NVIDIA's GB300 Posts 20x Agentic Leap as Abridge Healthcare Model Lands

NVIDIA reported that its Blackwell Ultra GB300 set records on the new AA-AgentPerf benchmark — a standardized measure of agentic AI inference — running such workloads roughly 20x faster than the prior Hopper generation, as the next-gen Rubin architecture nears launch. NVIDIA also touted leading performance on the industry's first agentic AI coding benchmark, signaling a push to define how agent capability is measured on its hardware.

On the application side, NVIDIA and Abridge are co-developing a healthcare-specific foundation model built on Blackwell infrastructure and the open Nemotron model family, powering Abridge's clinician-intelligence platform with an enterprise-wide rollout at Northwestern Medicine. NVIDIA additionally expanded AI-infrastructure partnerships in Korea, including next-generation memory work with SK hynix.

The combination — benchmark dominance plus vertical deployments — reflects NVIDIA's strategy of pairing raw compute leadership with sticky industry use cases. The 20x agentic claim, in particular, targets the fastest-growing inference category as agents move into production.

Watch for independent verification of the AA-AgentPerf results, Rubin's launch timing, and competitive responses from custom-silicon efforts.