What We Know About Energy Use at U.S. Data Centers Amid the AI Boom (www.pewresearch.org)

The rapid buildout of hyperscale data centers to support AI training and inference is already reshaping U.S. energy use and local infrastructure. Industry databases estimate over 4,000 U.S. data centers, with heavy clustering in Virginia, Texas and California; half of new builds sit within existing hubs. The IEA estimates U.S. data centers consumed 183 TWh in 2024 (about 4% of national electricity) and projects consumption could more than double to 426 TWh by 2030. A single AI-focused hyperscaler can draw as much power annually as ~100,000 households, and the biggest sites under construction may use 20× that amount—putting concentrated stress on regional grids and utility capacity markets. Key technical impacts: roughly 60% of data-center electricity powers servers (AI accelerators consume 2–4× the watts of traditional chips), while cooling accounts for 7–30% of use and drives large water demand—U.S. data centers directly used ~17 billion gallons in 2023, with hyperscalers projected to consume 16–33 billion gallons annually by 2028. The current energy mix supplying data centers is ~40% natural gas, ~24% renewables, ~20% nuclear and ~15% coal; natural gas is expected to remain dominant although companies are signing PPAs for advanced nuclear and states are weighing renewable mandates/reporting. These trends carry real costs: grid upgrades can raise residential bills (e.g., a $9.3B PJM capacity impact tied to data centers) and studies project up to an ~8% national bill increase by 2030, much higher in regional hotspots—making energy sourcing, efficiency and policy central concerns for the AI/ML community.