Nvidia is training an AI model with at least a trillion parameters that it does not intend to charge anyone for. The Information reported the plan on 11 August, citing several employees working on the project. The model family is called Nemotron 4, final training is not finished, no release date has been set, and people on the team think it could be ready as early as late autumn.
The strange economics of giving away a frontier model
Frontier models are the most expensive artefacts the software industry has ever produced. Companies build them because the model itself is the product. Nvidia is building one because the model is an advertisement for the product, and the product is silicon.
Every open-weight model that developers actually adopt creates downstream demand for GPUs, networking gear, inference capacity and Nvidia software. The company does not need Nemotron to out-earn Claude or GPT. It needs Nemotron to be good enough that thousands of teams build applications on it, tune it, serve it, and buy compute to do all three. If the model is optimised hard for Nvidia hardware along the way, so much the better.
Kari Briski, Nvidia vice president of generative AI, framed it in public-good terms: 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. Both things can be true. Open weights genuinely do widen access, and they also happen to widen the market for the only company selling the chips everyone needs.
Why now
Timing is the tell. Nvidia is one of very few large American firms shipping open-weight models at all, and the pressure to do so has come mostly from China. DeepSeek and Moonshot AI have shown that freely distributed models spread through developer communities faster than licensed ones, and that the capability gap with closed American systems has narrowed to something a cost-conscious enterprise will tolerate. Meta has treated open releases as a way to set standards. Nvidia now has a sharper version of the same incentive, because it profits regardless of who wins the model layer, as long as the winner runs on Blackwell.
There is a defensive angle too. If AI budgets keep inflating and cheap open models absorb the workloads that closed labs currently monetise, the value in the stack shifts toward whoever supplies the compute. Nvidia would rather accelerate that shift than wait for it.
The buildout underneath
The rest of this week supplied the physical evidence. IBM and Together AI signed a 240 million dollar multiyear agreement to build a US-based inference cluster on IBM Cloud, starting with roughly 2,000 Nvidia Blackwell-generation chips in HGX B300 systems connected by Spectrum-X networking. Together AI built its business helping enterprises run open models, which is the exact demand Nemotron 4 would amplify. The company expects much of the capacity to be committed before it is even deployed.
Foxconn reported the manufacturing side of the same story. Second-quarter net profit rose 35 percent year over year to NT$59.97 billion, about 1.86 billion dollars. More striking was the mix. Cloud and networking products, the division that contains AI servers, hit 51 percent of quarterly revenue for the first time. Consumer electronics, including the iPhone business that built the company, came in at 29 percent. Foxconn is preparing production of Nvidia Vera Rubin systems this quarter with shipments targeted for the fourth.
The constraint is moving upstream. Foxconn warned that next year’s server volumes depend partly on available CoWoS advanced packaging capacity, which comes largely from TSMC. Demand is not the bottleneck. Packaging is.
What a trillion parameters buys
Parameter count is a crude proxy for capability and a poor one for usefulness, and a trillion-parameter open model would still be smaller than several Chinese releases. What matters more is whether Nvidia ships weights that teams can realistically run, tune and serve without a hyperscaler contract. A model too large to deploy is a press release.
If Nemotron 4 lands in late autumn and performs anywhere near the top open models, the interesting consequence is not a new leaderboard entry. It is that the company selling the hardware has decided the fastest way to sell more of it is to remove the reason anyone would pay for a model at all. For more coverage of frontier hardware and future tech, visit Mylistingo.
