Deep Dive — Nvidia's $6 billion plan to beat China at open weights

For a company that historically made its money selling the shovels, Nvidia just bought a pickaxe. The chip giant is taking the extraordinary step of licensing the model-building software of the Paris startup Poolside and hiring 109 of the people who used it — a deal people familiar with the matter say Nvidia intends to turn into one of the world's most powerful open-weight AI models, built specifically to answer Chinese heavyweights like DeepSeek and Kimi K3. The number attached to that ambition is $6 billion, and the number attached to the pivot is bigger: a chipmaker that profits from every model ever trained is now quietly positioning itself as a frontier model maker in its own right.

The report, from the Wall Street Journal's Robbie Whelan, landed Friday evening against a $600 billion market-drop memory and a very different set of incentives than the ones that pushed Nvidia into the open-weights fight a year ago. When DeepSeek's R1 torpedoed Nvidia's valuation in January 2025 — CNBC and the WSJ both covered the carnage — the thesis was that cheaper, open Chinese models would dent hardware demand. Instead, open models got cheaper and Nvidia's revenue kept climbing. The Poolside move reads as the company's answer to a subtler threat: not that open weights kill GPU demand, but that China — not the US — gets to define what the most powerful freely available models look like.

What Nvidia actually bought

Let's be precise about the terms, because three numbers matter. Nvidia is paying $6 billion for a non-exclusive license to Poolside's Model Factory — the software platform, spanning data processing, training, reinforcement learning, and evaluation, that Poolside used to build its Laguna family of open-weight coding models. On top of the license, Nvidia is investing $1 billion in Poolside itself at a $12 billion pre-money valuation. And it is extending employment offers to 109 Poolside employees, reported to be the people who actually built Laguna (a team, per CEO Eiso Kant on the Latent Space podcast last month, that numbered fewer than 70 people, with fewer than 115 across engineering and research combined). Poolside's three co-founders stay in place, and both sides stress the same point: it is "not an acquisition and it is not an acquihire."

The placement of the word non-exclusive does the real work in that sentence. Nvidia has not bought the Model Factory outright, nor exclusive rights to it — it has bought access to the capability, and Poolside keeps the software and can license it elsewhere. As analysis site AI Tech Connect put it, the structure signals that Nvidia's value is in having the capability now rather than in denying it to anyone else. The $6 billion is slated to be distributed to Poolside's investors by the end of 2027, and Poolside continues to operate independently — a big cash check going to a company's backers, rather than the usual exit.

This is the third time Nvidia has reached for this specific shape. In December 2025 it paid Groq about $20 billion to license its inference technology and hire its top engineers (Groq stayed independent and has since raised more, with Nvidia participating). Before that it paid roughly $900 million to license the networking-chip startup Enfabrica's technology and hire its CEO and staff. The Groq, Enfabrica, and now Poolside deals are all variations on a template: take the technology and the team, leave the shell standing, avoid the regulatory review and integration cost of a full acquisition.

Why the software, not the models

The sharpest detail in the investor letter is the reason Poolside gave for agreeing. Continuing to compete in open-weight model development, the company said, required access to more Nvidia hardware than it could secure on its own. Poolside described a six-week window at the end of last year to raise $2 billion for a 40,000-GB300 cluster coming online in January — funding it did not close in time, as the site 00011000 reported from the letter, costing it the cluster.

Read that twice, because it reframes the whole arms race. The dividing line in frontier open-weight AI is no longer research ability; it is compute procurement. A lab can write a state-of-the-art training recipe and still not matter if it cannot get the accelerators. That is precisely the constraint Nvidia sits on top of — it is both the gatekeeper and, increasingly, the owner of the field. By licensing a proven training pipeline and absorbing the team that ran it, Nvidia gets the model-building know-how hardcoded into its own engineering, on hardware that will never be short of supply for itself.

What Nvidia hasn't bought is a finished frontier model. The license covers the means of production, not a product. That distinction matters for reading what comes next: Nvidia is not shipping Laguna under its own name; it is importing the machinery and the people to build the next generation of its own open family.

The strategic U-turn

This is the part of the story that deserves the long treatment, because it overturns the cleaner narrative. For two years the canonical open-vs-closed fight had clean flanks: Alibaba's Qwen and the Chinese labs pushing open weights, Anthropic and OpenAI defending closed APIs, and Nvidia selling compute to both sides. Nvidia profited from neutrality. Each open-weight drop was, in effect, advertising for Nvidia's chips — as we argued when Alibaba released its 2.4-trillion-parameter Qwen3.8 as open weights, the more powerful the free model, the more demand for the racks that run it.

The Poolside move breaks that neutrality. Nvidia is now a first-party participant in the open-model market, not just its landlord — and the WSJ report frames the intent explicitly as a "powerful U.S. alternative to Chinese AI." Nvidia already ships its own open Nemotron line, and The Information reported in mid-August that a Nemotron 4 model with more than 1 trillion parameters is under development, leapfrogging Nemotron 3 Ultra's 550 billion. Pair that trajectory with the Poolside pipeline and a "from 30B to 1T+" open lineup becomes a coherent ladder rather than a side project. Nvidia's own Hugo blog has been positioning the company as a force in open models and agentic AI all year.

The tension this creates is real and it sits unresolved at the director level. Nvidia's biggest customers — OpenAI, Anthropic, Microsoft, Google — are the same companies that pay Nvidia for training clusters. When the chip supplier starts fielding a competing open-weight model at scale, those customers are simultaneously Nvidia's revenue and Nvidia's rivals. The company has so far handled the conflict by keeping Nemotron aimed at the open tier rather than the closed frontier, but a 1-trillion-parameter open Nemotron built with Poolside's technology is a direct shot across the bow of the closed labs it now competes against on one axis and sells to on another.

What the open camp says — and the skeptics

The honest reaction from the open-model community is somewhere between excitement and a warning. SemiAnalysis, amplified by analyst Ramez Naam on X as the news broke, has been tracking that open-weight models are catching up to proprietary ones — and with each new wave of capability, faster. If Nvidia — the single richest actor in the accelerator economy — throws its hardware and distribution behind a frontier open model on American soil, the conventional wisdom that frontier open weights are a Chinese speciality collapses. The US would finally have a homegrown open answer to Qwen and DeepSeek of its own.

The skeptical reading is just as sharp, and it starts with the word everyone keeps using: answer. Answer to DeepSeek. Answer to Kimi. The framing concedes the initiative. For all of Nvidia's resources, its open model would be arriving in a field where the pace is set in Shenzhen and Beijing, and the weights it competes against are MIT-licensed and freely derivative-able. A non-exclusive license to another company's training software is an input, not a moat — Alibaba and the Chinese labs move faster on open releases, as the cadence of Qwen model drops has demonstrated repeatedly.

Then there's the deeper skepticism about motives. A company selling the compute and now the open models has to convince the market these are complementary, not conflicted. On the same day the Poolside deal surfaced, Bloomberg reported that Nvidia has told some of its largest customers that AI server prices will jump more than 15 percent — including on Vera Rubin and Grace Blackwell systems — starting in early 2027, as memory costs surge. Raising hardware prices while throwing $6 billion at your own model-making capability is, at minimum, a striking hand to play in the same week. It reads as a company seeking to own the two scarcest things in the market simultaneously: the silicon that runs every model and the open model that sets the benchmark price for the ones above it.

What to watch next

Four threads will decide whether this is a footnote or a pivot. First, what actually ships under the Nemotron brand in the next two quarters — whether the Poolside pipeline feeds a real frontier-scale open model, and whether that model is competitive with Qwen and Kimi or just another mid-tier open entry. Second, how the closed-lab customers react: a quiet recalibration of their dependence on Nvidia's roadmap for training clusters would be the tell that the conflict is biting. Third, whether the "license plus team, keep the shell" structure — now used three times — becomes Nvidia's default M&A replacement, and whether regulators start treating it as such. Fourth, and most important, whether Washington's emerging oversight framework for frontier open weights, which would now have to count Nvidia itself among the covered model makers, slows this down before it reaches the frontier.

The through-line is that Nvidia has stopped renting the game and started playing it. We covered the deal's structure the day it surfaced in Poolside strikes $6B Nvidia deal and a $1B investment, and it is worth setting this move beside the earlier read that Chinese AI labs stay locked into Nvidia via CUDA — the same flywheel that made Chinese labs dependent on Nvidia now pulls Nvidia into the open-model lane because that lane is where the dependency leads. If you want the building block primer, our explainer on what open-weight models actually are frames the stakes precisely: downloadable weights are the beginning of openness, not the end.

The deepest question is not whether Nvidia can build an open model — it plainly can. It's whether the company that owns the compute can be trusted to run an open ecosystem whose best interest sometimes means telling customers to spend less on that same compute. For nearly twenty years Nvidia's answer to every AI trend was the same: buy more GPUs. This $6 billion says the answer has finally changed.

If the company selling the shovels now digs its own mine, who wins — the open ecosystem, or Nvidia's shareholders? Tell us in the comments.

Sources: Wall Street Journal — Nvidia is spending $6 billion to build a U.S. alternative to Chinese AI · Newcomer — Poolside strikes $6 billion Nvidia deal · PYMNTS — Nvidia pays $6 billion to license Poolside software · The Next Web — Nvidia pays Poolside $6bn for its model factory · Bloomberg — Nvidia customers notified about AI-related price hikes