The United States is entering a new era of electricity demand, and the spark is artificial intelligence. After two decades of relatively flat consumption, the grid is now straining to keep up as AI data centers proliferate, chip densities soar, and workloads shift from conventional cloud to accelerated computing. In January 2026, federal forecasters said outright that U.S. electricity use would not just keep rising, it would set new records in 2026 and again in 2027.
That inflection is not just about extreme weather or electrification. It’s increasingly about where and how fast AI capacity is being built. Clusters of hyperscale campuses are concentrating gigawatts of new load in a handful of places, forcing grid operators, utilities, regulators, and tech giants to rethink timelines, rate design, and even their fuel mix. The result: record demand, contentious politics, and a scramble for firm, clean power.
The new records, and why AI is the tipping point
U.S. electricity consumption hit a record in 2025 and is forecast to climb to 4,256 billion kWh in 2026 and 4,364 billion kWh in 2027, the strongest four-year growth since 2000. Federal analysts explicitly link that rise to large computing facilities serving AI and other digital loads, alongside broader electrification.
Globally, the International Energy Agency estimates data centers’ electricity use could more than double by 2030 to roughly 945 TWh, with accelerated servers for AI driving about half the net increase. In the United States specifically, the IEA notes that data centers are on track to account for almost half of demand growth through 2030, illustrating why the U.S. grid is feeling the pinch first.
New records are not just annual totals, they also show up in regional peaks and local bottlenecks. Where AI clusters form, demand growth that used to take a decade can materialize in just a few years, outpacing how quickly power plants, transmission lines, and substations can be built.
Where the surge is hitting hardest: PJM and ERCOT
Two regions stand out. In PJM, the nation’s largest grid operator across the Mid-Atlantic and parts of the Midwest, peak demand is now projected to grow around 2% per year this decade, a sharp turn from prior flat forecasts, with planners pointing to explosive data center development as a key driver. PJM’s latest long‑term forecast shows summer peaks rising from roughly 151 GW in 2024 toward 179 GW by 2034.
Texas’ ERCOT is experiencing an even steeper wave. ERCOT officials disclosed that large‑load interconnection requests surged to more than 230 GW in 2025, nearly quadruple the end‑2024 level, with over 70% coming from data centers. Many proposed sites request more than 1 GW each, compressing timelines and stressing planning processes. ERCOT has reorganized internally to cope with rapid growth and to modernize interconnection.
Federal energy forecasters reflect that divergence: EIA sees electricity demand in ERCOT growing about 5% in 2025 and 9.6% in 2026, and in PJM about 3.3% in both years, rates not seen in decades. That concentrated growth amplifies the challenge of supplying firm capacity and moving power to AI clusters fast enough to maintain reliability.
Virginia and Georgia: case studies in AI-era planning
Northern Virginia’s “Data Center Alley” remains the epicenter. Dominion Energy reports its contracted data‑center capacity nearly doubled in late 2024 to about 40 GW, prompting an increase in its five‑year capital plan to more than $50 billion for 2025, 2029. The utility has also floated new large‑load rates as it races to expand wires and firm supply.
That growth cascades into transmission. A supplemental filing tied to Dominion’s long‑range plan shows winter peak demand in its territory could climb from about 17.4 GW in 2024 to 26.6 GW by 2039 with data center build‑out, requiring new high‑voltage lines to funnel power into load pockets. These local upgrades are now central flashpoints in siting debates.
Further south, Georgia’s regulators approved a 2025 Integrated Resource Plan that anticipates 8,500 MW of potential load growth within six years, preserves some existing thermal capacity for reliability, and adds up to 4,000 MW of new renewables and 1,500+ MW of storage by 2035, explicitly citing large loads like data centers. The plan also accelerates transmission investment and requires regular reporting on large‑load developments.
How hyperscalers are scrambling for firm, clean power
Corporate buyers are moving beyond traditional wind‑and‑solar PPAs to secure round‑the‑clock, carbon‑free electricity. In June 2025, Meta signed a 20‑year deal for the entire 1.121‑GW output of Constellation’s Clinton nuclear plant in Illinois starting in 2027, supporting the plant’s relicensing and a small uprate.
In January 2026, Meta went further, announcing agreements that could unlock up to 6.6 GW of nuclear capacity by the mid‑2030s via partnerships with TerraPower, Oklo, and Vistra, part of a broader strategy to anchor firm clean power near AI campuses. Such deals signal a shift toward baseload contracting to match 24/7 demand profiles.
Others are pursuing similar paths. Constellation says a 20‑year agreement with Microsoft underpins the restart of the 835‑MW Crane Clean Energy Center in Pennsylvania to serve PJM‑region data centers, while developers like Switch are inking direct contracts for geothermal supply. These moves complement surging grid‑scale storage and efficiency upgrades inside facilities.
Prices, politics, and community pushback
As utilities juggle unprecedented load growth, local communities are scrutinizing who pays for new substations, feeders, and peaking plants. Microsoft responded on January 13, 2026 with a “Community‑First AI Infrastructure” pledge, promising to avoid shifting grid costs to households by paying higher rates, cut water use, and be more transparent about sites and energy needs. The company paired the pledge with commitments on jobs and local taxes.
Still, backlash has canceled or delayed multiple projects as residents question secrecy in utility deals, potential bill impacts, and land‑use tradeoffs. Policymakers are weighing new rules on large‑load interconnections and rate design to protect existing customers while welcoming investment. In fast‑growing hubs, the politics of data centers now shape local elections and statehouse agendas.
Grid operators warn that without timely generation and transmission, reliability margins could erode during extreme weather. Some regions are exploring interim measures, such as requiring large loads to provide on‑site backup or curtail during peaks, while stakeholders negotiate longer‑term solutions.
The water and emissions footprint of AI facilities
AI‑optimized campuses pack extraordinary compute into tight footprints, pushing up both electricity and cooling needs. The latest LBNL/DOE assessment estimates U.S. data centers used about 4.4% of national electricity in 2023 and could reach 6.7%, 12% by 2028, depending on deployment and efficiency. That rapid climb concentrates environmental impacts in a few hot spots.
Water is a growing concern. LBNL’s 2024 report and subsequent summaries estimate U.S. data centers directly consumed about 17 billion gallons of water for cooling in 2023, with indirect consumption via power generation far higher; both could multiply by 2028 if facilities and supply remain fossil‑heavy. These figures are driving site shifts toward water‑abundant regions and efficiency upgrades like liquid cooling and heat reuse.
Emissions trends depend on what fills the supply gap. Preliminary 2025 estimates from Rhodium Group show U.S. power‑sector emissions rising on higher coal generation amid surging load and pricier natural gas, underscoring the stakes if record AI demand is met with fossil capacity rather than firm clean power.
Can innovation blunt the surge?
Efficiency still matters. The IEA projects electricity use by accelerated AI servers to grow roughly 30% per year this decade; improving server utilization, model efficiency, and facility PUE remains essential to reduce grid stress per unit of compute. Meanwhile, grid‑side upgrades, advanced conductors, dynamic line ratings, and faster interconnection studies, can unlock capacity sooner.
DOE’s playbook for large loads emphasizes on‑site generation and storage so data centers can be grid assets, repurposing retired coal sites with existing high‑voltage ties, and innovative rates that align consumption with system needs. That approach, paired with 24/7 clean energy procurement, can limit impacts on other customers while accelerating decarbonization.
Even with innovation, timing is critical. The record‑setting demand now forecast through 2027 will test whether utilities, regulators, and hyperscalers can add firm, clean supply and transmission as fast as AI capacity is deployed. The answer will determine whether AI’s power footprint becomes a catalyst for grid modernization, or a new source of volatility.
AI’s rise doesn’t have to mean higher emissions and higher bills. The path is emerging: procure firm low‑carbon power at scale, build wires faster, push efficiency across chips and models, and share costs transparently with host communities. Get those pieces right, and the U.S. can meet record demand while hardening reliability and accelerating the clean‑energy transition. Fall short, and record highs in demand may be followed by record‑high risks.
