Nvidia Says AI Data Centers Use Up to 40% of Power on Inefficiencies

Nvidia says AI data centers can use as little as 60% of their electricity for AI processing, with the rest lost to infrastructure and system inefficiencies. The figures were shared by Rod Evans, Nvidia’s vice president of Supercomputing, AI Cloud Infrastructure, and AI/HPC for EMEA, as reported by TechRadar.

For a 1-gigawatt data center, Evans said 20% of power is lost through cooling, power distribution, and conversion. Another 10% goes to incorrectly configured racks, failures, and restarts. Nvidia says total inefficiencies can reach 40%, though the share used for AI processing can be as high as 70%.

Evans said a facility contracted for 100 megawatts might deliver 93 megawatts, or less at some times of day. He also said data centers are designed for peak loads and outages, leaving unused capacity that could be recovered by adjusting power and cooling systems.

Nvidia and infrastructure company Vertiv have collaborated since early 2026 on the design of future data centers, with Vertiv providing power and infrastructure management for Nvidia GPUs. They say they aim to build more efficient, sustainable facilities.

At Vertiv’s recent annual EMEA event, company president for EMEA Paul Ryan described power availability and permits as separate constraints. Some sites have power but lack permits; others have permits but no power. Remote locations can also make equipment difficult to deliver. The report says some data centers wait several years for enough power to match their capacity.