Data centers use as little as 60% of power consumed for actual AI work, Nvidia claims — failures and restarts 'consume' a whopping 10% of every watt used

Nvidia claims that for each gigawatt an AI data center uses, only 70% is used for AI work Inefficiencies include cooling and power distribution, failures, and restarts While a chip shortage was considered to be the main bottleneck to AI data center capacity, it is now a facility’s connection to the grid AI data centers are inefficient, says Nvidia, with various physical and systems issues contributing to power use beyond AI processing.

This opinion, shared by data center builder Vertiv at its recent annual EMEA event, underpins future development steps, in which it plans to assemble sustainable facilities.

Vertiv and Nvidia have been collaborating since early 2026 on co-designing the next generation of data centers, with Vertiv providing power and infrastructure management for Nvidia’s GPUs.

With a maximum of 70% of electricity drawn from the grid used for AI processing, the remaining 30% represents inefficiencies that optimized data center designs for the future should reduce.

That's not efficient Although the various inefficiencies affecting AI data centers may not come as a surprise, the scale of them might.

According to Nvidia – via its Vice President of Supercomputing, AI Cloud Infrastructure, and AI/HPC for EMEA Rod Evans – that 30% consists of infrastructure and system issues.

For a one-gigawatt data center, 20% is lost through cooling, power distribution, and conversion.

A further 10% is lost to incorrectly configured racks, as well as failures and restarts.

Evans says that “You sign a contract for 100 megawatts, but you never actually get that.

With a bit of luck, you’ll get 93 megawatts, and at certain times of the day, even less.” Nvidia believes some of the 30% (which can be as high as 40%) can be recovered.

Evans notes that “A data center is actually designed for peak load and outages, not for what it typically uses.” This means that unused capacity can be tapped, and power and cooling can be revised to find the optimum configurations.

Bottlenecks of power Rod Evans states that Nvidia and Vertiv are planning data centers that are more efficient and sustainable.

“Together, we want to build sustainable AI data centers where the focus isn’t on using as much power as possible, but on getting the most out of every watt.” One of the most significant details from the press conference is the challenge posed to AI data centers not by a shortage of chips, but by a lack of power.

Bottlenecks are resulting in underpowered data centers, with some waiting for several years to receive the power to match their capacity.