Data Center Electricity Consumption and AI Workloads
Global data centers consumed roughly 460 TWh of electricity in 2022, with IEA projections of 620 to over 1,000 TWh by 2026 — the upper bound equivalent to all of Japan's annual electricity use. AI workloads are the principal swing factor, and consumption is geographically concentrated in a handful of grid pockets that are already showing strain.
The International Energy Agency's Electricity 2024 report estimated global data center electricity consumption at about 460 TWh in 2022, projecting a range of roughly 620 to more than 1,000 TWh by 2026 when cryptocurrency mining and AI workloads are bundled in. The high end is comparable to Japan's total national electricity demand. A 2025 follow-up modelling exercise puts the 2030 figure near 945 TWh. In the United States, a 2024 Lawrence Berkeley National Laboratory report tracked usage rising from 58 TWh in 2014 to 176 TWh in 2023 — about 4.4% of the national total — and projected 325 to 580 TWh by 2028, or 6.7% to 12% of U.S. electricity. The bulk of that growth is attributed to hyperscale buildout for AI rather than to conventional cloud or enterprise workloads. It helps to separate three workload classes. AI training runs are bursty, multi-week campaigns on tightly coupled GPU clusters that can swing power draw by 30 to 60 percent within milliseconds as kernels move between compute, communication, and synchronization. AI inference is steady-state and lower per-query, but cumulative across billions of requests it is now estimated to consume the majority of LLM lifecycle energy and is forecast to dominate accelerator deployments by the late 2020s. Traditional data center workloads — web, storage, virtualization — are comparatively flat and benefit from continued efficiency gains that mostly offset their growth. Geographic concentration matters more than aggregate totals for grid planning. Ireland's Central Statistics Office reported data centers consumed 21% of national metered electricity in 2023, up from 5% in 2015, with the share rising to 22% in 2024; roughly half of metered consumption in the Dublin and Meath region was attributable to data centers. EirGrid and the Commission for Regulation of Utilities imposed a de facto moratorium on new Dublin-area grid connections in 2021, only lifting it in late 2025 under rules requiring on-site generation or storage. In the United States, the Northern Virginia data center alley around Ashburn drew an estimated 32 TWh in 2023 — roughly a quarter of Virginia's electricity — and Dominion Energy's contracted data center capacity reportedly grew from about 21 GW to nearly 40 GW over the second half of 2024, with new connections in parts of Loudoun County delayed pending transmission upgrades. Rack-level power density is the proximate engineering pressure. AI GPU racks routinely exceed 60 kW versus 5 to 10 kW for traditional servers, forcing a shift to liquid cooling and pushing power usage effectiveness closer to its theoretical floor at the best sites while the industry-wide average has been stuck near 1.55 to 1.59 for a decade. See AI Power Consumption: Local vs Cloud and the Scale Problem for the demand-side companion.