Nvidia trains 1-trillion-parameter Nemotron 4 open model

Nvidia is training a Nemotron 4 family whose flagship is expected to exceed one trillion parameters, roughly twice the size of its current-largest Nemotron 3 Ultra, according to The Information. Two people familiar with the project told the outlet the model could arrive as soon as this fall, though final training is not yet complete.

The money underlines the intent. Nvidia's cloud-compute budget for Nemotron development is capped at $7 billion through fiscal 2028, per the same report. Kari Briski, Nvidia's vice president of generative AI, said in a company statement carried by Yahoo Finance that "Nvidia is investing in Nemotron because we believe every company and every country needs accessible frontier open models to strengthen safety and security, accelerate innovation, and provide a foundation they can rely on from one generation to the next."

The context is a market in which cheap Chinese open models have been closing on top systems from leading American labs Anthropic and OpenAI, and in which Nvidia's biggest customers, OpenAI and Microsoft among them, are quietly building their own AI chips. Shipping the world's best open weights would give any lab, cloud, or sovereign buyer that wants to move off closed APIs a native reason to keep buying Nvidia silicon. It landed the same day Nvidia shipped Nemotron 3.5 Lightning and open-sourced NeMo Switchyard, a smaller drop that reads as a warm-up act.

The reporting rests on unnamed Nvidia employees at a single outlet, and includes neither benchmark numbers nor license terms nor a training-data disclosure for the flagship. If Nvidia does deliver a frontier-grade open-weight model on that late-fall window, the pressure lands squarely on the other major open-weight players and on the Chinese open-weight labs whose models have become the default when teams do not want to pay closed-lab rent.