Nvidia's $6B Poolside deal isn't an acquisition. Officially.

Nvidia paid $6 billion for a non-exclusive license to Poolside's Model Factory — the internal pipeline the startup used to train its coding models — then hired 109 of its engineers and put another $1 billion into what's left, at a $12B pre-money valuation. Poolside's investor letter is blunt about the framing: "not an acquisition and it is not an acquihire." The three co-founders, including CEO Eiso Kant, stay put and keep running an independent company that's still free to license the same software to anyone else.

That's the third time in nine months. Nvidia paid Groq roughly $20B for inference tech and its top engineers in December, and ran a similar play on Enfabrica for about $900M. Total committed via license-plus-hire: around $27 billion, none of it structured as a merger.

The prize is Nemotron, Nvidia's open-weight model line, reportedly aimed at a trillion-parameter open model to go up against DeepSeek, Kimi and the US frontier labs. Kant's own number is the tell — fewer than 70 people built Poolside's model. Nvidia just picked up that team and the machine that made them fast.

Why it matters: the company selling everyone else's compute now wants to own the open-weight models running on it — and it's assembling them one license at a time instead of one acquisition.