Energy Efficiency
Energy efficiency measures the useful computational work completed per unit of energy consumed—or, equivalently, the energy required per task. In AI infrastructure, it is expressed through metrics such as tokens per kilowatt-hour or joules per token, reflecting how much total energy is needed to generate a given amount of AI output. Performance per watt is the closely related rate-based measure of how efficiently a system converts electrical power into throughput, such as tokens per second per watt. Together, these metrics help AI factory operators maximize output, reduce energy cost per task, and make better use of fixed power capacity.
Inference efficiency turns fixed power into more billable AI output by increasing throughput per watt across the full AI factory.
Lower precision cuts AI energy use by reducing data movement and compute cost while calibration and validation protect accuracy.
Lower AI serving energy cost by improving useful inference per watt across models, routing, utilization, rack design, and cooling.
Learn how to model AI data center power, heat, cooling, water use, and rack-scale efficiency before construction begins.
Operators are using accelerated computing, liquid cooling, rack-scale design, and smarter networking to increase AI throughput per watt.
Learn how data centers reclaim stranded power capacity with accelerated computing, efficient networking, and liquid-cooled AI infrastructure.
Power-constrained AI operators are choosing rack-scale, liquid-cooled AI factory infrastructure built for performance per watt.
Purpose-built NVIDIA accelerated AI infrastructure gets more training done per watt by optimizing compute, networking, cooling, and utilization.
Learn the key metrics for AI data center efficiency, from AI performance per watt to GPU utilization, cooling power, and water use.
Infrastructure determines AI energy use by shaping performance per watt across compute, networking, cooling, and the full AI factory.