Nvidia is preparing to spend roughly $6 billion through a licensing and investment deal with AI coding startup Poolside, in a move aimed at building open-weight models that can compete with China’s DeepSeek and Moonshot as well as the closed frontier systems from OpenAI and Anthropic. The arrangement was reported by The Wall Street Journal over the weekend and detailed by Silicon Republic.
According to Silicon Republic’s account of the WSJ reporting, the combined commitment totals about $7 billion, including roughly $1 billion in equity at a $12 billion pre-money valuation, a fourfold jump from Poolside’s 2025 valuation of $3 billion. More than 100 of Poolside’s roughly 109 employees are expected to join Nvidia and work on the company’s Nemotron project, which develops open-weight models, though the deal is described as non-exclusive so the startup’s leadership stays in place.
The strategic logic is laid out in an open letter from Nvidia founder and CEO Jensen Huang on open weights and American AI leadership. “Our AI leadership will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem that diffuses into every sector,” Huang wrote, arguing that open weights expand access to the AI economy and avoid concentrating advanced capability behind a small number of closed models.
Why open weights, and why now
Open-weight models publish their trained parameters, so developers can download, customize, and run them without paying frontier-model prices per task. That positioning has made Chinese labs a global force: Hugging Face data cited in the reporting shows Alibaba’s Qwen models were downloaded more than 3 billion times in six months, ahead of Google, OpenAI, and Meta. Nvidia’s bet is that a credible U.S. open-weight challenger protects its own installed base, because those models overwhelmingly run on Nvidia hardware and software.
Poolside, founded in 2023, builds the Laguna series of open-weight models sized to run on a single Nvidia DGX GPU. Pulling that team and technology toward Nemotron shortens the path for Nvidia to field models that rival its own customers’ offerings, a tension the reporting flags directly: Nvidia would compete with OpenAI and Anthropic, two companies it also backs. The pattern is not new. Nvidia hired much of Groq’s talent in a 2025 licensing deal and spent more than $900 million to license technology and bring on staff from chip startup Enfabrica.
The push fits the same thesis behind the local AI hardware case: as inference shifts toward models teams can host and tune themselves, the vendor that supplies the silicon underneath wins regardless of which model wins the headline benchmark. Nvidia’s open-weights investment is a way to make sure that silicon is its own.
The next marker to watch is whether Nvidia formally confirms the Poolside arrangement and what Nemotron releases follow. If the reported plan holds, the company will enter 2027 with a direct stake in both the most expensive AI hardware and one of the most accessible open model stacks on the market.
