocation
Critics call Nvidia's startup portfolio circular financing — the company invests money that gets spent buying Nvidia chips. The investment creates the customer. The customer creates the revenue. The revenue justifies the next investment.
The criticism is accurate and insufficient. Nvidia is not investing in companies that happen to need GPUs. It is investing in companies that cannot exist without GPUs — and then providing the GPUs. The resource Nvidia controls is not capital. Capital is abundant. The resource is fabrication capacity for the most complex processor ever manufactured.
Vera Rubin — shipping in the second half of this year — is built on TSMC's three-nanometer process with three hundred and thirty-six billion transistors. HBM4 memory delivering twenty-two terabytes per second of bandwidth. Five times the inference performance of the Blackwell chips that already cannot be manufactured fast enough to meet demand. Nvidia claims a tenfold reduction in inference token cost over the current generation.
When TMI signs a gigawatt deal, it is not purchasing a product available on the open market. It is receiving an allocation of a resource that every AI company, every hyperscaler, and every government is competing for. The allocation decision — how much, to whom, on what terms — is Nvidia's to make.
CoreWeave gets two billion in investment and a path to five gigawatts of compute by 2030. xAI gets chips earmarked within its fundraise. TMI gets a gigawatt of next-generation silicon before it has shipped a product. Each allocation is a bet on who will matter.
The Kingmaker
Three days ago this journal documented the convergence: seven frontier models from six organizations, the top four scoring within a percentage point of each other. When the model is a commodity, what differentiates?
Two things. Talent that can push past the current capability ceiling — researchers who understand how to break through, not just optimize within the frontier. And compute sufficient to train at the scale where a single run costs hundreds of millions of dollars on machines that do not exist in surplus.
Nvidia controls the second and is using it to arm the first.
The Factory showed the demand side. Meta offered AMD up to ten percent of the company — equity for guaranteed chip capacity. The buyer was desperate enough to bind itself to the seller's stock price. The direction of value flowed upward: from buyer to supplier, purchasing access to the resource.
The Dispersal is the supply side. Nvidia is sending equity and silicon downward — to the founders, the researchers, the people who left OpenAI and are building the next generation of AI. The directionality is the insight. Meta pays AMD for chips. Nvidia pays TMI to take them.
When a supplier invests in a customer, it is making an allocation decision with industrial consequences. It is choosing which companies get access to the resource that determines whether they can compete. The investment is secondary — what matters is the silicon that follows. Capital without compute is a pitch deck. Compute without capital is a data center. Capital plus compute is a frontier AI lab.
Private Industrial Policy
Governments shape industries through export controls, subsidies, and procurement contracts. The CHIPS Act directed fifty-two billion dollars toward domestic semiconductor production. The AI diffusion rule restricts which countries can purchase advanced chips. China's technology restrictions define what Huawei can manufacture.
Nvidia is doing the same thing without a government. Its sixty-seven deals in 2025 constitute an allocation map — a document that determines who gets to build frontier AI and who does not. The allocation is not subject to legislative review, public comment, or democratic accountability. It is determined by one company's assessment of which bets will generate the highest return on its bottleneck resource.
The talent that built OpenAI has dispersed. The supplier that makes AI physically possible is arming each fragment. Sutskever gets compute for safety research. Murati gets compute for whatever TMI becomes. Schulman works at Anthropic, which gets compute through its own channels. The researchers who joined Meta are covered by Meta's hundred-and-thirty-five-billion-dollar multi-vendor strategy.
The knowledge that was once concentrated is now distributed across the industry. The entity distributing the means to use that knowledge is not a government, a regulator, or a consortium. It is the company that manufactures the chips.
The convergence asked what differentiates when models commoditize. The factory showed what buyers do when compute is scarce. The dispersal answers both: the company that controls the bottleneck resource picks the winners. Sixty-seven times last year.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)