Starved of compute, Poolside sold itself to Nvidia

BURIED IN a letter Poolside sent its investors last week was the kind of admission most funding announcements are built to avoid. At the end of last year, the company wrote, it had a six-week window in which to raise $2 billion for a 40,000-chip cluster of Nvidia’s GB300 processors, due to come online in January. It missed the window. It lost the cluster. A startup that had raised more than half a billion dollars and pitched itself as the West’s answer to China’s open-weight leaders had been undone by the one input capital alone could not summon on demand: contracted access to Nvidia’s silicon.

Days later came the resolution. Nvidia, the world’s most valuable company, agreed to pay Poolside, a coding-model startup, roughly $6 billion to license Model Factory — the internal system Poolside built to produce its Laguna family of open-weight coding models — and to invest a further $1 billion at a $12 billion valuation, roughly four times the mark the company carried in October 2024. More than a hundred of its engineers were offered seats in Nvidia’s Nemotron program, the chipmaker’s own line of open models. Its founders, the former GitHub chief technology officer Jason Warner and Eiso Kant, would stay. Poolside would continue, on paper, to operate independently. The letter was firm that the arrangement was neither an acquisition nor an acquihire.

The house always builds

Set the two documents side by side and the transaction describes a closed loop. Poolside could not compete because it could not secure enough Nvidia hardware. Nvidia then bought the capability Poolside had built with the hardware it could not get more of, at a valuation the same shortage had suppressed. The firm that controls the scarce input absorbed the customer that scarcity had broken, at a price that scarcity had set. None of this needed malice, or even a plan. It is the ordinary physics of a market with one chokepoint and one company standing on it.

That chokepoint is the real subject of the deal. Poolside’s letter said as much: the binding constraint on frontier model-building, it wrote, is not only capital but physical data-center space and contracted compute. A lab can raise money in a week and hire researchers in a month; it cannot manufacture GB300s, or conjure the power and floor space to run them, on any schedule it controls. Those are allocated — by Nvidia, which decides who gets the newest chips and when, and increasingly by Nvidia’s own balance sheet, which has flowed into the neoclouds and model labs that then route the money back into GPU orders. The company sells the shovels, funds the diggers, and now, when a digger stalls, folds the claim back into the estate.

Nvidia has run this play before, and recently. In December, across the thin trading days after Christmas, it announced a non-exclusive licensing agreement with Groq, an inference-chip startup, worth a reported $20 billion; Groq’s founder and senior engineers decamped to Nvidia while the company carried on under a new chief executive. No press release, no regulatory filing, no acquisition — the same structure, the same emphatic language, eight months before Poolside. Analysts noted at the time that the wording kept what one called the fiction of competition alive, and that the design sidestepped the antitrust review a $20 billion purchase would ordinarily invite. Microsoft built the template with Inflection in 2024; Amazon used a version of it with Adept. Nvidia has now run it twice inside a year, on a chip company and a model company, which is to say across two different layers of the stack it already dominates.

Every independent model lab now operates on a bet it does not fully control: that it can secure compute, on workable terms, from a supplier that competes with it, invests in it, and has twice shown it will absorb the layer above its chips when the opening appears. Poolside’s Laguna models were credible — the smallest ran on a single desktop system and scored within reach of models many times their size — and credibility did not preserve the company’s independence. The point for an allocator is not that model labs are bad businesses. The equity in one is, in part, a call option on Nvidia’s willingness to keep selling it chips, and Nvidia holds the other side of that option.

There is an older shape to this. The most durable monopolies have rarely been the ones that made the best product; they were the ones that owned the thing every rival had to pass through — the rail line into the city, the pipeline out of the field, the port. Nvidia’s chips are, for now, the best available, but the deeper source of its power is that everyone building at the frontier has to route through them, and that routing is something Nvidia can meter, price, fund, and reclaim. The China framing on the Poolside deal — a powerful American open model to answer DeepSeek and Kimi — is real enough as a goal. It also supplies a patriotic gloss for a transaction whose plainer meaning is that the model layer has become something the chip layer can buy.

Poolside spent three and a half years trying to build a frontier lab and lost it in a six-week window it could not close. Nvidia paid $6 billion for what remained. What it bought was the demonstration that no one builds a frontier model without Nvidia’s permission first.