NVIDIA has announced Blackwell Ultra, its next-generation data-center GPU architecture built for large AI model training and inference. The company claims up to four times the performance of standard Blackwell chips for certain workloads.
Built for massive models
Blackwell Ultra increases memory bandwidth and supports larger model deployments in a single server. NVIDIA says it is optimized for trillion-parameter models, mixture-of-experts architectures, and real-time inference at scale.
Cloud availability
AWS, Google Cloud, Microsoft Azure, and Oracle Cloud have announced early access programs. Dell, HPE, and Supermicro are also building Blackwell Ultra systems.
Energy efficiency
NVIDIA emphasized performance per watt, claiming Blackwell Ultra is more efficient than competing solutions. This matters as data center power becomes a limiting factor for AI growth.