Nvidia's €5.6 Billion Move to Build Open-Weight AI

Picture this: a hush over Silicon Valley, then a single, decisive move that redraws the map. Nvidia has quietly agreed to a landmark deal worth €5.6 billion with startup Polyside, a partnership designed to push open-weight artificial intelligence into the mainstream.

It is not just a license. Nvidia will also back Polyside with roughly €930 million as part of a valuation tied to the startup. More than a hundred of Polyside's engineers and core staff will fold into a new effort centered on Nvidia's Nemotron project, bringing talent, code and momentum to a plan that aims at the clearest prize in AI right now: powerful, open, and usable models that anyone can run and adapt.

Inside the play

The ambition is blunt and bold. Nvidia wants a model family that can match China’s most impressive open systems — think DeepSeek and Kimi K3 — while offering a credible, efficient alternative to closed models from dominant players. The company is already shipping a lightweight Nemotron 3.5 Lightning. Next up: a training run that will exceed one trillion parameters.

Why does this matter? Because open-weight models change the economics and the politics of AI. They cost far less to deploy, let organizations tune behavior to real-world needs, and foster experimentation outside large labs. That lower barrier to entry has helped Chinese open models accelerate faster than many expected, and Nvidia’s move signals that the US tech stack is gearing up to compete on those same terms.

Polyside’s founders framed the deal plainly: the goal is to prevent general-purpose intelligence from becoming the exclusive property of a very small number of companies. They want a future where comprehensive AI is cultivated in common, not locked behind closed gates. That statement reads like a manifesto, but it is also a strategic orientation for both parties.

Executives at Polyside will not be absorbed into Nvidia’s executive roster. They will pursue other research projects, while their engineering teams join the Nemotron effort. That split preserves independent research pathways and also keeps a flow of ideas between an upstart culture and one of the world’s biggest chip makers.

There is a commercial logic too. Nvidia sells the hardware that powers modern AI. By investing in open, high-performance models, the company protects and expands demand for its processors while shaping the software architecture that will run on them. It is a two-way street: chips and code, together.

Call it a chess move. Or call it ecology building. Jensen Huang, Nvidia’s CEO, has repeatedly argued that winning global AI competition is not about a single model but about creating a healthy, open ecosystem. This transaction reads like a step in that direction, one that pairs deep-pocketed infrastructure with agile model teams.

There are risks. Open-weight models are attractive precisely because they are accessible, which raises questions about safety, misuse and governance. But the counterargument is practical: democratized models can also spur defensive, innovative uses faster than closed systems ever could.

For global readers tracking AI power shifts, the deal is a clear signal. It shows how infrastructure giants are no longer content to be passive suppliers; they want to help design the stack that sits on their silicon. Expect more deals like this, more talent migrations, and faster iterations on trillion-parameter, open-weight systems.

Nvidia has always been about compute. With Polyside, it is stepping into the software arena in a more deliberate way, nudging the future toward models that are powerful, cheaper to run, and more widely owned.