Nvidia has agreed to pay Poolside $6 billion for the rights to “Model Factory,” the internal tooling the startup used to build its Laguna models, and it is extending job offers to 109 of the company’s engineers. The chipmaker is separately putting in $1 billion of fresh capital at a $12 billion pre-money valuation, and Poolside’s three founders are staying put rather than joining Nvidia.
The terms come from an investor letter that Eric Newcomer’s newsletter first surfaced, according to The Decoder, whose fuller writeup on the letter is the basis for this account. Poolside has set itself a deadline: the entire $6 billion license fee is to reach its investors before next year closes, effectively converting the software payment into a payout for the venture backers who funded Laguna’s development.
The letter is explicit that Nvidia is not buying Poolside. It states the arrangement is “not an acquisition and it is not an acquihire,” a term for hiring sprees built around headcount rather than a company’s products or revenue. That denial sits awkwardly next to the mechanics: Nvidia gets the software, gets the staff who built it, and leaves behind a shell that pays its investors and lets the founders keep their titles.
The structure matters because a straight acquisition of Poolside would trigger antitrust review, the kind of scrutiny that can stall a deal for months while regulators examine market concentration. Licensing the technology and extending individual job offers sidesteps that process entirely, since no single transaction consolidates ownership of the company itself. Big buyers have leaned on this playbook before precisely because it delivers the technology and the talent without the paperwork of a merger filing.
Nvidia has run this play twice already this year. It struck a $20 billion licensing arrangement with Groq and a $900 million deal with Enfabrica, both similarly built around technology access and staff offers rather than outright purchases. The Poolside terms sit between those two in size, and they follow the same template: pay for the system, hire the people who know how to run it, leave the corporate entity intact on paper.
There is a second tension the letter does not address. Nvidia already ships its own open-weight models under the Nemotron line, and it sells the GPUs that companies like Poolside rent to train their models in the first place. Licensing a system built specifically for constructing AI models puts Nvidia in more direct competition with the same model-building customers who buy its chips, a conflict that a hardware vendor moving into model tooling will have to manage carefully as more labs weigh whether their infrastructure supplier is also becoming their rival.
For AI Insiders readers who have tracked Poolside’s Laguna releases and its research staff growth over the past year, this is the exit: the model line’s engineering core moves to Nvidia, the company that supplied the compute it trained on. Startups building proprietary model-training infrastructure should treat this as a signal that the buyers most likely to want that infrastructure are the chipmakers already selling them the hardware underneath it, and that the deal terms on offer increasingly avoid the word “acquisition” even when the outcome functions like one.
Reported by Maximilian Schreiner for The Decoder on August 21, 2026.
