Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency

Power is an inescapable constraint in AI infrastructure, determining how many tokens an AI factory can generate within a fixed power budget, which in turn affects revenue and profitability. Modern AI models, such as the NVIDIA Blackwell NVL72 platform, require co-design across all system layers to deliver high performance per watt, maximizing revenues and minimizing token costs. Each new generation of models introduces architectural changes that necessitate new optimizations. NVIDIA GB300 NVL72 systems achieve up to 25x performance per watt compared to the NVIDIA Hopper generation, showcasing continuous improvement.