The NVIDIA Blackwell Ultra, the next development of the NVIDIA Blackwell AI factory platform, was unveiled byit, opening the door to the era of AI reasoning.
Organizations worldwide may now speed applications like AI reasoning, agentic AI, and physical AI with NVIDIA Blackwell Ultra’s enhancement of training and test-time scaling inference, which is the skill of using extra computing during inference to improve accuracy.
Blackwell Ultra NVIDIA, which is based on the revolutionary Blackwell architecture that was unveiled a year ago, features the NVIDIA HGX B300 NVL16 system and the NVIDIA GB300 NVL72 rack-scale solution. In addition to providing 1.5 times more AI performance than the NVIDIA GB200 NVL72, the GB300 NVL72 boosts Blackwell’s revenue potential for AI factories by 50 times as compared to those constructed with NVIDIA Hopper.
NVIDIA designed Blackwell Ultra for this moment, which is a single versatile platform that can easily and efficiently do pretraining, post-training, and reasoning AI inference. AI has made a giant leap reasoning and agentic AI require orders of magnitude more computational power.
NVIDIA Blackwell Ultra Enables AI Reasoning
The NVIDIA GB300 NVL72 links 36 Arm and 72 Blackwell Ultra NVIDIA GPUs.
Neoverse-based NVIDIA Grace CPUs that function as a single, enormous GPU designed for test-time scalability in a rack-scale configuration. AI models can use the NVIDIA GB300 NVL72’s enhanced computing power to investigate various problem-solving strategies and deconstruct intricate requests into several phases, producing results of superior quality.
NVIDIA DGX Cloud, an end-to-end, fully managed AI platform on top clouds that optimizes performance with software, services, and AI expertise for changing workloads, is also anticipated to provide the GB300 NVL72. The GB300 NVL72 rack architecture is used by NVIDIA DGX SuperPOD with DGX GB300 computers to give clients a turnkey AI factory.
In comparison to the Hopper generation, the NVIDIA HGX B300 NVL16 offers breakthrough performance for the most difficult applications, including AI reasoning, with 11x quicker inference on big language models, 7x more computation, and 4x larger memory.
The Blackwell Ultra NVIDIA platform is also perfect for the following applications:
- Agentic AI handles complex, multi-step problems utilizing advanced reasoning and iterative planning. AI agent systems do more than just follow instructions. They are able to plan, reason, and act to accomplish particular objectives.
- Physical AI, which enables businesses to produce artificially realistic, real-time videos for the mass training of robots and self-driving cars.
NVIDIA Scale-Out Infrastructure for Optimal Performance
A key element of AI infrastructure that can provide optimal performance while lowering latency and jitter is advanced scale-out networking.
Through an NVIDIA ConnectX-8 SuperNIC, Blackwell Ultra NVIDIA systems can interface with the NVIDIA Spectrum-X Ethernet and NVIDIA Quantum-X800 InfiniBand platforms, offering 800 Gb/s of data throughput per GPU in the system. With the best remote direct memory access capabilities available, cloud data centers and AI factories can handle AI reasoning models without experiencing bottlenecks.
Multi-tenant networking, GPU computing elasticity, rapid data access, and real-time cybersecurity threat detection are all made possible by NVIDIA BlueField-3 DPUs, which are also found in Blackwell Ultra NVIDIA systems.
Global Technology Leaders Embrace Blackwell Ultra NVIDIA
It is anticipated that partners will begin offering Blackwell Ultra NVIDIA-based goods in the second half of 2025.
In addition to Aivres, ASRock Rack, ASUS, Eviden, Foxconn, GIGABYTE, Inventec, Pegatron, Quanta Cloud Technology (QCT), Wistron, and Wiwynn, a variety of servers based on Blackwell Ultra NVIDIA products are anticipated to be delivered by Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo, and Supermicro.
providers of cloud services Among the first to provide Blackwell Ultra-powered instances are GPU cloud providers CoreWeave, Crusoe, Lambda, Nebius, Nscale, Yotta, and YTL, as well as Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure.
NVIDIA Software Innovations Reduce AI Bottlenecks
The full-stack NVIDIA AI platform supports the whole range of NVIDIA Blackwell products. It was also announced that the NVIDIA Dynamo open-source inference framework scales up reasoning AI services, offering the most effective way to expand test-time compute, resulting in leaps in throughput while lowering response times and model serving costs.
The new AI inference-serving program NVIDIA Dynamo was created to optimize token revenue generation for AI factories that use reasoning AI models. In addition to using disaggregated serving to divide the processing and generation stages of big language models over multiple GPUs, it coordinates and speeds up inference communication among thousands of GPUs. This guarantees optimal GPU resource consumption and enables each phase to be separately optimized for its unique requirements.
The NVIDIA AI-Q Blueprint and the new NVIDIA Llama Nemotron Reason models, which are supported by the NVIDIA AI Enterprise software platform for production-grade AI, perform best on Blackwell systems. In addition to AI frameworks, libraries, and tools that businesses can implement on NVIDIA-accelerated clouds, data centers, and workstations, NVIDIA AI Enterprise now incorporates NVIDIA NIM microservices.
NVIDIA’s ecosystem of robust development tools, NVIDIA CUDA-X libraries, more than 4,000 applications that scale performance across thousands of GPUs, and more than 6 million developers are all built upon by the Blackwell platform.
