NVL72
An NVIDIA rack connecting 72 Blackwell GPUs into one NVLink domain that behaves as a single huge accelerator.
Current numbers
3.6 TB/sNVLink 6 per-GPU bandwidth (Rubin); 900 GB/s Hopper, 1.8 TB/s Blackwell — ~260 TB/s per NVL72 rack
~10× (up to 10×)Vera Rubin NVL72 tokens/s per megawatt vs GB200 NVL72 (CoreWeave, measured on DeepSeek-R1 at matched interactivity; ratio only)
~1/10 (one-tenth)Vera Rubin NVL72 inference cost per million tokens vs GB200 NVL72 (ratio only; no absolute $/token published)