Nvidia Building Trillion-Parameter Open-Source AI Model to Counter Chinese Rivals

Nvidia is building a bigger open-source AI model, and the timing tells you exactly why.

The Information reported, citing two people familiar with the project, that Nvidia is training a new model family called Nemotron 4. The flagship version is expected to exceed 1 trillion parameters, according to the outlet's sourcing as relayed by Reuters on August 11. That would roughly double the size of Nvidia's current largest model, Nemotron 3 Ultra, a 550-billion-parameter system released in June 2026.

Nvidia has not set a release date and final training is not complete. Employees told The Information the model could be ready as early as late fall 2026. Nvidia previously shipped a smaller Nemotron-4 340B series, so this is a major step up in scale, not a first attempt.

The money backs up the ambition. Nvidia's cloud-compute budget for Nemotron development is capped at $7 billion through fiscal 2028, according to The Information's reporting as cited by AI Weekly. That is a serious, board-approved bet on open-source AI, not a side project.

Kari Briski, Nvidia's vice president of generative AI, put the rationale plainly in a company statement carried by Yahoo Finance. "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."

Why Nvidia needs this

Nvidia's biggest customers, including OpenAI and Microsoft, are quietly building their own AI chips, according to AI Weekly. If those companies succeed in weaning themselves off Nvidia silicon, Nvidia's core business takes a hit no matter how good its GPUs are.

Shipping the best open-weight model in the world changes that math. Any company, cloud provider, or government that wants to run a top-tier AI model without paying rent to a closed lab like OpenAI or Anthropic still needs somewhere to run it. Nvidia wants that somewhere to be Nvidia hardware.

There is also a China angle. Chinese open-source models, including Kimi K3, have closed the gap on top American systems from Anthropic and OpenAI this year, according to Korea Economic Daily. Most U.S. AI labs still do not release open-source models at all. Nvidia is positioning itself as the American open-source counterweight.

National security officials and technologists have raised concerns that if the best free AI models come only from Chinese labs, American companies, allied governments, and open-source developers will build critical infrastructure on Chinese-origin technology by default. Nvidia formed a coalition with Microsoft and other tech companies last month specifically to develop AI safety and cybersecurity tools, and the group issued a letter supporting open-weight models to keep AI development from moving overseas, according to Korea Economic Daily. That is a legitimate policy argument, even though it also happens to serve Nvidia's bottom line.

There's a genuine security wrinkle too. Open-source models carry no built-in restrictions on how they get used, including for cybersecurity purposes. Interest in open models has risen following recent disclosed hacking cases involving autonomous AI agents, according to Korea Economic Daily. An unrestricted trillion-parameter model is a tool that cuts both ways.

What actually shipped today

While Nemotron 4 remains unfinished, Nvidia did release something real on August 11: Nemotron 3.5 Lightning, a 30-billion-parameter model using a mixture-of-experts architecture with only 3 billion active parameters per task. Nvidia says it runs on a single GPU and delivers up to 4x faster output than comparable models, cutting agent task completion time by 30%.

According to Nvidia's own developer documentation, Lightning is built for the "execution layer" of AI agents, handling repetitive tasks like tool calls and result validation while larger models like Nemotron 3 Ultra handle complex planning. Nvidia also open-sourced NeMo Switchyard, a routing tool that automatically assigns tasks to whichever model handles them best.

Companies including CrowdStrike, Harvey, and CodeRabbit have started customized testing of the smaller model, according to BigGo Finance. Nvidia released the weights, training data, and recipes for Lightning under a permissive license called OpenMDW-1.1.

What's unverified

The trillion-parameter claim for Nemotron 4 rests on unnamed Nvidia employees speaking to a single outlet, The Information. There are no published benchmark numbers, no license terms, and no training-data disclosure for the flagship model yet, as AI Weekly's reporting noted. Nvidia has not confirmed a release date publicly.

If Nvidia hits its "late fall" target, the test comes fast: does Nemotron 4 actually outperform Chinese open models like Kimi K3, or does it just close the parameter-count gap without closing the capability gap. That answer will not be known until the model ships and independent researchers get to run it.