NVIDIA has unveiled the NVIDIA Blackwell Ultra AI factory platform, marking a significant advancement in AI reasoning capabilities. This new platform enhances training and inference efficiency, allowing organizations to accelerate applications in AI reasoning, agentic AI, and physical AI. Building on last year's Blackwell architecture, Blackwell Ultra integrates the NVIDIA GB300 NVL72 rack-scale solution and the HGX™ B300 NVL16 system, offering 1. 5 times more AI performance than the previous generation and increasing potential revenue opportunities for AI factories by 50 times compared to those using NVIDIA Hopper™. Jensen Huang, NVIDIA's CEO, emphasized that the Blackwell Ultra has been designed to meet the demands of advanced AI performance, capable of efficiently handling pretraining, post-training, and inference tasks. The GB300 NVL72, featuring 72 Blackwell Ultra GPUs and 36 Arm-based NVIDIA Grace™ CPUs, acts as a unified GPU designed for scaling during inference. This setup allows AI models to leverage enhanced compute capacity for more comprehensive problem-solving and complex requests. The platform will also be available on NVIDIA DGX™ Cloud, offering an end-to-end managed AI solution, while the HGX B300 NVL16 boasts 11 times quicker inference and significantly improved compute and memory capabilities compared to its predecessor, making it suitable for complex workloads. Blackwell Ultra supports various applications including agentic AI for autonomous problem solving and physical AI for generating synthetic videos for training robots and autonomous vehicles. Advanced networking capabilities are integral to the Blackwell Ultra systems, featuring the NVIDIA Spectrum-X™ Ethernet and Quantum-X800 InfiniBand platforms for optimal performance and low latency.

The inclusion of NVIDIA BlueField®-3 DPUs further enhances networking and cybersecurity features. Products based on Blackwell Ultra are expected to be available from partners in the second half of 2025, with companies like Cisco, Dell, and Lenovo planning to release servers based on this technology. Additionally, major cloud and GPU service providers will offer Blackwell Ultra-powered instances. Finally, NVIDIA supports the Blackwell platform with its full-stack AI ecosystem, including the new open-source NVIDIA Dynamo framework designed for efficient AI inference and enhanced model serving performance. This ecosystem enables extensive resource optimization for scalable AI solutions, further solidifying NVIDIA's leadership in accelerated computing innovation. For more information, NVIDIA encourages attending the GTC keynote and related sessions through March 21.