Nvidia and Vertiv Outline Heavy Power Losses Across AI Data Centers

Data centers built for artificial intelligence workloads direct as little as 60 percent of their incoming electrical power toward active computing operations, according to executive remarks reported by TechRadar Pro (https://www.techradar.com/pro/data-centers-use-as-little-as-60-percent-of-power-consumed-for-actual-ai-work-nvidia-claims-failures-and-restarts-consume-a-whopping-10-percent-of-every-watt-used). The remaining 30 to 40 percent of incoming energy is absorbed by infrastructure overhead, thermal management, and operational disruptions.

Speaking at an annual EMEA event held by data center infrastructure provider Vertiv, Rod Evans, Nvidia's vice president of supercomputing, AI cloud infrastructure, and AI/HPC for EMEA, outlined the distribution of those losses. Evans stated that in a typical one-gigawatt facility, approximately 20 percent of total energy is lost during cooling, power distribution, and power conversion. An additional 10 percent is consumed by incorrectly configured server racks, hardware faults, and system restarts.

Evans also pointed to discrepancies between contracted grid capacity and real-world power delivery. Operators contracting for 100 megawatts frequently receive around 93 megawatts under favorable conditions, with delivered power dropping further during specific times of the day. Because data centers have historically been designed around peak load requirements and outage contingencies rather than average operational baselines, Evans noted that facilities maintain unutilized headroom that could be recovered through revised power delivery and cooling designs.

Nvidia and Vertiv have been collaborating since early 2026 on co-designing data center architectures, with Vertiv providing power management and infrastructure systems for Nvidia GPUs. The joint effort seeks to maximize computational throughput per watt rather than relying purely on higher aggregate power consumption.

The operational inefficiencies emerge as data center growth faces grid constraints rather than processor manufacturing shortages. Infrastructure operators in multiple regions face multi-year waiting periods to secure grid connections that match their planned capacity.

Vertiv EMEA President Paul Ryan cited regional regulatory and logistical hurdles complicating new construction. Ryan noted that some jurisdictions grant grid connections without building permits, while others issue permits quickly but lack available power. In other instances, locations that secure both power access and permits are situated in remote regions that complicate equipment transit.