Nvidia revamps Blackwell architecture with Ultra variant optimized for agentic AI workloads.
Artificial intelligence agents dominated the conversation at Nvidia Corp.'s GTC 2025 event, with the company lifting the veil on Nvidia Blackwell Ultra, its latest GPU platform evolution designed to support next-generation AI reasoning workloads. The company announced the platform alongside its first-ever Blackwell-powered desktop computers for AI developers. Blackwell Ultra provides a significant upgrade to the original Blackwell architecture announced one year earlier, improving performance and scale for both AI training and inference workloads and enabling so-called AI agents to operate autonomously, performing multiple kinds of tasks for users without more effort than a simple prompt describing the desired outcome.
Blackwell Ultra bundles the Nvidia GB300 NVL72 rack-scale solution, which packs 72 of the latest Blackwell GPUs with 36 Arm Neoverse-based Nvidia Grace central processing units running together as if they were a single super-high-performance chip. Combined with the Nvidia HGX B300 NVL16 baseboard system, this configuration delivers 1.5 times better AI performance than last year's GB200 NVL72 rack-scale setup and boosts the "revenue opportunity" for so-called AI factories by up to 50 times when compared with those powered by older Hopper architecture-based GPUs. Nvidia Chief Executive Jensen Huang justified the push for greater compute power, stating: "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 inference."
The Blackwell Ultra architecture incorporates Nvidia's Spectrum-X Ethernet and Quantum-X800 InfiniBand networking systems, which provide up to 800 gigabytes per second of data throughput for each of the 72 GPUs in the system, along with a cluster of Nvidia's BlueField-3 data processing units to handle nonessential computing tasks and free GPUs to focus solely on training and inference. The GB300 NVL72 system will be made available through the Nvidia DGX Cloud platform, a managed AI infrastructure service running on leading cloud providers including Amazon Web Services and Google Cloud, as well as on-premises through Nvidia DGX SuperPOD hardware systems. Nvidia is offering the Blackwell Ultra architecture to key server partners including Cisco Systems Inc., Dell Technologies Inc., Hewlett Packard Enterprise Co., Lenovo Group Ltd. and Supermicro Computer Inc., with additional vendors such as AsusTek Computer Inc., Gigabyte Technology Co., Ltd., Pegatron Corp. and Inventec Corp. planning to sell Blackwell Ultra-based servers later this year. The platform will also be available on Amazon Web Services, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and GPU cloud providers CoreWeave and Crusoe.
Nvidia is also launching its first "personal AI supercomputers" powered by Blackwell GPUs to enable agentic AI development in diverse locations. These systems come in two forms—DGX Spark and DGX Station—developed as part of Nvidia's Project DIGITS and designed for AI developers, researchers and data scientists to prototype, fine-tune and experiment with large language models locally. Built by companies including Asus, Dell, HP and Lenovo, these systems feature a single Nvidia GB10 Grace Blackwell Superchip combined with 5th-generation tensor cores and FP4 support, enabling them to process up to 1,000 trillion operations per second of AI compute. This gives AI developers, researchers and students local access to AI computing power comparable to what is typically found only in data centers.