NVIDIA released new performance data for its Blackwell Ultra platform, specifically the GB300 NVL72 systems. The platform delivers up to 50 times higher throughput per megawatt compared to the previous Hopper architecture. These efficiency gains result in a 35-fold reduction in the cost per token for agentic AI and coding assistant workloads.
Performance improvements stem from a combination of new system architecture and software optimizations. Microsoft, CoreWeave, and Oracle Cloud Infrastructure are currently deploying these systems at scale. The deployment targets applications requiring low latency and long context windows.
The announcement precedes NVIDIA’s upcoming earnings report, where analysts expect strong revenue growth from AI data-center solutions. Market sentiment remains positive as the Blackwell platform is viewed as a primary driver for broader enterprise AI adoption.