Nvidia is developing an open-source artificial-intelligence model with at least a trillion parameters, according to The Information, a scale that would put the chipmaker in direct competition with the largest open models in the field. The report, published Monday, landed on the same day Nvidia released a smaller open model, Nemotron 3.5 Lightning, in a pairing that shows the company pressing on two fronts at once.
The trillion-parameter project, reported to be called Nemotron 4, would make Nvidia the first hardware company to ship a frontier-scale open model. The threshold matters because it is the size at which models begin to display capabilities that smaller systems lack — reasoning, planning, and the kind of emergent behavior that has defined the top of the field. DeepSeek and other Chinese labs have already crossed it with open models, and Nvidia’s entry would give the open-source world a competitor with an advantage no lab can match: unlimited access to its own chips.
The strategic logic is visible from the outside. Nvidia sells the processors that train and run every serious model, open or closed, and its business does not obviously need a model of its own. But the company has been pulling the model layer into its orbit for years, arguing that its hardware is best used with software that knows how to exploit it. A flagship open model trained on Nvidia silicon, released under an open license, would give developers a reference point that runs best on Nvidia systems — and would give the company a hedge if the closed labs that dominate the field ever decide to build their own chips.
The release of Nemotron 3.5 Lightning on the same day underscores the strategy. Smaller models have become the workhorses of enterprise AI, running on servers and devices where cost and speed matter more than raw capability. By shipping a compact model alongside the trillion-parameter project, Nvidia is covering the full range of the market it wants to define: the frontier for ambition, the small models for deployment, all of it tuned to run on the company’s own hardware.
The move also reshapes the competitive map. OpenAI and Anthropic have built their businesses on closed, proprietary models, and both are preparing public listings whose valuations assume the moat holds. An open model at trillion-parameter scale from the company that supplies their chips would not erase that moat, but it would lower the floor under open alternatives, giving enterprises a credible option they do not have to license. For the open-source community, Nvidia’s entry brings an unusually well-funded patron to a movement that has often run on volunteer labor and research grants.
The technical questions are as large as the strategic ones. Training a trillion-parameter model requires tens of thousands of accelerators and months of time, and the costs run into the hundreds of millions of dollars. Nvidia can absorb those costs more easily than any lab, but it will face the same challenge every open-model developer faces: the difference between releasing weights and building a product. DeepSeek’s models are open and widely used, yet the companies trying to commercialize them have found that distribution, not parameters, is the scarce resource.
Analysts said the project is best read as a statement about the market Nvidia sees coming. The company’s revenue depends on AI computing continuing to expand across every layer of the economy, and a future in which a handful of closed labs control the model layer is one in which that expansion is gated by their budgets. An open model at the frontier keeps the market open — and open markets buy more chips. The move does not threaten Nvidia’s business; it insures it.
The enterprise question will decide how much the project matters. Open models have made steady inroads into corporate deployment, where cost control and data control argue for models companies can run on their own infrastructure. A trillion-parameter open model tuned to run efficiently on Nvidia hardware would give enterprises a reason to stay on Nvidia silicon, and it would give the company an answer to customers who say they need open weights to feel safe about their dependence on a single vendor. If Nemotron 4 ships and runs well, the open-source race stops being a sideshow and becomes part of the hardware company’s sales motion.
The Information did not say when the model would ship, and Nvidia has not confirmed the project publicly. What is already clear is the direction: after years of selling the shovels in the AI gold rush, Nvidia is now digging alongside the miners. The trillion-parameter question is no longer whether open models can reach the frontier. It is whether the company with the most chips in the world will get there first.
