Nvidia buys the production line: $6B for Poolside's factory, plus offers to 109 engineers

What transfers here is a process, not a set of weights. Poolside describes its Model Factory as an integrated collection of data pipelines, distributed training software, evaluation systems, inference infrastructure and reinforcement-learning tools, built to replace slow, manually coordinated training projects with something repeatable ^[7]. The supporting evidence is narrow but concrete: Laguna M.1 and Laguna XS.2 were both trained from scratch inside it on more than 30 trillion tokens, and Poolside says XS.2 went from the start of training to release in five weeks ^[8]. Cycle time is what Nvidia bought. A license to finished models would be worth a fraction of that, which is why the offers to 109 people who know why the number is five weeks and not fifteen sit inside the same transaction ^[2].

Per head, the license alone runs about $55 million ^[14]. Add the $1 billion equity cheque and the outlay is roughly $64 million for each person offered a job ^[15]. Against everything Poolside has disclosed raising, a $26 million seed, a $100 million Series A and a $500 million Series B ^[11], the licensing payment is about 9.6 times the company's lifetime capital ^[16]. Nvidia paid more than $900 million to license Enfabrica's technology and recruit its people, and ran the same structure on Groq in December 2025 while hiring founder Jonathan Ross ^[12]; the Poolside figure is at most about seven times the Enfabrica one ^[17]. The mechanism is house style now. The price of it has moved.

Nvidia was already inside this company, as an investor in the Series B ^[11], and already supports the Laguna architecture in its NeMo AutoModel documentation ^[9]. Supporting someone else's models on your silicon is a customer relationship. Holding the factory and the staff is something else: Nvidia gains the machinery and personnel to produce additional models under its own direction ^[9], aimed, according to the Wall Street Journal, at Chinese labs including DeepSeek and Moonshot AI as well as the closed-model providers OpenAI and Anthropic ^[3]. Open weights shipped by the firm that sells the accelerators do not need to top a leaderboard to earn their keep. They need to be good enough, free, and supported first-party in Nvidia's own stack ^[9].

The performance case for the factory, meanwhile, is still Poolside's own. Laguna S 2.1, a 118-billion-parameter mixture-of-experts model with 8 billion active parameters per token and a context window of up to 1 million tokens, is said to compete with much larger models on agentic coding benchmarks; those figures come from Poolside's evaluations, though it published its final evaluation trajectories for outside inspection ^[1].

Non-exclusivity is the part worth sitting with. Poolside keeps using the technology and both co-founders are expected to stay ^[6][5], so two organisations now operate the same production method with very different balance sheets behind them. Nvidia did not buy sole use. It bought a running start, and the seller of the compute is now also a supplier of the thing the compute is bought to make.