Crypto news

On June 4, 2026, NVIDIA made an unexpected but strategically sound move: it released its largest AI model, Nemotron 3 Ultra, to the public under an open license. Not only the model weights, but also the training data and training methodologies were published. At first glance, this seems like a generous gesture. But behind it lies a harsh market logic that allows the giant to earn more than its closed competitors.

Hybrid Architecture: Mamba-2, Attention, and Latent MoE

Nemotron 3 Ultra is not just another "scaled-up transformer." It is based on a hybrid architecture combining three approaches: Mamba-2 layers, classic attention layers, and a Latent MoE (Mixture of Experts) mechanism. Mamba-2 layers process long texts quickly and efficiently, as their costs grow linearly rather than exponentially. Attention layers, in turn, accurately retain large amounts of context in memory. And Latent MoE compresses data before passing it to "specialists," allowing each expert to work narrowly and precisely without requiring unnecessary computations.

The result is impressive: with a total of approximately 550 billion parameters, only about 55 billion are activated for processing each token. This means the model thinks like a giant system but behaves like a much more compact one in terms of cost. Combined with a context window of 1 million tokens and a speed exceeding 300 tokens per second, this provides 5–6 times greater throughput and roughly 30% lower task costs compared to analogs.

Ecosystem Strategy: A Free Model as a Magnet for Sales

The main value of this release is not the model itself, but the ecosystem that NVIDIA is building around its hardware. The logic is simple: whoever runs Nemotron almost certainly does so on NVIDIA graphics cards, fine-tunes it using its software tools, and deploys it on its software. Openness here is not charity, but a way to steer developers back toward purchasing the company's hardware.

NVIDIA can afford this approach because its financial capabilities are incomparable to the cost of the model. With a market capitalization exceeding $5 trillion, training Nemotron 3 Ultra, which likely cost hundreds of millions of dollars, is almost 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 also adds weight to the release. An open American model can be inspected, modified, and run on one's own servers—this has made it attractive to countries building independent national AI, from Europe to Southeast Asia. Such a model cannot be remotely disabled, and this is especially valuable in light of recent restrictions surrounding closed models.

Weaknesses and the Future

Despite all its advantages, Nemotron 3 Ultra is not the smartest model on the market. In the independent Artificial Analysis Intelligence Index ranking, it scored 48 points—the best result among open US models, but globally it lags behind leaders such as Kimi K2.6 (54 points) and DeepSeek. According to analysts, open models trail closed ones by three to seven months.

However, this gap matters less and less if the open model is sufficient for real-world tasks. A bank deploying Nemotron 3 Ultra to process loans on its own servers doesn't need flagship-level intelligence—it needs a model that can be fine-tuned on private data, kept within a secure perimeter, and not share confidential information with outsiders.

My expert opinion: NVIDIA's bet on efficiency rather than test records may prove more far-sighted than it seems. With mass AI adoption, the cost of running a model comes to the forefront, and one that is almost as smart but five times cheaper wins in real-world operation. I expect NVIDIA's open ecosystem to only strengthen: the company has the resources, motivation, and distribution channels to release increasingly powerful open models faster than any other company.