Nvidia has agreed with startup Poolside on a deal worth approximately $7B: $6B will go for a non-exclusive license to the Model Factory model development system, and another $1B is an investment at a pre-money valuation of $12B for Poolside. More than 100 Poolside engineers will join the Nemotron team — Nvidia's project for open open-weight models. According to The Wall Street Journal, the goal of the deal is to make Nemotron competitive with Chinese DeepSeek and Kimi (Moonshot AI) and closed models from OpenAI and Anthropic.
What happened
The technical asset of the deal was not the model weights, but Model Factory — the system, i.e., the pipeline and infrastructure for model development built by Poolside. Nvidia is paying $6B for it and receiving a non-exclusive license: formally, Poolside retains the right to sell access to the same system to other clients. The deal also includes $1B in investment from Nvidia into Poolside at a pre-money valuation of $12B for the company. According to Forbes, 109 Poolside engineers are receiving offers and joining the Nemotron team. The company's founders, Aiso Kant and Jason Warner, remain at independent Poolside, and in a letter to shareholders, Nvidia emphasizes that the deal is 'not an acquisition and not an acquihire.'
Context
Nemotron is Nvidia's project for open open-weight models, and this deal moves the company from the position of a compute provider to a direct participant in the race for the quality of open models. The 'license plus hiring' structure is not new for Nvidia: the deal repeats the company's recent scheme, applied after Groq (valuation of $20B) and Enfabrica (approximately $900M), when Nvidia bought ready-made expertise without a classic acquisition. Before the deal, Poolside remained an independent startup whose team trained models on a cluster of 40,000 GB300 GPUs; in the fall, the company failed to close a $2B round within six weeks to expand this cluster.
Why this matters for the industry
For the industry, this is the first case where Nvidia is paying billions not for compute, but for the technology of building models itself. The 'license plus hiring' structure, which formally is not an acquisition, is estimated to allow bypassing antitrust review and could become a template for AI market consolidation. The $6B license price and the $12B valuation of Poolside become a new anchor point for negotiations in similar deals. The strongest signal is the talent flow: engineers who built training on a cluster of 40,000 GB300 GPUs are moving to Nemotron, and this is exactly the layer of expertise — data pipelines, scaling recipes, training infrastructure — on which DeepSeek and Kimi achieved success. At the same time, Nvidia is directly competing with the same labs to which it sells GPUs.
Why this matters for users
There is no direct impact today: there is no new model, API, weights, or benchmarks yet, nothing to deploy. If Nvidia's plan is realized, the Nemotron family will be supplemented with open weights that can be downloaded, fine-tuned, and deployed locally without OpenAI or Anthropic subscriptions. The point of observation is Nemotron releases on Hugging Face and GitHub: a team that has been joined by more than 100 Poolside engineers is now working on the models. Product teams should prepare in advance scenarios for comparing API and local open-weight models by latency, cost, and quality on their use cases.
What is still unknown / limitations
The announcement contains no technical evidence: no architecture, no model sizes, no weights, no benchmarks, no technical report. The statement about Nemotron's competitiveness with DeepSeek, Kimi, and closed models from OpenAI and Anthropic — according to WSJ — is a declaration of ambitions, not a measurable result. Forecasts of a 'strong open Nemotron family in the coming months' remain untestable until weights and independent benchmarks appear. There is no release schedule or public documentation for Model Factory in the sources, nor independent confirmation that the deal structure actually bypasses antitrust review.
Sources
Author
Look at AI, editorial team