Why Are AI Data Centers Consuming So Much Electricity?
Training frontier models grabs attention, but running billions of daily AI queries on ultra-dense GPU racks is what is straining regional power grids.
For years, efficiency gains in server chips kept total data center electricity growth surprisingly flat even as cloud computing exploded. Generative AI broke that curve. Modern AI campuses now draw as much power as mid-sized cities, forcing utilities to revise decades-old demand forecasts and pushing tech giants into direct deals for nuclear, geothermal, and natural gas generation.
The surge comes down to the physics of dense parallel computing. Traditional web servers spend much of their time waiting on databases or serving lightweight requests. AI accelerators run flat out for weeks during model training and stay heavily loaded all day answering reasoning queries, generating images, and powering autonomous coding agents.
Rack density and the cooling multiplier
A conventional enterprise server rack might draw 7 to 15 kilowatts of electricity. By contrast, a modern liquid-cooled rack packed with frontier AI accelerators easily crosses 100 kilowatts, concentrating extraordinary heat inside a tiny footprint. Air conditioning alone can no longer strip heat away fast enough, forcing operators to build direct-to-chip liquid cooling loops, heat exchangers, and massive chiller plants.
Every watt pushed into a silicon chip turns directly into heat that must be pumped out of the building. Even in highly efficient facilities, cooling systems, power conversion equipment, and high-speed optical networking switches add significant overhead on top of the chips themselves.
Inference scales relentlessly with usage
Early coverage focused heavily on the power needed to train one giant model. In practice, inference, which is the ongoing cost of running trained models for millions of users, now accounts for a huge and steadily growing share of total energy consumption. Reasoning models that think through dozens of hidden steps before outputting an answer multiply the electricity spent on a single prompt.
Because building new electrical substations, transformers, and long-distance transmission lines can take five to ten years, access to reliable gigawatts of electricity has quietly become the hardest bottleneck in AI expansion, rivaling the supply of chips themselves.