AI energy refers to the electricity consumed by AI systems, and in practice it is dominated by the power draw of data centers - the warehouses of processors that both train large models and serve their answers to users. As AI capability has scaled with compute, the electricity needed to supply that compute has grown into one of the field’s hardest physical constraints, sitting alongside chip availability as a question every large operator must now plan around.
The clearest first-party measure comes from the International Energy Agency’s April 2025 report “Energy and AI.” The IEA found that data centers accounted for around 1.5 percent of the world’s electricity consumption in 2024, or 415 terawatt-hours, and projected that data-center electricity consumption is “set to more than double to around 945 TWh by 2030.” The agency named “AI” as “the most important driver of this growth, alongside growing demand for other digital services.”
Within that total, the demand splits into two phases. Training a frontier model is a concentrated, weeks-to-months burst of very high power draw across thousands of accelerators. Inference - running the trained model to answer each query - is individually far smaller, but it runs continuously across enormous user populations, so as deployment scales, inference becomes the larger and steadier share of total energy demand. This is why the energy question did not shrink as training costs were better understood: serving a popular model is itself a heavy, ongoing load.
The result has been a wave of dedicated power deals. The most striking was Constellation Energy’s September 2024 agreement to sell power to Microsoft and restart an undamaged reactor at Three Mile Island under a 20-year contract, with Constellation framing the deal explicitly in terms of “data centers” that require “an abundance of energy that is carbon-free and reliable every hour of every day.” Operators have pursued nuclear, including small modular reactors, alongside large renewable purchases, partly for carbon-free supply and partly because round-the-clock reliability matters as much as price. Large infrastructure programs such as the Stargate Project are as much about securing power and grid capacity as about acquiring chips.
Why business readers should care: power is becoming a gating factor for AI, not just a cost line. Grid interconnection queues, local generation capacity, and the pace of new power projects increasingly determine where and how fast AI data centers can be built. For anyone planning AI infrastructure, energy availability - measured in megawatts and delivery dates - is now a strategic input on the same footing as compute itself.