THE CRUNCH
NVIDIA is pushing a new metric for AI infrastructure: tokens per megawatt. At the AI Infra Summit, the company and partners showcased how the Vera Rubin platform and DSX software can boost efficiency. Lambda reported a 23% improvement in performance per watt using DSX MaxLPS, while Emerald AI demonstrated a flexible-load program that lets AI factories throttle power during grid demand spikes. The goal is to make AI a
The goal is to make AI a controllable grid resource rather than a fixed energy load.
The DSX MaxLPS software dynamically shifts power across racks to match workload needs. Lambda showed this can fit 19 nodes into the power budget of 16, increasing token throughput by 24%. For the upcoming Vera Rubin NVL72 systems, NVIDIA claims this approach could unlock up to 40% more GPU capacity within the same megawatt envelope.
Beyond software, NVIDIA is designing the Vera Rubin platform to handle power spikes. Intelligent Power Smoothing software and expanded energy buffering absorb short surges, letting systems run closer to sustained demand. This helps prevent power caps from throttling performance during bursts of activity.