Ahmed Shaami
Nvidia has agreed to pay AI startup Poolside $6 billion for a non-exclusive licence to its model-building system, and to invest a further $1 billion in the company at a $12 billion pre-money valuation. Roughly 109 Poolside staff received offers to join Nvidia and work on Nemotron, the chipmaker’s open-weight model family. The terms come from a letter to Poolside investors first reported by Eric Newcomer on August 20, 2026.
The shape of the agreement
The licensed asset is Poolside’s “Model Factory”, the internal system the startup used to build and train its own models. Because the licence is non-exclusive, Poolside remains free to license the same software to other buyers.
Poolside’s three founders, chief executive Eiso Kant, co-founder Jason Warner and operations executive Margarida Garcia, are not moving to Nvidia and will continue independent research at the company. The investor letter was explicit about the structure, stating it is “not an acquisition and it is not an acquihire.”
The Wall Street Journal reported the arrangement as a sweeping technology and investment agreement in which more than 100 engineers transfer into Nvidia’s Nemotron effort. Reporting indicates the deal came together after a fundraising round fell short, leaving Poolside without the computing capacity needed to keep training at the frontier.
Why Nvidia is buying its way into model building
Nvidia has spent the AI boom selling the hardware other companies use to train models. This deal moves it upstream into building the models themselves. Nemotron is already shipping, and beingguru covered the release of Nemotron 3.5 Lightning as Nvidia’s first open-source AI model earlier this month.
The competitive pressure is coming from China. Developers including DeepSeek and Alibaba’s Qwen team have released capable models with downloadable weights, giving builders a degree of control that closed systems do not offer. A stronger American open-weight family gives Nvidia a counterweight, and, conveniently, every organisation that downloads and fine-tunes those weights needs GPUs to run them.
The deal structure is itself a signal. Rather than acquiring a startup outright and inviting regulatory scrutiny, large technology companies are increasingly assembling licensing agreements, minority investments and mass hiring into a single package that delivers the intellectual property and the team without a takeover.
What downloadable weights change for smaller developers
For a developer in Lahore or a two-person agency in Dubai, the practical difference between a closed API and a model with published weights is the shape of the bill. Closed access is metered per token, so costs scale with usage and stay unpredictable. A downloadable model can be run on rented GPU time or on your own hardware, which turns a variable cost into a compute budget you control.
Downloadable weights also matter for client work involving sensitive data. If a client will not permit their documents to leave a controlled environment, a model you can host yourself is often the only viable option. Fine-tuning on a narrow domain, which is where most small-agency AI work actually lives, is far easier when you hold the weights.
One caution before anyone builds a business on this. “Open-weight” and “open-source” are not interchangeable terms. Weights being downloadable says nothing about what the licence permits commercially, and several prominent model licences carry restrictions on scale or use case. Read the licence before you ship a client product on top of it.
The founders who stayed behind
Kant, Warner and Garcia keeping Poolside independent while its engineering core moves to Nvidia is an unusual outcome for a company valued at $12 billion pre-money. It leaves an open question about what Poolside becomes next, with its model factory licensed non-exclusively to the world’s largest chipmaker and the engineers who built it now working on someone else’s model family. Additional detail on the deal is available from The Next Web.
