NVIDIA is 'not acquiring' a startup in a $6 billion deal
NVIDIA has entered into a $6 billion deal with the startup Poolside, securing access to its AI models and a team of 109 engineers. This strategic move is aimed at strengthening the company's position in the field of 'open weights' models.
At the end of 2025, NVIDIA announced the largest deal in its history, in which it paid $20 billion to 'not acquire' the AI chip startup Groq (Groq, not Elon Musk's Grok). What does 'not acquire' mean? As part of the deal, NVIDIA paid for a comprehensive license for Groq's hardware and technology, and brought on board the founder of Groq and one of the developers of Google's TPU. Since this is not a standard acquisition, it avoided all the regulatory headaches that come with a full acquisition of a competitor. Now, NVIDIA is again announcing a similar deal in the billions with very interesting implications.
$6 billion for infrastructure and team, without the founders
As part of the new deal, NVIDIA will pay $6 billion to the American startup Poolside in exchange for the use of its AI models. The deal also includes the transfer of 109 of Poolside's employees to the chip giant. The parties also decided that NVIDIA will invest an additional $1 billion in Poolside at a pre-money valuation of $12 billion. Unlike the Groq deal, the founders of Poolside, including the former CTO of GitHub, will not join NVIDIA at this stage, unlike a significant portion of the company's engineers.
The startup, founded in 2023, specializes in developing large language models for automating programming tasks in natural language. The technology includes the Laguna model and the Model Factory training infrastructure – for which NVIDIA paid the huge sum.
But why, actually?
Why is NVIDIA willing to pay so much money for access to the systems and team of a young startup? A report in the Wall Street Journal reveals the motive: the chip manufacturer wants to use this deal to become the leading player in the field of 'open weights'.
An 'open weights' model is an LLM that allows users and developers to download, run, and fine-tune the model independently, unlike a 'closed' model. While training data usually remains confidential, the result of the training (the weights) reaches the general public. NVIDIA has been investing for years in the Nemotron open-source model series – a development based, among other things, on the technology of the Israeli company Deci AI, which was acquired in 2024.
According to the report, NVIDIA wants to compete with the dominance of Chinese companies like Moonshot (developer of Kimi models), DeepSeek, and Alibaba, which are already managing to offer impressive performance that is very close to the most advanced closed models of OpenAI and Anthropic.
Partner or competitor?
It is important to note that if NVIDIA reaches this goal, it will not only compete with the Chinese companies, but will position itself as a direct competitor to its customers and partners, including Anthropic and OpenAI. The WSJ notes that the new giant model that NVIDIA will work on will remain under the Nemotron branding. One of these models even served as the infrastructure for a giant language model trained on Hebrew by Dicta.
The move continues the strategy that NVIDIA presented last March at the GTC conference, when it unveiled the Nemotron coalition – a collaboration of companies like Mistral, Perplexity, and Mira Murati's Thinking Machines Lab for sharing data, expertise, and computing resources.
