Nvidia has agreed to pay Poolside $6 billion to license its AI technology and invest a further $1 billion in the startup at a $12 billion pre-money valuation, according to the Wall Street Journal. Roughly 109 Poolside employees will join Nvidia to work on Nemotron, its open-weight model program.
The target is explicit. Nvidia wants a model that competes with DeepSeek, Moonshot’s Kimi K3 and Alibaba’s Qwen, the Chinese open-weight families that have taken a large share of developers who download and run models themselves rather than calling an API. Jensen Huang framed the reasoning in national terms: “Our AI leadership will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem” that spreads across sectors.
Poolside was founded in 2023 by Jason Warner, formerly GitHub’s chief technology officer, and Eiso Kant. It builds the Laguna series of open-weight coding models, sized to run on a single Nvidia DGX system, and its $12 billion price tag is roughly four times the $3 billion valuation it carried at its Series B. Nvidia backed that round too, which makes this an escalation of an existing relationship rather than a new one. The licensing arrangement is non-exclusive and Poolside stays independent, with both founders remaining at the company. What Nvidia is buying is the technology, the engineers and the head start.
The move puts Nvidia in competition with its own customers. OpenAI, Anthropic and Meta all buy its hardware; a capable Nvidia-backed open-weight model competes for the same developers. It also aligns Nvidia’s model strategy with its hardware business in a specific way: open-weight models get downloaded and run on infrastructure the user provides, which is a different demand pattern from closed models served from a handful of large clusters.
Chinese labs have dominated that category by releasing aggressively. Moonshot published Kimi K3’s full weights in July as Washington debated restrictions on Chinese models. Z.AI broke from the pattern this month, holding GLM-5.3’s weights back for a two-week safety review after its cyber capabilities grew faster than the company expected, the first time a Chinese lab has delayed an open release on those grounds.
Nvidia’s own China business, meanwhile, keeps generating denials. The company said last week that a report it would ship a China-specific inference chip by year-end was incorrect, the second such denial in 2026. Building an American open-weight rival is a strategy that works regardless of how that question resolves, since a model that anyone can download does not need an export license.
Neither Nvidia nor Poolside has said when a Nemotron model built on the licensed technology might ship.
Sources: Silicon Republic, Tech Startups
–
By the Control Plane Editorial Team