On a summer afternoon when air conditioners are running full blast across the mid-Atlantic, the electrical grid serving 65 million people from New Jersey to Illinois will need every megawatt it can muster. PJM Interconnection, the organization that coordinates that grid, is warning that it may not have enough. The reason: artificial intelligence data centers are devouring electricity at a pace that has outrun every recent forecast, and new power plants are not being built fast enough to keep up.
PJM’s load forecast revisions and reserve margin analyses point to a capacity shortfall on the order of 6 gigawatts by the late 2020s, enough electricity to power roughly 4.5 million homes. The gap has already shown up in market signals. PJM’s most recent capacity auction, covering the 2025/2026 delivery year, cleared at $269.92 per megawatt-day, nearly tenfold the prior auction’s price and a record for the grid operator. That price spike reflects a market screaming for new supply that does not yet exist.
Federal regulators, grid analysts, and the tech companies driving the demand surge are all scrambling to figure out what comes next.
Where the demand is coming from
Northern Virginia already hosts the densest cluster of data centers on the planet. Dominion Energy, the utility serving much of that territory, has publicly reported that data center interconnection requests in its queue exceed available transmission capacity. New facilities are under construction or permitted across Maryland, Ohio, and Pennsylvania, with Microsoft, Amazon, and Meta among the companies expanding aggressively in PJM’s footprint.
The Lawrence Berkeley National Laboratory published a 2024 report modeling scenarios in which data centers account for roughly 6.7% to 12% of total U.S. electricity consumption by 2028. Even the low end of that range represents a dramatic jump from the roughly 4% share data centers held just a few years ago. The Department of Energy released a companion evaluation tying the surge directly to AI and machine-learning workloads, calling it a significant increase in electricity demand from data centers that grid planners had not anticipated.
What makes AI workloads different from traditional cloud computing is their intensity. Training a single large language model can consume as much electricity as thousands of households use in a year, and the inference phase, when millions of users query the model simultaneously, creates sustained around-the-clock load. Unlike a streaming video server that idles at 3 a.m., an AI cluster runs hot 24 hours a day.
Why supply is not keeping up
Building a power plant takes years. Permitting a natural gas combined-cycle facility in PJM territory typically requires three to five years from proposal to commercial operation. Utility-scale solar and wind projects face their own permitting and interconnection delays, and PJM’s interconnection queue has been backlogged, with hundreds of projects waiting years for grid studies to be completed.
At the same time, older generators are retiring. Coal plants across Ohio, Pennsylvania, and West Virginia have been shutting down as they become uneconomic against cheaper gas and renewables. The U.S. Energy Information Administration published analysis showing that faster-than-expected data center demand could actually cause fossil generation to rise in markets including the mid-Atlantic, a finding that cuts against the broader retirement trend. Some aging plants may need to keep running simply to prevent blackouts.
Battery storage, often cited as a bridge technology, remains limited in scale within PJM’s territory. As of early 2026, the installed battery capacity across PJM’s 13-state footprint represents a small fraction of the projected shortfall. Batteries can help manage peak demand for a few hours, but they cannot replace the sustained baseload generation that data centers require around the clock.
Federal regulators are paying attention
The Federal Energy Regulatory Commission convened a Commissioner-led Technical Conference focused on resource adequacy challenges across regional grid operators. PJM’s anticipated capacity shortfalls were among the central issues reviewed, placing the grid operator’s warnings on the official federal record.
That conference marked a shift. When FERC commissioners publicly question whether a grid operator can keep the lights on, the concern has moved beyond industry chatter into formal oversight. Commissioners heard testimony about the mismatch between load growth projections and the pace of new generation development, and several pressed for faster interconnection processes and clearer demand forecasting from utilities.
State regulators in PJM’s territory are also responding. Virginia’s State Corporation Commission has been reviewing Dominion Energy’s integrated resource plans with heightened scrutiny, and Maryland and Pennsylvania regulators have opened proceedings examining how data center load growth affects residential ratepayers. The core worry at the state level: if utilities must build or retain expensive generation to serve data centers, who pays for it?
What this means for electricity bills and reliability
The record-high capacity auction prices are not abstract. Utilities pass capacity costs through to customers, and the 2025/2026 auction results are expected to translate into noticeable increases on residential and commercial electricity bills across PJM’s territory starting in mid-2026. Estimates vary by utility and state, but ratepayer advocates in several PJM states have flagged potential annual increases in the range of $100 to $300 per household.
Reliability is the deeper concern. PJM’s reserve margins, the buffer between available generation and peak demand, have been thinning. If a severe heat wave coincides with planned generator outages and the continued growth of data center load, the grid operator could be forced to call for emergency demand reductions or, in a worst case, implement rolling blackouts. PJM has not had to order widespread load shedding in decades, but the shrinking margin leaves less room for error.
For context, 6 gigawatts is roughly the output of six large nuclear reactors or about 12 natural gas combined-cycle plants. Closing a gap that size in two to three years would require an unprecedented pace of construction, permitting, and grid interconnection.
What to watch through the rest of 2026
Several concrete milestones will signal whether the supply gap is narrowing or widening. PJM’s next capacity auction results, expected later in 2026, will show whether high prices are attracting enough new generation commitments. State-level decisions on power plant siting and transmission line approvals in Virginia, Ohio, and Pennsylvania will determine how quickly new supply can physically come online. And FERC’s ongoing review of interconnection queue reforms could accelerate or further delay the hundreds of generation projects waiting to connect to PJM’s grid.
Tech companies are also making moves that bear watching. Some, including Microsoft and Amazon, have signed or explored direct power purchase agreements with nuclear plant operators, effectively reserving existing generation for their own use. Whether those deals reduce the supply available to everyone else, or whether they fund life extensions for plants that might otherwise retire, is a question regulators have not fully answered.
The 65 million people who depend on PJM’s grid did not choose to live at the epicenter of the AI power boom. But the decisions being made right now by grid operators, regulators, utilities, and tech companies will determine whether the lights stay on and what the bill looks like when they do.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
