Nvidia Bought Open Weights So It Can Rent Them Back

Jensen Huang bought the open-weight AI ecosystem so he can rent it back.

Eiso Kant, Jason Warner, and Margarida Garcia can explain the kinder interpretation themselves. In their letter to Poolside’s shareholders, the founders said the deal was meant to ensure that artificial general intelligence would not remain “a closed technology controlled by few,” but would instead be “one built by many out in the open.” Nvidia’s Huang made the same argument in his first post on X, titled “Open Weights and American AI Leadership,” arguing that American leadership would be judged “not by one frontier AI model, but by whether the United States builds a strong, open ecosystem.”

Read separately, the statements are coherent. Read against the structure of the transaction, they are a prospectus for enclosure.

Nvidia will pay $6 billion to license Poolside’s technology, invest another $1 billion at a $12 billion pre-money valuation, and take on more than a hundred of Poolside’s engineers. The company’s founders and operations chief will remain at a residual Poolside to pursue unspecified research projects, but the company the three of them founded is no longer a company in any meaningful sense. It is a contracting party to Nvidia’s roadmap. Newcomer first reported the arrangement, and The Wall Street Journal confirmed it.

Cui bono.

Poolside’s letter supplied the mechanism. At the end of last year, the company had six weeks to raise $2 billion for a cluster of 40,000 Nvidia GB300 servers, planned to come online in January. A GB300 is a Grace Blackwell server, one of the basic units from which a frontier-training cluster is assembled. Forty thousand of them are not equipment a three-year-old laboratory rents for the weekend. They are a commitment to chips, interconnects, data-centre capacity, power, cooling, and the vendor’s software stack for the duration of the run.

Poolside could not close the raise. The cluster was reassigned. The company had lost the compute before Nvidia licensed its technology and hired its engineers.

The open-weight version of the story is therefore the one in which the foundry owns the forge, the anvil, the hammer, and most of the smiths, while the public is invited to admire the wrought iron in the foundry gift shop. The weights are available. The capacity to produce the next generation of weights is not.

Cory Doctorow’s chokepoint capitalism describes the arrangement with suitable precision: a powerful intermediary places itself between producers and the resources they need, then extracts rent on every transit. Here the producers are independent open-weight laboratories. The indispensable resource is frontier compute. The intermediary is the firm selling the compute. The rent is paid in equity, talent, and ideology; Poolside’s residual future, its engineering workforce, and the useful ambiguity of “open” now all serve Nvidia’s position at the centre of the market.

Amazon and Audible performed the same conversion in audiobooks. A handful of platform-and-label combinations did it to recorded music. Live Nation and Ticketmaster did it to live-event ticketing.

The names change. The shape does not.

Tim Wu’s The Master Switch traced the recurring movement from open, distributed systems to infrastructure captured by a small number of firms. In The Curse of Bigness, he identified the structural question that follows: when one company controls an indispensable input, can it determine who competes, and on what terms?

The AI industry is in the enclosure phase of that cycle. Poolside’s Laguna S was one of the most popular open-weight models in the West before the deal. Its weights were downloadable, finetunable, and deployable on hardware selected by the user. That openness was real. What remained closed was the industrial system required to train its successor at the frontier.

Nvidia launched the Nemotron Coalition in March with Mistral, Thinking Machines Lab, and Perplexity, presenting it as a convening of the open-weight community rather than an acquisition of it. The distinction matters, although not as much as Nvidia’s ownership of the substrate. Coalition members train on Nvidia hardware. Their tooling depends on CUDA, Nvidia’s proprietary programming layer, and their development schedules inherit the release cadence of the company supplying both.

CUDA is not merely a brand name attached to the chips. It is the operating language in which a great deal of frontier-model training is written, along with the libraries, debugging tools, compilers, and accumulated engineering knowledge surrounding it. Leaving that layer means recompiling the low-level code that makes the chips perform the required work, retuning the tooling, and rebuilding the validation systems against which a model is measured. That is a multi-year undertaking measured against an open-weight release cycle operating in months.

The Coalition is a market of licensees.

There is evidence that Nvidia is assembling the surrounding pieces. Its recently released Nemotron 3.5 Lightning is, by independent accounts, genuinely competitive. Nvidia-backed Reflection AI, which investors have called the “DeepSeek of the West,” was reported to be in advanced talks to raise $2.5 billion at a $25 billion valuation. Nvidia is also reported to be developing a trillion-parameter Nemotron model that would be among the largest open-weight systems in the world.

The engineering is the engineering. None of that is fake.

The question is not whether Nvidia’s engineers are competent. The question is whether the substrate on which they work is contestable. The engineering becomes more impressive as consolidation makes it less contestable.

The Poolside agreement is only one layer of the same structure. Nvidia arranged $500 billion in AI financing with Wall Street firms. Filings from the largest buyers of its chips show roughly $3 trillion in off-balance-sheet AI commitments. Nvidia-backed SB Energy agreed to a ten-gigawatt Ohio data-centre arrangement with OpenAI. These are not separate stories. They are the same story at different layers: one firm increasingly sits between the capital financing frontier AI, the chips training it, the data centres hosting it, the electricity powering it, and the open-weight ecosystem distributing it.

Huang is right about the dependency. He is not the firm that should be trusted to define what “open” means, because his firm is the dependency.

The Chinese open-weight systems complicate the picture in precisely the way political rhetoric tends to flatten. DeepSeek, Moonshot’s Kimi, and Z.AI’s GLM have narrowed the gap with the American frontier to a matter of months. That is a genuine competitive and geopolitical problem, and it deserves a serious response. OpenAI, Anthropic, and Google’s research arm are also designing custom silicon; if those efforts succeed at scale, they could reduce the largest customers’ dependence on Nvidia and place downward pressure on the price of the indispensable input.

This is the market pressure capable of disciplining Nvidia. It is also not helpfully timed for an independent laboratory whose next training cluster had to be financed inside a six-week window. The closed laboratories are pursuing their own chokepoints. They are not building an open-weight alternative that the Poolside founders and Huang’s post describe.

Two chokepoints do not an open ecosystem make.

The list of beneficiaries is short. Companies writing applications on Nemotron and the technology licensed from Poolside remain customers of the same supplier. Buyers dependent on Nvidia’s GPUs can continue buying them, because the independent frontier that might have produced a substitute is being absorbed into Nvidia. The public, meanwhile, retains the comforting instruction that open weights are the safeguard against closed-model capture.

Kate Crawford’s Atlas of AI supplies the necessary physical correction. Artificial intelligence is not a cognitive system floating free of material support. It is a logistical-extractive system built from minerals, energy, data-centre capacity, training labour, and chips. The actual power sits where the chips are.

There is no reason to doubt the Poolside founders’ sincerity. Founders frequently believe the best about the company closing the cheque. There is every reason to doubt the structure. Open weights controlled by a single upstream supplier are one artefact; open weights produced by many independent upstream suppliers are a market. The first is a product Nvidia offers. The second is a market in which Nvidia must compete.

The foundry has bought the mill.

There is a Canadian parallel, and it is less comforting than the familiar claim that national-security policy requires concentration. In Commissioner of Competition v. Rogers/Shaw, an 18-day trial produced 40 lay and expert witnesses. On December 30, 2022, the Competition Tribunal dismissed the Commissioner’s application to block the transaction. The divestiture of Shaw’s wireline assets to Quebecor and Vidéotron was governed by an agreement negotiated between the parties rather than imposed as the Tribunal’s remedy, while the wireless concentration at the substantive centre of the case went unaddressed. The Federal Court of Appeal dismissed the Commissioner’s appeal on January 24, 2023.

That record is not proof of capture. It is something narrower and more useful: a demonstration that formal approval can coexist with the non-resolution of the competition question. The Canadian lesson for AI infrastructure is that a regulator should not confuse the existence of conditions with the existence of a remedy.

The structural pattern is older than artificial intelligence. The same Canadian industrial belt saw Co-Steel, Stelco, Algoma, and Dofasco pass through leveraged ownership, restructuring, and the steady conversion of existing capability into a larger firm’s return. In 1995, Gerdau bought Manitoba Rolling Mills, the Selkirk bar mill where I grew up. My father kept his job. Several uncles did not. He worked another 16 years and retired in 2011 with a pension substantially reduced from the promise attached to the earlier bargain.

Not every promise was revoked. The bar kept rolling. The people who had built the place lost control of how it would be used.

Nvidia is not loading Poolside with acquisition debt, selling its real estate, or cutting its pension obligation. It is licensing intellectual property, purchasing equity, and moving the engineering workforce into a larger roadmap. The mechanism differs in form. The operational result is recognisable: the venture with real capability becomes a contracting party to the larger firm’s plan. The thing its people built gets dissolved into the acquiring organisation.

There is no need to declare the open-weight ecosystem doomed. Its models are real. Its engineering is real. The rapid release of genuinely useful model weights is one of the substantial successes of the current investment cycle. The open-weight answer to DeepSeek and Kimi is also a national-capacity question deserving of an answer more serious than leaving the supplier of the indispensable input in charge of assembling the public alternative.

Proceedings at the US Department of Justice, the Federal Trade Commission, the European Commission under the Digital Markets Act, and the forthcoming public consultation on competition in AI infrastructure all touch on parts of the question. Bipartisan members of Congress have asked the agencies to treat the chip layer as a competition problem. None of this will produce a structural remedy inside the six-week financing window that determined whether Poolside remained an independent frontier laboratory.

Lina Khan’s account of vertical foreclosure is useful here because the transaction is not a conventional merger between horizontal competitors. Nvidia has not bought another chipmaker. It has bought access to downstream technology, engineering labour, and an open-weight narrative while controlling the upstream input required by its rivals. The foreclosure can operate without Nvidia becoming the only laboratory in the market; it is enough that Nvidia decides which laboratories can obtain the capacity to become credible competitors.

The remedies follow from the mechanism. Absorbed technology should be licensed on non-discriminatory terms. Nemotron Coalition members should have public training-compute pricing schedules rather than private arrangements whose terms depend on the supplier’s approval of their roadmaps. Future distress-driven acquisitions in the AI-infrastructure stack should receive presumptive structural review rather than being treated as ordinary hiring or intellectual-property transactions.

Public and cooperative compute capacity, open standards, and a genuine ability to inspect and repair research hardware would lower the entry price in the same way. Interoperability is not a decorative addition supplied after a supplier has become indispensable. It is how an indispensable supplier becomes contestable again.

The consultation may not contain all of those questions. Deadlines are the only part of regulatory processes the regulated actually respect, and submissions are the only part of the regulatory record that subsequent governments have to read.

To be fair to Huang, he has described the dependency more candidly than most of his competitors. His position is honest in the limited sense that the open ecosystem he describes runs on Nvidia’s hardware, software, financing, and infrastructure. The remaining question is whether that dependency is a feature the country intends to preserve.

The bill of materials does not care how the press release reads. The weights may be open; the forge is not.