Nvidia spent $27 billion buying AI startups — without acquiring a single one

Nvidia has agreed to pay Poolside, an AI startup best known for its coding models, $6 billion to use the methodology the company created for building and training them. The arrangement is a license — non-exclusive, meaning Poolside can offer the same rights to Google, Amazon, or any other buyer. Along with the license, Nvidia is investing $1 billion in what remains of the company, at a $12 billion pre-money valuation, and hiring 109 of its roughly 115 engineers and researchers.

What Poolside built is called the Model Factory: a proprietary system for constructing large language models through reinforcement learning from code execution. Rather than training on human preference ratings, the system runs millions of coding tasks against real execution results, letting the model correct itself by observing what actually works. Poolside applied this method with a team of fewer than 70 engineers to produce Laguna S 2.1, a 118 billion-parameter model that outperforms rivals ten times its size on standard benchmarks and runs in full on hardware available to individual developers — not just data centers.

The non-exclusive clause in the license is the detail that changes its meaning. Nvidia has not bought the recipe; it has bought the right to use it. Any competitor who approaches Poolside can negotiate the same deal, and Poolside’s three co-founders — who are staying on with the remaining company — have not disclosed what they plan to build next. The $6 billion will be distributed to investors before the end of next year, at which point the commercial logic of what remains is unclear.

For developers who currently rely on Poolside’s Laguna models for software engineering tasks, the departure of most of the technical team raises a practical question that the company has not answered: who maintains the roadmap going forward. The models remain available, but the engineers who built them will be working for Nvidia.

This is the third time Nvidia has acquired AI capability through a structure that stops short of formal ownership. It reached a $20 billion agreement to license inference technology from the chip startup Groq and absorb most of its engineering team. It also closed a comparable arrangement with Enfabrica, a hardware company, for roughly $900 million. Across the three deals, Nvidia has committed approximately $27 billion to AI capability without triggering merger review, carrying goodwill on its balance sheet, or requiring shareholders to vote. The deal with Poolside, like the previous two, is expected to close before the end of the year. The acquired engineers will work on Nemotron, Nvidia’s internally developed open-weight model family aimed at the top coding and reasoning benchmarks.

Tags: nvidia, ai, Machine Learning, Groq, Poolside, tech deals