NVIDIA GTC 2026 Opens: Vera Rubin NVL72 Platform Formally Launches — 10x Lower Inference Cost Than Blackwell, H2 2026 Availability

NVIDIA Newsroom·high signal

NVIDIA's GTC 2026 keynote by Jensen Huang on March 16 formally kicked off the Vera Rubin era: 72-GPU NVL72 rack with 50 PFLOPS/GPU (NVFP4), 260 TB/s NVLink 6 bandwidth, BlueField-4 AI-native storage for agentic KV-cache sharing, and 10x lower inference cost per token vs. Blackwell. New NVIDIA Inference Context Memory Storage Platform introduced to accelerate agentic AI reasoning. Available from AWS, Google Cloud, Microsoft, CoreWeave, and Nebius in H2 2026.