NVIDIA GTC 2026 Keynote: NemoClaw Enterprise Agent Platform Debuts Alongside Vera Rubin NVL72 — 5x Inference Over Blackwell, 288GB HBM4

DEV Community·high signal

Jensen Huang's March 16 keynote introduced NemoClaw, an open-source enterprise agent deployment platform designed to mirror CUDA's lock-in strategy at the application layer. Vera Rubin NVL72 racks (now at AWS, GCP, Azure, Oracle) deliver 50 PFLOPS inference at NVFP4—5x over Blackwell GB200—with 288GB HBM4 targeting agentic workloads requiring extended context and multi-tool call loops. Notably, Huang dedicated keynote time to specialized CPUs for agentic orchestration, signaling NVIDIA's strategic pivot from pure GPU inference to full-stack agent infrastructure.