Poolside AI: Laguna Models and Nvidia's $6B License Deal

Who started it, and why code

Poolside was founded in May 2023 by Jason Warner and Eiso Kant, who had met in 2017. Warner had been GitHub's chief technology officer through the launch of Copilot and had earlier led engineering at Heroku and Canonical; Kant had built several developer-focused startups, including Athenian. The company's premise is that source code, because it can be executed and checked for correctness, is a better substrate for teaching a model to reason than text alone, and that reasoning built on code generalizes to problems well outside programming. [1][3]

Its products are the Laguna family of open-weight models, trained through what the company calls the Model Factory, its internal pipeline for training, evaluating and iterating on models rather than a single one-off release. Poolside publishes research on that pipeline, including work on GPU efficiency, evaluation methodology and reinforcement learning from code execution at scale, and sells access to Global 2000 companies and public-sector customers alongside the open releases. [3]

Funding in two very different sizes

Poolside raised a $126 million Series A in 2023 and then a $500 million Series B in October 2024, led by Bain Capital Ventures with eBay Ventures, Nvidia and others participating, at roughly a $3 billion valuation. The Series B funded 10,000 Nvidia GPUs to train future models, the scale of compute the company needed to compete with much larger labs on model quality rather than just product features. [4]

The NVIDIA deal

In August 2026, NVIDIA agreed to license Poolside's Model Factory for $6 billion, the pipeline behind Laguna, and separately invested roughly $1 billion in a growth round that reporting valued Poolside at $12 billion to $13 billion depending on the outlet. NVIDIA also extended offers to more than 100 Poolside employees who had worked on model development, reportedly to work on NVIDIA's own open-weight Nemotron models. The license fee is reported to be paid out to Poolside's investors over time, through the end of 2027. [5][6]

The deal is structured as a license and a minority investment, not an acquisition or an acqui-hire: Poolside's three co-founders stay in place and the company continues to operate independently, still building and releasing its own models. Reporting on investor communications describes NVIDIA's reasoning as partly defensive, that continuing to compete in frontier model development would have required Poolside to secure more NVIDIA hardware than it could otherwise obtain, which the license and investment address directly. [6][7]

Atlas interpretation: For NVIDIA, the deal is a way to buy access to a working model-training pipeline and the people who run it without absorbing the company, useful groundwork for its own model lineup while a customer it also sells chips to gets a large capital infusion at a valuation multiple times its previous round. For Poolside, it trades some independence in who its top engineers ultimately work for in exchange for capital and a closer relationship with its most important hardware supplier, a trade a smaller lab racing much larger ones has limited room to refuse. [6]