The rapid evolution of artificial intelligence has been driven by relentless advances in computing power. NVIDIA has taken another significant step in this journey with the launch of its Blackwell Ultra AI factory platform, a move that could reshape how AI systems reason, learn and interact with the world. The platform, designed to support the increasing demands of AI reasoning, agentic AI, and physical AI, offers a new level of computational efficiency and scalability.
Blackwell Ultra builds on the Blackwell architecture introduced last year, enhancing training and inference capabilities to deliver faster and more accurate AI processing. Central to this development is the NVIDIA GB300 NVL72, a rack-scale system that connects 72 Blackwell Ultra GPUs and 36 Grace CPUs, creating a single, massive GPU. This enables AI models to scale inference dynamically, improving their ability to handle complex queries and solve problems through multi-step reasoning.
Scaling AI beyond traditional limits
The shift toward AI reasoning represents a fundamental change in how AI models operate. Rather than merely following pre-set instructions, reasoning AI systems must engage in iterative planning and problem-solving. The Blackwell Ultra platform is built for this new phase, providing up to 1.5 times the AI performance of its predecessor, the GB200 NVL72, and significantly expanding the computational potential for AI factories.
With this capability, AI can generate higher-quality responses by exploring multiple solution pathways in real time. This is particularly critical for applications such as agentic AI, where models must autonomously break down complex requests into logical steps, and physical AI, which enables real-time synthetic video generation for robotics and autonomous vehicle training.
Performance is further enhanced through NVIDIA’s scale-out networking solutions. The Blackwell Ultra platform integrates with the Spectrum-X Ethernet and Quantum-X800 InfiniBand systems, offering 800 Gb/s throughput per GPU. This level of data processing power helps AI factories avoid bottlenecks, ensuring efficient performance at scale. The inclusion of NVIDIA BlueField-3 DPUs also enables multi-tenant networking, accelerating AI operations while strengthening cybersecurity.
Preparing for AI’s next leap
The launch of Blackwell Ultra has already drawn commitments from some of the world’s leading technology firms. Companies including Cisco, Dell Technologies, Hewlett Packard Enterprise, and Lenovo plan to deliver server solutions based on the new architecture. Cloud providers such as AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure will also be among the first to offer Blackwell Ultra-powered instances, making high-performance AI accessible to a global audience.
At the heart of this advancement is software innovation. NVIDIA Dynamo, a new open-source AI inference framework, is set to streamline AI reasoning by optimising throughput while reducing response times and operational costs. The framework is designed to scale across thousands of GPUs, ensuring efficient AI processing while maximising hardware utilisation.
The expansion of AI reasoning capabilities marks a significant shift in how businesses and researchers will engage with artificial intelligence. With Blackwell Ultra, NVIDIA is not simply pushing the limits of hardware performance but redefining what AI can achieve in an era of increasing complexity and autonomy. As AI systems move from simple instruction-following to more advanced reasoning and decision-making, the infrastructure supporting them must evolve. Blackwell Ultra provides a glimpse into that future, offering the scale, speed, and efficiency required for AI to become a more intelligent, adaptable, and indispensable part of our world.
