Labs

Nvidia pays $7B
for Poolside's model
factory in its
most aggressive 'non-acquisition'
yet

A $6 billion license to Poolside's Model Factory, job offers to 109 engineers, and a separate $1 billion investment at a $12 billion valuation — the third time Jensen Huang has used the same license-plus-hire-plus-stake structure to absorb an AI startup without triggering merger review.

Nvidia will pay Poolside $6 billion for a non-exclusive license to the startup’s Model Factory, extend job offers to 109 engineers, and inject a separate $1 billion at a $12 billion pre-money valuation, according to a Poolside investor letter first surfaced by Newcomer on August 20 and confirmed the same day by Bloomberg and The Information’s Amir Efrati.

It’s the third time Jensen Huang has assembled a deal in exactly this shape. Roughly $20 billion went to Groq in December 2025 for a non-exclusive license to its inference designs, after which Groq closed a $350 million round at a $3.5 billion valuation. Earlier in 2025, Nvidia struck an approximately $900 million arrangement with Enfabrica along similar lines. The pattern is now legible enough that lawmakers have already begun describing it as an end-run around merger review.

The mechanics reward attention. Because the $6 billion is a license fee rather than a purchase price, the payment routes to existing backers, Bain Capital Ventures, eBay, Citi Ventures, Redpoint, and Adams Street, expected by the end of 2027, while the $1 billion equity check quadruples Poolside’s prior $3 billion mark. The 109 engineers moving to Nvidia represent close to the entire technical cohort. Co-founder Eiso Kant has said fewer than 70 people built the model itself, and fewer than 115 were involved across engineering and research. Whatever is left at Poolside after the transfer is a shell wrapped around a spun-out infrastructure company, separated in January 2026, that’s building a 1.2GW data center in Texas.

The asset being licensed is real. Poolside’s open-weight Laguna family, built under former GitHub CTO Jason Warner and Kant, posted 72.5% on SWE-bench Verified with Laguna M.1 and 70.9% with the smaller Laguna XS 2.1, competitive with Qwen-3.5 and DeepSeek V4-Flash. The Next Web reports Nvidia is preparing a trillion-parameter Nemotron release, and the Model Factory is presumably how it gets there faster.

What Huang has engineered is the acqui-hire re-priced for the antitrust era. A conventional acquisition of Poolside at these numbers would trigger months of review and, given Nvidia’s position in the AI supply chain, potentially block. The license-plus-hire-plus-stake structure delivers the same outcome, the people, the IP, and an equity claim on whatever’s left, while remaining, on paper, three separate commercial agreements between arm’s-length parties. The third instance is where a maneuver stops being clever and starts being policy. Regulators tend to notice on the third one.

Sources