Nvidia’s next expansion is moving directly into the AI model layer. The company is developing the Nemotron 4 family of open models, with its largest version expected to contain at least 1 trillion parameters, according to The Information as cited by Reuters. Final training has not been completed, and a release could come as early as late fall.
The number is significant, but the larger story is Nvidia’s expanding position across the AI value chain. The company already dominates GPUs for AI computing, has built a powerful networking business around its infrastructure and established CUDA as a core software platform for developers. Nemotron 4 pushes Nvidia further into a market occupied by OpenAI, Google, Meta and a growing group of open-model developers.
Nvidia had already confirmed in March that Nemotron 4 was under development. It also created the Nemotron Coalition, bringing together Mistral AI, Black Forest Labs, Cursor, LangChain, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab to develop frontier open models. The coalition’s first base model, co-developed by Nvidia and Mistral AI and trained on Nvidia DGX Cloud, is intended to underpin the Nemotron 4 family.
Why Nvidia Needs Its Own Models
Nvidia does not need to defeat OpenAI or Google in consumer chatbots for Nemotron 4 to matter. Its strategic advantage lies elsewhere: models can drive more workloads onto Nvidia’s own computing platform.
The company has already built an open-model portfolio spanning agentic AI, robotics, autonomous vehicles, physical AI and biomedical research. Nvidia says its models are being adopted by companies including ServiceNow, CrowdStrike, Cursor and Perplexity, while Jensen Huang has described open models as critical to allowing companies and countries to build specialized AI systems.
Nemotron 4 therefore strengthens a loop that Nvidia has spent years building. More capable models create more AI applications; more applications require more inference; more inference consumes more computing capacity; and much of that capacity still runs on Nvidia GPUs, networking and software.
That is a different strategy from simply selling chips to whichever model developer wins.
It also comes as competition in open models intensifies. Reuters reported this week that U.S. companies including Nvidia and Meta are increasing their efforts as lower-cost, customizable open-weight models gain ground. Enterprises are increasingly looking beyond the largest frontier models toward systems that can deliver adequate performance at lower inference costs and can be customized for specific workloads.
From GPU Company to AI Infrastructure Company
Nemotron 4 becomes more important when viewed alongside Nvidia’s expansion beyond semiconductors. The company is increasingly connecting chips → networking → CUDA and AI software → models → data centers → capital.
That changes the competitive map. At the chip level, Nvidia faces AMD and custom accelerators developed by hyperscalers. At the model level, it increasingly overlaps with Google, Meta, OpenAI and other AI developers. In data centers and cloud infrastructure, its technology is becoming intertwined with hyperscalers and specialized AI cloud providers.
The strategy does not mean Nvidia needs to own every layer. Its stronger position may come from making the layers reinforce one another. A company using Nemotron models can train and run them on Nvidia hardware, connect them via Nvidia networking, optimize them with Nvidia software, and deploy them in infrastructure built on the same ecosystem.
This is also why Nvidia’s recent push into AI infrastructure financing matters. As AI data centers become increasingly capital-intensive, the ability to sell more computing hardware depends on whether customers can finance the facilities and secure the electricity needed to operate them. Models create computing demand; financing and power determine whether that demand can become physical capacity.
The strategic risk is equally clear. Nvidia is moving into markets occupied by some of its largest customers. Microsoft, Google, Meta and other hyperscalers buy large amounts of Nvidia hardware while simultaneously developing their own AI chips, models and software. The deeper Nvidia moves into models and infrastructure, the more the AI ecosystem becomes a mixture of partnership and competition.
For South Korea, the immediate opportunity remains hardware. Expanding AI model development and inference should support demand for high-bandwidth memory from SK hynix and Samsung Electronics. But Nvidia’s expansion also shows where a growing share of AI value may be created: not in a single semiconductor component, but in connecting chips, software, models, data centers and capital.
Nemotron 4 is therefore more than another large AI model. Nvidia is building a structure in which each new layer of AI can reinforce demand for the others. The company that became dominant by supplying the computing engine of the AI boom is now trying to extend that position into the models running on those machines — and the infrastructure and capital behind them.
