NVIDIA says performance per watt is the defining metric for AI infrastructure efficiency

NVIDIA said performance per watt has become the defining metric for AI infrastructure because power is the inescapable constraint on how many tokens an AI factory can generate within a fixed power budget, which determines revenue and profitability. The company said mixture-of-experts models now dominate the frontier and serving them at rack scale demands codesign across every layer of the system and software stack, plus operational depth from running these models under real production load. NVIDIA said its Blackwell NVL72 platform delivers that rack-scale foundation today, and the upcoming Vera Rubin platform builds on it to further elevate rack-scale energy efficiency. The company said its GB300 NVL72 system delivers up to 25 times the performance per watt of the prior Hopper generation across the newest generation of leading open models, though it noted these numbers reflect where Blackwell stands today and continue to improve. Different workloads demand different operating points, some optimize for latency, others for throughput and cost, and most need to move between them. NVIDIA said it presents Pareto curves for each model rather than a single number and provides tools such as DynoSim to help teams find their optimal point on the Pareto frontier.