NVIDIA is making two major moves that could reshape how AI gets built and deployed: releasing a powerful open-source reasoning model while its Blackwell GPU chips continue to dominate enterprise infrastructure. On June 4, 2026, the company unveiled Nemotron 3 Ultra, a 550-billion-parameter model specifically engineered for AI agents that need to plan, reason, and work across hundreds of steps. Meanwhile, Blackwell chips are driving a global data center construction boom, with companies from Amazon Web Services to Google Cloud racing to deploy them at scale.

What Makes Nemotron 3 Ultra Different From Other AI Models?

Nemotron 3 Ultra isn't designed to win chatbot competitions. Instead, NVIDIA built it to orchestrate complex workflows where AI agents need to call tools, delegate to sub-agents, recover from errors, and keep working across hundreds of turns without wasting computational resources. According to NVIDIA's benchmarks, the model achieves up to 5x higher throughput and up to 30% lower cost to complete agentic tasks compared to other open models in its class.

The model carries 550 billion parameters but only uses roughly 55 billion active parameters per token, meaning it has the knowledge capacity of a very large model while paying the inference cost of a much smaller one. This efficiency matters because it lets the model run fast and cheap relative to denser alternatives. NVIDIA released the model weights, training data pipeline, and recipes under the permissive OpenMDW-1.1 license from the Linux Foundation, making it fully open in ways that most closed models prohibit.

On the Artificial Analysis Intelligence Index, Nemotron 3 Ultra scored 48, positioning it as the most capable open model from a US lab. The model lands in what analysts call the most attractive quadrant on their chart, combining high intelligence with fast output speed.

How to Deploy Nemotron 3 Ultra Across Your Infrastructure

• Single-Node Setup: Requires 8 B200 GPUs with roughly 1.5 terabytes of aggregate memory for organizations running smaller workloads or testing the model before scaling.
• Multi-Node Deployment: Runs on 8 or more H100, H200, GB200, or GB300 GPUs for enterprises needing to process larger volumes of agent tasks across distributed systems.
• Software Framework Support: Ships with deployment recipes for vLLM, SGLang, and TensorRT-LLM, with TensorRT-LLM currently limited to Blackwell hardware for maximum performance.
• Quantization Options: Available in a base model, post-trained instruct variant, and NVFP4 quantized version that shrinks memory requirements while maintaining accuracy across Hopper, Blackwell, and Ampere GPUs.

The model supports configurable reasoning modes, including a full thinking mode, medium effort mode, and reasoning-off mode toggled through the chat template. For teams watching token spending closely, the model also supports a hard reasoning budget where you set a token ceiling on the thinking trace, forcing the model to stop deliberating and answer once it hits the cap.

Why Blackwell Chips Are Driving a Global Data Center Construction Wave

While Nemotron 3 Ultra represents NVIDIA's software strategy, Blackwell GPUs are reshaping the physical infrastructure of AI. The flagship GB200 NVL72 system packs 72 B200 GPUs into a single rack, delivering up to 30x faster inference than the previous-generation H100 for large language model tasks and up to 4x faster training on AI models.

Each B200 GPU contains 192 gigabytes of HBM3e memory and delivers 9 petaflops of FP8 performance. These aren't consumer graphics cards; they're purpose-built computing engines designed from the ground up for the AI era. The demand is extraordinary. Amazon Web Services, Google Cloud, Microsoft Azure, Meta, and Oracle have all confirmed they're deploying Blackwell at scale, with NVIDIA CEO Jensen Huang describing demand as "insane" based on order backlogs.

The infrastructure requirements for Blackwell are substantial. A single GB200 NVL72 rack can draw 120 kilowatts of power at peak and generates enormous heat requiring liquid cooling, higher power density, and stronger floors. Traditional data centers weren't designed for 100-plus-kilowatt racks, triggering a multi-billion-dollar global construction wave. Microsoft and other hyperscalers are investing massively in new data center infrastructure to handle Blackwell clusters, with countries like India seeing significant investment in digital infrastructure to support the AI boom.

What Do These Releases Mean for the AI Industry?

Nemotron 3 Ultra and Blackwell represent NVIDIA's dual strategy: dominating the hardware layer while building credibility in open-source software. The Nemotron family, which includes smaller Nano and Super models alongside Ultra, is designed to work together as a tiered system. The larger Ultra model teaches and coordinates the smaller ones, reducing overall token consumption across workflows. Because Nemotron is open with a permissive license, its outputs can legally be used to post-train the smaller models, something the license terms of most closed frontier models prohibit.

NVIDIA's Nemotron 3 Ultra is built on several architectural innovations that push efficiency and accuracy in its favor. The model interleaves Mamba-2 layers with attention layers, allowing it to process long sequences far more efficiently than traditional attention-only models. This hybrid approach makes the 1-million-token context window practical, meaning the model can process roughly 100,000 words at once. The model also uses LatentMoE, which projects tokens into a smaller latent dimension before routing them to experts, making expert routing more efficient without the overhead a conventional mixture-of-experts design would incur.

Blackwell's efficiency gains don't automatically mean lower total energy use. When you deploy more powerful chips and run them harder, aggregate consumption can still grow. However, Blackwell is more efficient than its predecessor on a per-computation basis. A single GB200 NVL72 rack can draw 120 kilowatts at peak, and data centers deploying hundreds of these racks are adding meaningful load to regional power grids. This is already triggering debates about renewable energy sourcing and whether the AI industry's carbon footprint is being properly accounted for.

NVIDIA's market position remains dominant despite competition from AMD's MI300X, Google's TPU v5, and Amazon's Trainium 2. The company holds an estimated 70 to 90 percent share of the AI accelerator market, a moat built not just on hardware but on CUDA, the software ecosystem that millions of AI developers have built on for over a decade. Switching away from NVIDIA means retraining workflows, rewriting code, and accepting a less mature software stack, which is why Blackwell is expected to widen NVIDIA's lead even further.