The technology company NVIDIA has announced the evolution of its artificial intelligence factory platform, NVIDIA Blackwell Ultra, during the GTC event. This new advancement promises to usher in a new era in AI development, focused on reasoning, autonomous planning, and complex decision-making.
Blackwell Ultra builds on the Blackwell architecture presented last year but significantly multiplies its capabilities. With its flagship system, the NVIDIA GB300 NVL72, the company offers performance that is 1.5 times higher than the previous model and up to 50 times increased revenue potential for AI factories compared to installations based on the Hopper architecture.
“We’re witnessing a giant leap,” said Jensen Huang, founder and CEO of NVIDIA. “Reasoning AI and autonomous agents demand orders of magnitude more computing power, and Blackwell Ultra has been designed for this moment.”
A Large-Scale Infrastructure for a New Type of AI
The GB300 NVL72 system integrates 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs based on Arm, acting as a single giant processor capable of performing scaled inferences during the testing phase. This approach allows AI models to explore alternative solutions and break down complex requests into intermediate steps, thereby improving the quality and accuracy of responses.
Meanwhile, the NVIDIA HGX B300 NVL16 system provides 11 times more speed in large-scale language model inferences, with four times more memory and seven times the computing power compared to the previous generation.
Key Applications: Agentive and Physical AI
The platform is particularly geared toward applications that require iterative planning and autonomous decision-making (agentive AI), as well as physical AI capable of generating photorealistic synthetic videos in real-time. This has a direct impact on areas such as robotic systems training and autonomous vehicles.
High-Performance Networks and Advanced Security
To avoid bottlenecks and ensure optimal performance, Blackwell Ultra systems integrate connectivity with NVIDIA Spectrum-X Ethernet and NVIDIA Quantum-X800 InfiniBand platforms, offering up to 800 Gb/s bandwidth for each GPU. Additionally, they include NVIDIA BlueField-3 DPU processors, which enable multi-client management, real-time cybersecurity, and accelerated data access.
Adoption by Tech Giants and Cloud Providers
The first products based on Blackwell Ultra are expected to be available in the second half of 2025. Leading companies such as Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo, and Supermicro are already preparing compatible servers, along with companies like ASUS, GIGABYTE, Quanta, and Foxconn.
The major cloud service providers —Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud— will also be among the first to offer instances powered by Blackwell Ultra, along with specialized providers like CoreWeave, Lambda, and Nebius.
NVIDIA Dynamo: The New Software for Scaling AI Services
The introduction of Blackwell Ultra was accompanied by the announcement of NVIDIA Dynamo, a new open-source inference framework designed to maximize performance and reduce costs in implementing reasoning AI models. Dynamo allows the orchestration of processes across thousands of GPUs, optimizing each phase (processing and generation) independently to achieve maximum resource utilization.
Moreover, Blackwell Ultra is the ideal environment for the Llama Nemotron Reason models and the NVIDIA AI-Q Blueprint architecture, included in the NVIDIA AI Enterprise platform for enterprise AI solutions.
With this move, NVIDIA not only consolidates its leadership in artificial intelligence hardware but also positions itself as the key player defining the infrastructure for the next great technological revolution: an AI capable of reasoning, planning, and acting autonomously.
via: Nvidia