The grid operator responsible for keeping the lights on across 13 states and the District of Columbia is staring at a math problem it cannot solve fast enough. PJM Interconnection, which coordinates wholesale electricity delivery for 65 million people from Illinois to Virginia, has warned in its load-forecast updates that demand in its territory is growing so quickly that available generating capacity could fall roughly 6 gigawatts short of what is needed by the 2027-2028 delivery year. The single largest driver of that demand surge: data centers built to run artificial intelligence workloads.
The warning is not hypothetical. PJM’s 2024 Load Forecast Report revised peak-demand projections sharply upward, adding gigawatts of expected consumption that had not appeared in prior planning cycles. Much of that new load traces to large-scale computing facilities clustered in northern Virginia, central Ohio, and northern New Jersey, all inside PJM’s footprint. The forecast revisions landed just months before PJM’s 2025/2026 Base Residual Auction cleared at $269.92 per megawatt-day, nearly ten times the prior year’s price, a market signal that available supply is thinning fast relative to expected demand.
Virginia’s metered reality
Federal data confirm that the demand spike is not a modeling abstraction. The U.S. Energy Information Administration has documented a pronounced rise in commercial electricity sales in Virginia, attributing the increase specifically to data center operations rather than broader commercial or retail growth. Virginia, and Loudoun County in particular, hosts the densest concentration of data center campuses on the planet, according to commercial real estate firms CBRE and JLL that track the sector.
The Dominion Virginia Power zone, which serves the northern Virginia corridor, has recorded some of the steepest load increases in PJM’s entire footprint. EIA figures, compiled from mandatory utility filings that reflect actual metered kilowatt-hours, draw a direct line between server racks in Loudoun County and reserve-margin calculations for a grid stretching from the Appalachian foothills to the Atlantic coast.
The physics of data center demand make the problem especially stubborn. A single large campus can draw 100 to 300 megawatts of constant baseload power. Unlike residential air conditioning, which peaks on summer afternoons and drops overnight, a data center runs at near-full load around the clock, 365 days a year. When dozens of these facilities come online within a few years, the cumulative hit to regional capacity reserves is enormous.
A supply pipeline that cannot keep pace
New power plants are not arriving fast enough to match the load. PJM’s interconnection queue, the formal process through which new generators apply to connect to the grid, has been backlogged for years. As of early 2025, the queue held hundreds of gigawatts of proposed projects, mostly solar, wind, and battery storage, but the average time from application to commercial operation stretched well beyond five years. PJM has undertaken queue-reform efforts in response to FERC Order 2023, but the reforms are still being phased in and have not yet cleared the backlog.
Natural gas plants, which can provide the firm, dispatchable power that complements intermittent renewables, face their own permitting and financing hurdles. Meanwhile, retiring coal and older gas units are removing capacity from the system. PJM’s own Energy Transition in PJM report flagged the risk that retirements could outpace additions, a concern that has only intensified as data center load accelerates the timeline.
The result is a narrowing reserve margin. PJM targets a reserve margin above roughly 14.7 percent to maintain reliability. Internal planning scenarios show that margin compressing toward or below that threshold in the late 2020s if demand growth tracks the upper range of current forecasts and new supply does not accelerate.
What AI changes about the calculus
Traditional cloud computing and enterprise hosting have driven data center growth for more than a decade. AI workloads add a new variable. Training a large language model can require tens of thousands of GPUs running at full tilt for weeks or months, consuming electricity at rates that dwarf conventional server operations. Inference, the phase where a trained model responds to user queries, draws less power per request but scales with adoption. As AI tools spread into search engines, business software, and consumer apps, inference demand is climbing rapidly.
No public dataset yet separates AI-specific electricity consumption from general data center load at the utility-zone level. That gap makes it difficult to pinpoint exactly how many of the megawatts flowing into Virginia’s data centers are powering AI versus traditional workloads. What is clear from PJM’s interconnection requests is that the newest and largest facilities are being designed with AI-scale power densities in mind, often requesting grid connections of 200 megawatts or more per campus.
The geographic concentration matters, too. AI training tends to cluster where fiber connectivity, land, and power are all available at scale. Northern Virginia checks every box, which is why it keeps attracting new campuses even as the local grid strains. Whether future AI buildouts diversify to less congested regions, or continue piling into the same corridors, will shape where supply gaps hit hardest.
What it means for electricity bills and reliability
The tightening supply picture has direct consequences for the 65 million people who depend on PJM’s grid. When capacity-auction prices spike, as they did in the 2025/2026 auction, those costs flow through to retail electricity rates. State consumer advocates in Pennsylvania, New Jersey, and Maryland have already flagged the potential for higher bills tied to rising capacity charges. Dominion Energy has cited data center-driven load growth in Virginia regulatory filings as justification for new transmission and generation investment, costs that are typically shared across the utility’s customer base.
Reliability is the deeper concern. A reserve margin that dips below PJM’s target does not guarantee blackouts, but it shrinks the buffer available during extreme weather events, unexpected plant outages, or fuel-supply disruptions. The combination of a thinner margin and a load profile dominated by facilities that cannot simply turn off during emergencies makes the system less forgiving of surprises.
State regulators are beginning to respond. Virginia lawmakers debated reforms to the state’s data center tax incentive program during the 2025 legislative session, weighing the economic benefits of the industry against infrastructure strain. PJM launched its Critical Issue Fast Path (CIFP) process to expedite rule changes addressing resource adequacy. And FERC has opened proceedings on interconnection-queue reform and transmission planning that could, over time, accelerate the supply pipeline.
A grid region at an inflection point
The documented surge in data center-driven electricity consumption, the backlog of new generation projects, and the sharpest capacity-price increase in PJM’s history all point in the same direction. The eastern U.S. power system is entering a period of tighter margins, not yet a crisis, but a structural shift that demands faster action from utilities, regulators, and the technology companies whose facilities are reshaping the load curve.
Whether the projected 6-gigawatt gap materializes in full depends on variables that no single forecast can pin down: how many announced data centers actually reach full power draw on schedule, how quickly new plants clear the interconnection queue, and whether AI demand growth follows the steep trajectory that current investment trends suggest. What the federal consumption data already show is that the pressure is real, it is concentrated, and it is growing. The decisions made in utility boardrooms, state capitols, and data center design meetings over the next 12 to 18 months will determine whether the grid bends or breaks under the weight of the AI boom.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
