Global data centre electricity demand rose 17% in 2025, while consumption from AI-focused facilities alone jumped 50% in the same year, according to the International Energy Agency. The agency now projects total data centre electricity use to roughly double from 485 terawatt-hours in 2025 to about 950 terawatt-hours by 2030, close to 3% of global electricity demand. In the United States specifically, McKinsey & Company estimates data centre consumption will climb from 147 terawatt-hours in 2023 to 606 terawatt-hours in 2030, taking data centres from 3.7% to nearly 12% of total US power draw. The US Energy Information Administration's 2026 outlook adds a longer horizon, projecting server electricity use in standalone data centres could reach 581 billion kilowatt-hours by 2050 under its high-demand case.
The strain is showing up first in interconnection queues and capacity markets. PJM Interconnection, the grid operator serving 67 million people across 13 states, failed for the first time in its history to secure enough supply in a capacity auction, leaving a shortfall of roughly 6.8 gigawatts for the 2028-29 delivery year. PJM now expects data centres to add more than 30 gigawatts of peak demand by 2030, and its market monitor found that data centres drove 63% of the price increase in the 2025-26 capacity auction, adding an estimated $9.3 billion in costs later billed to ordinary ratepayers.
Capacity prices in most PJM zones jumped from $28.92 per megawatt-day for 2024-25 to $329.17 per megawatt-day for 2026-27, an elevenfold increase in two years. Texas is under similar pressure: ERCOT received applications for 198 gigawatts of large-load interconnection in the first quarter of 2026 alone, a figure close to the grid's entire current peak load.
Governments are intervening directly. In May 2026, the US Department of Energy granted PJM emergency authority to curtail power to data centres running backup generation during grid emergencies, an unprecedented step for facilities long treated as untouchable critical infrastructure. PJM has also approved a $6.7 billion transmission investment plan and is preparing a one-time backstop capacity auction. At the state level, the World Resources Institute notes that five-year utility forecasts for summer peak demand growth more than tripled nationally, rising from 38 gigawatts in 2023 to 128 gigawatts in 2024, a pace utilities were not built to plan around. In Virginia's Loudoun County, home to the densest cluster of data centres in the US, facilities already account for 21% of local power consumption, according to the World Economic Forum, exceeding household use in the county.
Unable to wait for utilities alone, the largest cloud providers are now acting as power developers. Microsoft signed a 20-year, $16 billion power purchase agreement with Constellation Energy to restart the Three Mile Island reactor, renamed the Crane Clean Energy Center, targeting full output by 2027. Google contracted for up to 500 megawatts of small modular reactors from Kairos Power and separately partnered with NextEra Energy to draw power from Iowa's Duane Arnold plant. Meta signed a 20-year, 1.1-gigawatt agreement for the Clinton Clean Energy Center in Illinois, and Amazon committed more than $20 billion to convert its Susquehanna, Pennsylvania campus into a nuclear-adjacent AI hub. Collectively, corporate nuclear power purchase agreements tied to data centres surpassed 16 gigawatts of contracted capacity by the end of 2024, per BloombergNEF figures cited in industry analysis, and capital expenditure across the five largest hyperscalers exceeded $400 billion in 2025, with a further 75% jump forecast for 2026.
Not every trendline points toward runaway demand. The IEA reports that the electricity required per AI task has fallen by at least an order of magnitude annually in recent years, a pace it calls unprecedented in energy history — a routine text query now typically draws less power than running a television for the same stretch of time. Even so, McKinsey projects global data centre capital expenditure will exceed $1.7 trillion by 2030, with AI-driven capacity needs nearly tripling and accounting for roughly 70% of total demand growth. The gap between falling per-task energy costs and rising aggregate consumption is the central tension defining this build-out: efficiency is improving, but total appetite is expanding faster.
The next four years will be decided less by chip supply than by transformers, transmission lines, gas turbine backlogs, and permitting timelines. Lawrence Berkeley National Laboratory estimated grid evaluation bottlenecks had created a potential capacity backlog of 2,600 gigawatts of stalled projects nationally. Whether reforms at PJM, ERCOT, and other grid operators can clear that backlog fast enough — or whether hyperscalers' own nuclear and gas investments arrive first — will largely determine if power supply keeps pace with AI demand through the end of the decade.