Nvidia data centers exceed utility power promises, raising concerns about AI’s energy appetite

A single Nvidia GB300 rack reportedly draws over 180 kilowatts. For context, that’s roughly the electricity load of 60 average American homes, crammed into a server cabinet. Scale that up to a full AI cluster exceeding 100 megawatts, and you’re looking at something that rivals the electrical demand of a small city. The numbers paint a stark picture US data centers consumed approximately 176 terawatt-hours (TWh) in 2025, accounting for about 4.4% of total national power generation. But projections suggest it could balloon to somewhere between 466 and 580 TWh by 2030, pushing data centers to consume 9-12% of all US electricity generation. Goldman Sachs forecasts US data center power demand climbing from 31 gigawatts in 2025 to 66 gigawatts by 2027. That’s more than a doubling in just two years. The supply side looks far less ambitious. Utilities can realistically deliver about 93 GW of additional practical power supply, leaving a substantial shortfall relative to AI data center demand. The bottleneck isn’t chips anymore A new constraint has jumped the queue: physical infrastructure. Transformers, switchgear, and other grid equipment are now in critically short supply. Without them, e...