NVIDIA confirms Vera Rubin full production; Rubin GPU leads AgentPerf with 20x efficiency over Hopper

NVIDIA CEO Jensen Huang confirmed at GTC Taipei (Computex 2026 on June 1) that Vera Rubin GPU has entered full production, with partner availability beginning H2 2026. The Rubin GPU features 336 billion transistors on TSMC N3, 288GB HBM4 (double Blackwell capacity), 22 TB/s memory bandwidth, delivering 50 PFLOPS NVFP4 inference and 35 PFLOPS training per single GPU.

Vera Rubin's benchmark leadership is now tangible: NVIDIA's Blackwell Ultra NVL72 rack (72 GPUs) led AgentPerf, the first agentic AI benchmark from Artificial Analysis, running 20x more agents per megawatt than Hopper in equivalent configurations. This efficiency gain signals that Rubin—delivering 1.5x theoretical performance per GPU over Blackwell Ultra—is positioned to reshape cost-per-inference economics for production AI services running long-duration agentic workloads.

NVIDIA announced a multiyear memory partnership with SK hynix (June 7) to codevelop next-generation memory for Vera platforms, spanning Rubin GPUs, Vera CPUs, RTX Spark PCs, and Jetson Thor. SK hynix will use NVIDIA CUDA-X and PhysicsNeMo for semiconductor simulation. This supply-side lock-in reflects the structural memory crunch: hyperscalers cannot afford spot allocation risk.

For infrastructure buyers and capacity planners, Vera Rubin's H2 2026 availability and AgentPerf leadership imply a 12–18 month transition window to compete on inference cost. Hyperscalers with existing Hopper fleets face the ROI math of consolidating capital toward Rubin, while smaller AI service providers must evaluate whether holding pre-Rubin capacity into 2027 is viable given production demand and rising memory costs.