Nvidia will pay Poolside AI $6 billion for a non-exclusive license to the Model Factory system, on which the startup built its models, including the coding model Laguna, and will additionally invest $1 billion in the remaining part of the company at a $12 billion pre-money valuation. In a letter to investors, Poolside emphasizes that this is “not an acquisition and not an acquihire.”
What happened
Newcomer received Poolside’s letter to investors, and Bloomberg confirmed the deal. The $6 billion payment is specifically for a non-exclusive license to Model Factory, meaning the full training stack, not a single finished model. As part of the deal, Nvidia is making offers to 109 employees who worked on Laguna. Three Poolside co-founders will remain with the company, and the startup will distribute the $6 billion among its investors by the end of next year.
Context
Poolside is a startup that developed the coding model Laguna, positioned as a “Western answer to DeepSeek and Qwen,” and built all of its models on its own Model Factory system. According to the company’s letter, Poolside stopped in the race for frontier models: by January, the startup had not managed to raise $2 billion in 6 weeks for a cluster of 40,000 GB300s, which effectively cut off its access to frontier scale. In January, the company spun out Poolside Infrastructure, which is building a 1.2 GW data center in Texas, and the former lab is effectively repositioning as a data center player.
Why this matters for the industry
This is Nvidia’s third deal in recent months under the “license + hiring + equity” pattern, with a total volume of about $27 billion: previously there were Groq (about $20 billion, December) and Enfabrica (about $900 million). Nvidia is taking technologies and engineers without a full merger and without antitrust review, and members of Congress are already calling such deals a way to bypass merger review. The approach cements a new AI M&A format: companies are not buying a single finished model, but an entire model “factory” — data, training, RL, evaluation. The license remains non-exclusive, so Poolside will be able to license the system to others. For founders, this is a signal that the “own frontier models” business model is no longer working for labs that cannot quickly raise billions: the training pipeline is becoming a commodity, and value is shifting to applied layers and proprietary data.
Why this matters for users
There is no direct new product for readers and builders: there is no new API, no pricing, no latency data, and Model Factory remains an internal Nvidia system. The practical effect is visible on the open-models line: hiring the Laguna team strengthens the open Nemotron line, which is moving toward an open model with one trillion parameters, and this increases the chances of strong open coding models emerging in comparison with DeepSeek and Qwen on a horizon of about six months. Those building on open models should track Nemotron releases and their public benchmarks, and those considering a non-exclusive Model Factory license as a potential source of a training pipeline should note that its terms have not been disclosed.
What is still unknown / limitations
The sources contain no technical evidence: no benchmarks, no papers, no architectural details of Model Factory, so all claims about the capabilities of Laguna and the “factory” remain marketing. The statement about the open line moving toward a one-trillion-parameter model is a forward-looking roadmap without benchmarks or a description. The terms of the non-exclusive license have not been disclosed, and it is unknown whether Model Factory components, such as evaluation and RL, will be publicly available or licensed to third parties. Expectations about Nemotron releases and their comparison with DeepSeek and Qwen are a forecast on a horizon of about six months, not confirmed facts.
Sources
Author
Look at AI, editorial team