NVIDIA unveils Blackwell Ultra for advanced AI reasoning

NVIDIA has announced the next phase of its Blackwell AI platform, known as NVIDIA Blackwell Ultra, which is designed to facilitate advancements in AI reasoning capabilities.

The NVIDIA Blackwell Ultra platform aims to enhance both training and inference scaling, allowing organisations to improve the accuracy of AI reasoning, agentic AI, and physical AI applications.

This development is predicated on the Blackwell architecture introduced last year and includes new systems such as the NVIDIA GB300 NVL72 and NVIDIA HGX B300 NVL16.

Jensen Huang, Founder and CEO of NVIDIA, stated: "AI has made a giant leap — reasoning and agentic AI demand orders of magnitude more computing performance. 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 utilises a rack-scale design that connects 72 Blackwell Ultra GPUs and 36 Arm Neoverse-based NVIDIA Grace CPUs. This configuration allows the system to operate as a single large GPU designed explicitly for test-time scaling, potentially leading to higher-quality responses from AI models.

NVIDIA's GB300 NVL72 is anticipated to be integrated with NVIDIA DGX Cloud, a comprehensive managed AI platform that aims to optimise performance. Additionally, NVIDIA DGX SuperPOD utilises GB300 NVL72 rack design to facilitate the creation of turnkey AI factories for customers.

The NVIDIA HGX B300 NVL16 system offers significant improvements, boasting 11 times faster inference on large language models and various increases in computing and memory capabilities compared to previous generations.

The advancements in the Blackwell Ultra platform are also expected to assist applications like agentic AI, where systems use sophisticated reasoning to solve complex problems autonomously. NVIDIA claims the platform aids in processing real-time photorealistic videos for training applications such as robotics and autonomous vehicles, aligning with physical AI developments.

In terms of AI infrastructure, NVIDIA is focusing on advanced scale-out networking to reduce latency and jitter. Blackwell Ultra systems are integrated with NVIDIA Spectrum-X and Quantum-X800 InfiniBand platforms, ensuring efficient data throughput and supporting AI reasoning models without bottlenecks.

The platform also incorporates NVIDIA BlueField-3 DPUs, which offer benefits such as multi-tenant networking, GPU compute elasticity, and real-time cybersecurity threat detection, enhancing their utility in data-centre environments.

Products based on the Blackwell Ultra platform are expected to be available in the latter half of 2025. Companies like Cisco, Dell Technologies, and Hewlett Packard Enterprise, along with other notable technology leaders, are expected to introduce servers that leverage Blackwell Ultra products.

Furthermore, major cloud service providers including Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure will be among the first to offer Blackwell Ultra-powered instances. This is complemented by the support from GPU cloud providers such as CoreWeave and Lambda.

NVIDIA also revealed the NVIDIA Dynamo open-source inference framework, which is designed to maximise efficiency in scaling reasoning AI services. It aims to enhance throughput and reduce response times and operating costs for AI factories.

The Blackwell systems appear to be particularly suitable for new NVIDIA Llama Nemotron Reason models and are supported by the NVIDIA AI Enterprise software, which provides a suite of tools for deploying AI solutions across various platforms.

The Blackwell platform builds on NVIDIA's extensive ecosystem, including its suite of development tools, CUDA-X libraries, and applications targeted at scaling performance across a substantial developer base.