NVIDIA today announced the next evolution of its AI platform, the NVIDIA Blackwell Ultra, designed to advance AI reasoning and enhance the performance of AI applications. The introduction of Blackwell Ultra aims to significantly boost training and test-time scaling inference, which applies more compute power during inference to improve accuracy. This advancement is set to accelerate AI reasoning, agentic AI, and physical AI applications.
📣 Announced at #GTC25, EPRI, NVIDIA, and collaborators launch the Open Power AI Consortium (#OPAI).
⚡The consortium is building open datasets, #AI models, and software for use cases such as interconnection and rate case studies.
➡️ Learn more: https://t.co/xL5zy30qvW pic.twitter.com/gfbSsx40Ku
— NVIDIA AI (@NVIDIAAI) March 20, 2025
Built on the groundbreaking Blackwell architecture, Blackwell Ultra includes innovations such as the NVIDIA GB300 NVL72 rack-scale solution and the NVIDIA HGX B300 NVL16 system. The GB300 NVL72 delivers 1.5 times more AI performance than its predecessor, the NVIDIA GB200 NVL72, and increases revenue opportunities for AI factories by 50 times compared to those built with the previous NVIDIA Hopper architecture. “AI has made a giant leap — reasoning and agentic AI demand orders of magnitude more computing performance,” said Jensen Huang, founder and CEO of NVIDIA.
Is there a new Token Boom taking place in #ai that will drive @nvidia revenues? https://t.co/8XJz7W94tm
— Karl Freund (@karlfreund) March 22, 2025
“We designed Blackwell Ultra for this moment — it’s a single versatile platform that can easily and efficiently do pretraining, post-training, and reasoning AI inference.”
The NVIDIA GB300 NVL72 connects 72 Blackwell Ultra GPUs and 36 Arm Neoverse-based NVIDIA Grace CPUs in a rack-scale design, essentially creating a single, powerful GPU system for test-time scaling. This setup allows AI models to access greater compute capacity to solve complex problems more effectively, resulting in higher-quality responses. The GB300 NVL72 will offer an end-to-end, fully managed AI platform optimized for evolving workloads.
The NVIDIA HGX B300 NVL16 system offers 11 times faster inference on large language models, with 7 times more compute power and 4 times larger memory compared to the previous Hopper generation. This leap in performance supports complex workloads such as AI reasoning.
Evolving AI performance with Blackwell Ultra
Blackwell Ultra is particularly suited for applications such as agentic AI, which uses advanced reasoning and iterative planning to autonomously solve complex, multi-step problems. Additionally, physical AI applications will benefit, enabling real-time generation of synthetic, photorealistic videos for training robots and autonomous vehicles at scale. A key feature of Blackwell Ultra is its advanced scale-out networking capabilities, which are crucial for optimizing AI infrastructure.
The platform integrates seamlessly with NVIDIA Spectrum-X Ethernet and NVIDIA Quantum-X800 InfiniBand, offering 800 Gb/s data throughput for each GPU. This infrastructure significantly reduces latency and jitter, enabling AI factories and cloud data centers to manage AI reasoning models without bottlenecks. NVIDIA BlueField-3 DPUs, also featured in Blackwell Ultra systems, support multi-tenant networking, GPU compute elasticity, accelerated data access, and real-time cybersecurity threat detection.
Blackwell Ultra-based products are expected to be available from partners starting from the second half of 2025. Companies such as Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo, Supermicro, Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure are among those expected to offer servers and instances powered by Blackwell Ultra. The NVIDIA Dynamo open-source inference framework, announced alongside Blackwell Ultra, aims to scale up reasoning AI services by improving throughput while reducing response times and model serving costs.
Dynamo orchestrates and accelerates inference communication across multiple GPUs, optimizing each phase of large language models independently to ensure maximum resource utilization. Blackwell systems are compatible with new NVIDIA Llama Nemotron Reason models and the NVIDIA AI-Q Blueprint, supported in the NVIDIA AI Enterprise software platform for production-grade AI.
Image Credits: Photo by Mariia Shalabaieva on Unsplash
April Isaacs is a news contributor for DevX.com She is long-term, self-proclaimed nerd. She loves all things tech and computers and still has her first Dreamcast system. It is lovingly named Joni, after Joni Mitchell.
