At a Vertiv event, Nvidia said infrastructure operation and system inefficiencies can consume 30% of AI data-center power, and in some cases 40%. Its example of a 1 GW facility assigned 20% to cooling, power distribution and conversion, and 10% to rack-configuration issues, failures and restarts. It also noted that contracted power is not always fully delivered and some facilities may wait years for supply. These are vendor examples and estimates, not measured averages for all centers, and necessary cooling power cannot all be classified as avoidable waste.