Nvidia Just Paid $6 Billion for a Startup It Didn’t Buy

Poolside spent three and a half years building AI coding models, then hit a wall that had nothing to do with talent. In a letter to shareholders reviewed by The Wall Street Journal, the company says it had a six-week window at the end of last year to raise $2 billion for a 40,000-GPU cluster coming online in January. It missed the window. It lost the cluster.

Last week, Nvidia paid Poolside $6 billion. Not to buy the company. To license the software Poolside used to build its models, hire more than 100 of its engineers, and take a $1 billion stake at a $12 billion pre-money valuation.

What Nvidia actually got

The licensed technology is Poolside's Model Factory, the internal system the startup used to develop its Laguna series of open-weight coding models. The license is non-exclusive, so Poolside remains free to license the same software to anyone else. More than 100 Poolside engineers received offers to join Nvidia and work on Nemotron, Nvidia's own family of open-weight models. Founders Eiso Kant and Jason Warner, the former GitHub CTO, are staying at Poolside.

The shareholder letter is blunt about the framing. It says the arrangement is "not an acquisition and it is not an acquihire." Poolside plans to distribute the $6 billion to its investors by the end of next year.

Nvidia has run this play before

The structure is a pattern now: buy a non-exclusive license, hire the engineers, take a minority stake, leave the company standing. Nvidia did it with Groq for $20 billion and with Enfabrica for $900 million. None of the three companies was bought outright, which means none of the three deals triggered the regulatory review a conventional acquisition would have.

Why a chip company wants to build models

Nvidia's stated goal, per the Journal, is to build one of the most capable open-weight models in the world, something that can go up against Chinese releases like DeepSeek and Kimi K3 while offering a cheaper, more customizable option than the closed models from OpenAI and Anthropic.

That puts Nvidia in direct competition with several of its own biggest customers, which is an uncomfortable place for a supplier to stand. The logic holds anyway: every capable model built and optimized around Nvidia silicon, CUDA, and networking creates more demand for Nvidia silicon, CUDA, and networking. Nemotron does not have to beat GPT for the strategy to work. It has to anchor an open ecosystem that runs best on Nvidia's stack.

It also gives Nvidia somewhere to point its own GPUs if outside demand ever softens, which is a hedge worth roughly $7 billion on its own.

What this says about everyone else

The Poolside letter is the most interesting document in the whole deal, because it describes a company that was, by its own account, technically on the right track and financially outrun. Capital requirements went vertical. The letter argues the binding constraint has become physical data center space and contracted compute as much as money.

That is the situation for the AI middle class right now. You can be genuinely good and still lose because you could not sign for 40,000 GPUs fast enough. Nvidia, standing on the supply side of that shortage, gets to collect the pieces on terms it writes itself.