Nvidia Pays $6 Billion for Poolside's Model Factory and Places Nemotron Against DeepSeek and Kimi
The license for the Model Factory and the arrival of over 100 engineers from Poolside positions Nvidia as the owner of an open-source model, not just the silicon that makes the model run.
Nvidia made its largest strategic repositioning since the deal with OpenAI in 2025 on Monday. It will pay $6 billion to license Poolside's Model Factory, the internal system that the French startup used to train its Malibu and Point models, and invest an additional $1 billion in capital, at a pre-money valuation of $12 billion. The license is non-exclusive. Poolside can continue selling the same technology to other buyers, which means Nvidia's hefty check pays for speed, not monopoly.
Approximately 109 engineers from Poolside are transitioning to Nvidia, including the team responsible for operating the training pipeline. Their mission is to accelerate the Nemotron, the family of open models that Nvidia began publicly promoting in August, which is currently competing for mindshare with DeepSeek V3, Kimi K3, and Qwen, in addition to Meta's Llama 4. The three founders of Poolside—CEO Eiso Kant, co-founder and former GitHub CTO Jason Warner, and COO Margarida Garcia—remain separate. They continue to lead what remains of the company, with a declared focus on unspecified research.
Why Nvidia Bought an Entire Kitchen
The obvious interpretation is vertical: Nvidia stops just selling the training hardware and now has the software that decides how the training occurs. This is not a short-term play. It is a response to a problem that became public in 2025: the greatest model improvement curves began coming from data engineering and the training pipeline, not from more GPUs. If the customer paying for the cluster can also run a rival Model Factory, commoditization begins at the top.
The second reason is geopolitical. DeepSeek and Kimi K3 became too proficient in code for Washington's liking, and the current Nemotron is lagging in software engineering benchmarks. According to Reuters, Nvidia plans to leverage Poolside's technology and expertise to improve Nemotron while competing against Chinese models and American offerings from OpenAI and Anthropic. For a consulting firm like Capgemini, which has €700 million in restructuring planned by 2027 and states that generative AI and agentic have jumped from 5% to 10% of their bookings in a single quarter, a competitive American open model in code is the difference between passing license costs onto clients and absorbing them.
Where the $7 Billion Check Lands
In the United States, the acquisition opens up the opportunity for corporate clients to run a significantly better Nemotron in code within their own data centers. This directly targets the proposition from Anthropic and OpenAI for the software engineering segment. Anthropic had been converting investment banks and Big Four firms into multiyear contracts, with annualized revenue projected to exceed $30 billion by 2026, according to market data, but now sees the prospect of a cost-free substitute for token use at clients like Goldman or Deutsche Bank.
In Europe, the effect is distinct. Poolside is headquartered in Paris, and France has been showcasing the company as a flagship case for the EU's technological sovereignty, alongside Mistral. A $6 billion license, engineers migrating to California, and a headquarters reduced to three founders dedicated to research place the narrative in question in Brussels and Paris. The timeline is tight: the full enforcement phase of the AI Act for GPAI began on August 2, and the Chips Act 2 enters review in October.
In India, the interpretation is at the level of suppliers. TCS announced in July its intention to hire 8,900 AI engineers while cutting over 23,000 positions in a declared shift to an AI-first model. If the Nemotron, enhanced by the Model Factory, improves its coding capabilities, Indian firms reduce their reliance on paid APIs from Anthropic and OpenAI, thereby enhancing margins on outsourcing contracts that currently absorb part of the cost in client pass-throughs.
What Remains Open
Nvidia spends $7 billion, absorbs 109 engineers, licenses a pipeline, and gains an argument to sell clusters alongside models. In return, it incorporates the same thesis that part of the open-source community has been contesting since the beginning of the year: that investing billions in assets that any competitor can download for free is good capital usage. The counterargument is that Nvidia does not need to monetize the model; it just needs it to exist and to perform well on Nvidia GPUs. In this scenario, the value is not in what Nemotron sells, but in what it prevents others from selling. It's a defensive thesis, and defensive plays usually do not pay well for those demanding a growth premium. The first test comes on Wednesday with the release of Nvidia's second-quarter results.