The "hook and line" strategy: how NVIDIA gives away AI for free and makes the most money
On June 4, 2026, NVIDIA released Nemotron 3 Ultra, the largest open-source AI model in the Nemotron lineup. The release includes not only the model weights under a free license but also training data and training methodologies. This is not just "another open-source LLM" — it is a subtle and well-calculated move that strengthens the company's position in the hardware market.
Architecture: A Hybrid That Changes the Game
Nemotron 3 Ultra is not a scaled-up transformer. It is based on a hybrid architecture combining three approaches: Mamba-2 layers, classic attention layers, and a latent mixture of experts (Latent MoE).
Mamba-2 layers process long texts quickly and efficiently: their costs grow linearly with length, rather than exploding like the standard attention mechanism. Attention layers, in turn, accurately retain large amounts of context in memory. Latent MoE compresses data before passing it to experts — each of which works narrowly and precisely, without unnecessary computations.
The result: with a total volume of around 550 billion parameters, only about 55 billion are activated for processing each token. This provides a context window of 1 million tokens, a speed of over 300 tokens per second, and, by my estimates, 5-6 times greater throughput at a cost roughly 30% lower than comparable models.
Strategy: A Free Model as a Driver for Hardware Sales
The main value of this release is not the model itself, but the ecosystem NVIDIA is building around its hardware. Anyone running Nemotron is almost certainly doing so on NVIDIA graphics cards, fine-tuning it with NVIDIA tools, and deploying it on NVIDIA software. Openness here is not charity — it is a way to steer developers back toward purchasing hardware.
With a market capitalization exceeding $5 trillion, training Nemotron 3 Ultra, which likely cost hundreds of millions of dollars, is nearly a negligible expense for the company. Graphics card sales more than cover the research, so NVIDIA can give away the model for free and still earn more than closed competitors charge for paid access.
The political context adds weight: an open American model can be inspected, modified, and run on one's own servers. This makes it attractive for countries building independent national AI — from Europe to Southeast Asia. Such a model cannot be remotely disabled, which is especially valuable given recent restrictions around closed models.
Limitations and Prospects
Nemotron 3 Ultra is not the smartest model on the market. In the Artificial Analysis Intelligence Index, it scored 48 points — the best result among open US models, but trailing leaders like Kimi K2.6 (54 points) and DeepSeek. Open models generally lag behind closed ones by three to seven months.
However, this gap matters less and less if an open model is sufficient for real-world tasks. A bank deploying Nemotron 3 Ultra for processing loans on its own servers does not need flagship-level intelligence — it needs a model that can be fine-tuned on proprietary data, kept within a secure perimeter, and not expose confidential information to outsiders.
My view: NVIDIA's bet on efficiency rather than benchmark records may prove more far-sighted than the race for the "smartest AI." In mass adoption, the cost of running a model takes center stage. A model that is almost as smart but five times cheaper wins in real-world operation. Given NVIDIA's resources, motivation, and distribution channels, I expect its open ecosystem to only grow stronger. The company has everything it needs to release increasingly powerful open models faster than anyone else.