Nvidia is now building models to compete with its own customers

Nvidia, the chipmaker whose GPUs power most of the world's AI, has agreed to pay roughly $6 billion to license open-weight models from Poolside, a startup founded in 2023 and best known for its Laguna series of coding-focused systems small enough to run on a single Nvidia DGX server. The deal moves 109 of Poolside's engineers onto Nvidia's payroll and, on a combined basis including a $1 billion equity investment, is worth close to $7 billion. The models Nvidia is licensing are designed to compete with Chinese rivals DeepSeek and Moonshot. They are also designed to compete with OpenAI and Anthropic, two of Nvidia's largest existing investments.

"Open-weight" is the technical term for this class of model. In plain English: the underlying settings are published, so any developer can run, inspect, or customize them, generally at much lower cost than the closed systems from OpenAI or Google.

The shift the deal conceals is more durable.

Nvidia already takes equity in OpenAI, Anthropic, and a long list of other AI labs. The chips those labs use to train and run their models are the same chips Nvidia sells. Now the chipmaker is placing its own model line on the same shelf as the labs it underwrites, while still selling them the silicon underneath. The Poolside arrangement is structured as a non-exclusive license rather than an acquisition, so Poolside's founders Jason Warner and Eiso Kant stay at the company and Poolside can still sell to other buyers.

In a letter published on Nvidia's own site, Jensen Huang argued that US leadership in AI will be judged by whether an open ecosystem diffuses into every sector, not by which company owns the single strongest frontier model. The Poolside deal is the operational expression of that thesis: a chip company funding a publicly inspectable model line as a counterweight to both closed US labs and open Chinese ones.

Alibaba's Qwen family was downloaded more than 3 billion times in the past six months on Hugging Face, leading Google, OpenAI, Meta, and the rest of the field by that measure. DeepSeek and Moonshot have made similar inroads, per Forkast's reading of the deal. If open-weight becomes the default deployment tier for cost-sensitive customers, the closed frontier model stops being the product. The model that comes bundled with the chips takes its place.

That is the inversion Huang is now banking on. The next buyer of a DGX rack could be running a Poolside-derived model in production, with the closed-model license layered on top. The chip and the model are sold together, and the developer chooses the open-weight path because it is cheaper and inspectable. The frontier lab still gets paid; it just stops being the only place the model lives.

Newcomer first reported the deal on August 20 based on an investor letter. The Wall Street Journal and Bloomberg confirmed the terms through their own reporting, with TASS and Silicon Republic carrying the wire. Poolside's reported benchmarks include 72.5% on SWE-bench Verified and 70.9% on Terminal-Bench 2.1, sourced from Forkast and not yet cross-checked against Poolside's own technical reports.

The $1 billion equity investment values Poolside at roughly $12 billion pre-money, about four times the startup's $3 billion valuation earlier in 2025, according to the Silicon Republic coverage. Backers include Bain Capital Ventures, eBay, Citi Ventures, Redpoint, and Adams Street. The Laguna line is the foundation-model product Nvidia is now licensing.

The Poolside agreement is the third time Nvidia has used this licensing-plus-hire playbook in roughly eight months, after a $20 billion deal with Groq in December 2025 and a roughly $900 million arrangement with Enfabrica. That is the moment a one-off becomes a strategy, and the moment the inversion inside the AI stack becomes durable. Other hyperscalers will be reading the structure carefully; it is a way to absorb a model lab's technology and talent without the regulatory profile of an acquisition.

What the Poolside deal changes is not who owns the strongest single model. It is who gets to decide what ships on the next ten million DGX racks.