NVIDIA gives away AI models for free but makes more money from it than anyone else: breakdown of the Nemotron 3 Ultra strategy
On June 4, 2026, NVIDIA released Nemotron 3 Ultra, the largest open AI model in the Nemotron 3 lineup. Under a permissive license, the company released the model weights, training data, and methodologies. This is not just a "gift" to the community, but a calculated step within a long-term strategy where openness serves as a tool to strengthen its own ecosystem.
Architecture: A Hybrid of Speed and Accuracy
Nemotron 3 Ultra is not just another "scaled-up transformer." It is based on a hybrid architecture of three components: Mamba-2 layers, Attention layers, and Latent Mixture of Experts (Latent MoE). Mamba-2 processes long texts quickly with linear cost growth, while Attention retains large context volumes. Latent MoE compresses data before passing it to experts, allowing each to work narrowly and efficiently.
With a total of approximately 550 billion parameters, only about 55 billion are activated for processing each token. This gives the model 5–6 times higher throughput than comparable models and roughly 30% lower task costs. The context window is 1 million tokens, with a speed exceeding 300 tokens per second.
Strategy: A Free Model as a Catalyst for Hardware Sales
The main value of the release is not the model itself, but the ecosystem around it. Anyone running Nemotron almost certainly does so on NVIDIA GPUs, fine-tunes it using its tools, and deploys it on its software. Openness here is not charity, but a way to steer developers back to purchasing the company's hardware.
With a market capitalization exceeding $5 trillion, the cost of training Nemotron 3 Ultra (likely hundreds of millions of dollars) is a nearly negligible expense for NVIDIA. GPU sales more than cover the research, so the company can give the model away for free and still earn more than closed competitors charging for access.
The political context adds further weight to the release. An open American model can be inspected, modified, and run on one's own servers—making it attractive for countries building independent national AI, from Europe to Southeast Asia. Such a model cannot be remotely disabled, which is especially valuable amid recent restrictions around closed models.
Limitations and Prospects
Despite its strengths, Nemotron 3 Ultra is not the smartest model on the market. On 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 lag behind closed ones by three to seven months.
However, this gap matters less if an open model is simply sufficient for real-world tasks. A bank deploying Nemotron 3 Ultra for loan processing on its own servers doesn't 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.
Analytical Conclusion: NVIDIA's bet on efficiency rather than benchmark records may prove more far-sighted than it appears. With mass AI adoption, the cost of running a model takes center stage, and one that is nearly as capable but five times cheaper wins in real-world deployment. The open ecosystem will only strengthen: NVIDIA has the resources, motivation, and distribution channels to release increasingly powerful open models faster than any other company.