Nvidia is building a monster open-source model to sell more chips

Nvidia is no longer content with just selling the picks and shovels for the AI gold rush.

The chip giant is developing Nemotron 4, an open-source AI model packing at least one trillion parameters. Led by VP Bryan Catanzaro, the project aims to double the size of its current flagship, Nemotron 3 Ultra, and directly rival the best open-source models in the world.

A $28B compute bet and a new coalition

Training a trillion-parameter model takes massive horsepower. To support the initiative, Nvidia has tripled its cloud-compute commitments to $28 billion through 2031.

Nvidia is also not working in a vacuum. It has formed the Nemotron Coalition, bringing in heavyweight partners including Mistral, Cognition, Reflection AI, and Thinking Machines to contribute training data and architecture ideas.

Why it matters

Nvidia wants to break its reliance on a handful of deep-pocketed frontier labs like OpenAI. If only a few players can afford to train world-class AI, Nvidia's GPU business remains vulnerable to their spending cycles. By releasing a top-tier open-source model, Nvidia gives every other enterprise a foundation to build on, broadening demand for heavy GPU clusters across the entire market.

We think assembling an alliance with hot startups like Mistral and Cognition is a brilliant way to crowdsource scarce data and secure developer loyalty. However, Nvidia walking the fine line between hardware supplier and frontier model creator might raise eyebrows among its biggest cloud customers who sell competing models.