Calvin Butler runs the largest electric utility in the United States, and he is not softening the message. In a recent interview with the Financial Times, the Exelon chief executive said parts of the country could face blackouts as soon as 2027, with the danger concentrated in the Northeast and Midwest where much of his company operates. It was not a distant hypothetical. This past winter, Butler said, Exelon came within a whisker of cutting power to roughly 400,000 customers during the coldest stretch of the year.
What makes the warning unusual is the cause. The grid is not buckling under a single hurricane or a once-in-a-decade freeze. It is straining against demand that, after nearly two flat decades, has started climbing faster than utilities can build. Artificial intelligence is the accelerant. The server farms that train and run large AI models draw electricity on a scale the American grid was never designed to absorb this quickly, and they are landing in tight clusters — dozens of facilities in a single county — long before new transmission lines and power plants can be permitted and switched on.
Exelon sits in the middle of the region absorbing much of that load, serving customers through ComEd in Illinois, PECO in Pennsylvania, Baltimore Gas and Electric in Maryland and sister utilities across the Mid-Atlantic. Butler's argument is simple and uncomfortable: the arithmetic of supply and demand no longer balances on the worst days of the year, and the cushion is thinning fast.
The near-miss that rattled the grid's biggest operator
Curtailment is the industry's term for deliberately cutting power to some customers to stop the wider system from collapsing. It is the outcome utilities exist to prevent, which is why Butler's account of last winter carries weight.
"We came very close, this past winter, to having to curtail power for about 400,000 customers on some of the coldest days of the year. And it's only getting worse." — Calvin Butler, CEO, Exelon
At Fortune's Brainstorm AI conference in San Francisco in December, Butler reached for a mechanic's metaphor. "We are telling policy makers the warning lights are on," he said, cautioning that the country would "keep pushing this and no one's going to pay attention until it breaks down." He has spent about 25 years in the industry and said the past four decades never produced "this amount of load growth." Utilities, he added, are meant to be the quiet foundation, not the risk-takers: "We're 5% of the economy, but we power the next 95%." Coming that close to curtailment in the winter of 2025-26 — before the heaviest AI build-out has even arrived — is why executives are now naming 2027 rather than 2035.
Why demand is climbing after two flat decades
For most of this century, U.S. electricity demand barely moved. Efficiency gains from LED lighting and better appliances roughly offset economic growth. That era has ended. The U.S. Energy Information Administration, in a January 2026 forecast, projected national electricity consumption would rise about 1% in 2026 and 3% in 2027 — the strongest four-year expansion since 2000 and the first run of four straight annual increases since 2007.
The agency named the driver without hedging. Electricity demand is rising through 2027 "driven largely by increasing demand from large computing facilities," said EIA Administrator Tristan Abbey. Data centers are the headline, but they stack on top of factory reshoring, EV charging and the electrification of home heating and transport — every trend pulling in the same direction at once.
The data-center figures are the ones that stop you cold. These facilities consumed roughly 1.9% of U.S. electricity in 2018; by 2023 the share had reached about 4.4%. The Electric Power Research Institute has put the 2030 range anywhere from 4.6% to 9.1% of national consumption, while Lawrence Berkeley National Laboratory estimates data centers could account for as much as 12% of U.S. peak demand by 2028.
The AI build-out, by the numbers
The footprint of a single site explains the strain. A conventional server rack might draw 5 to 15 kilowatts; a rack packed with AI accelerators can pull 30 to 100 kilowatts or more. Scale that across a campus and one hyperscale facility can rival the load of a small city. BloombergNEF projects U.S. data-center power demand will reach about 106 gigawatts by 2035, up from roughly 25 GW in 2024 — an increase of more than 300%, and 36% higher than the same analysts forecast in early 2024.
Concentration compounds the problem. When many facilities share one corner of the grid, a single fault can ripple outward. In a 2024 disturbance in Northern Virginia — the densest data-center market on earth — a voltage dip knocked roughly 60 data centers offline at once, leaving the grid with about 1,500 megawatts of sudden surplus and a scramble to rebalance. Where demand clusters, wholesale prices spike; industry reporting has pointed to increases of more than 260% near major data-center hubs.
| Metric | Figure | Source |
|---|---|---|
| U.S. electricity demand growth, 2027 | +3% (strongest 4-year run since 2000) | EIA |
| U.S. data-center power demand, 2035 | ~106 GW (from ~25 GW in 2024) | BloombergNEF |
| PJM peak-demand growth, 2024-2030 | +32 GW (all but 2 GW from data centers) | PJM |
| PJM 2027-28 capacity shortfall | ~6,600 MW below reliability target | PJM |
| Record PJM capacity price | $333.44/MW-day (3rd straight record) | PJM auction |
| Data-center share of Dec. auction cost | $6.5B (40% of $16.4B) | Monitoring Analytics |
The forecasts vary in detail, yet they all point the same way: a step-change in load arriving over a handful of years, into a system where a single large transmission line can take most of a decade to permit and build.
PJM's alarm bell: a record auction and a thinning cushion
The sharpest signal came from PJM Interconnection, the grid operator for 13 states and Washington, D.C., including most of Exelon's territory. In its December 2025 capacity auction — which secures supply for the delivery year that begins June 1, 2027 — prices slammed into a $333.44 per megawatt-day cap, a record for the third auction running. Without a temporary price-cap deal in Pennsylvania, PJM said, the clearing price would have hit nearly $530, roughly 60% higher. The total bill reached $16.4 billion.
More telling than the price was the shortfall. PJM came up about 6,600 megawatts short of the reserve it targets to meet a one-event-in-ten-years reliability standard. Only around 774 MW of new generation cleared the auction — evidence that plants are not being built fast enough to keep pace. The operator now expects peak demand across its footprint to grow 32 gigawatts between 2024 and 2030, with all but 2 GW of that coming from data centers.
PJM's independent market monitor, Monitoring Analytics, was blunt about the cause: "Data center load growth is the primary reason for recent and expected capacity market conditions." Data-center demand accounted for about $6.5 billion — 40% — of the December auction's costs.
What it means for your electric bill
Higher capacity costs do not stay in a spreadsheet; they flow to households. Butler has said flatly that "the prices are going to go up," and they already have. Reporting on his remarks pointed to electricity price increases of around 7% nationally over the past year, with sharper jumps across several Exelon markets — roughly 17% in New Jersey, 16% in Maryland and 13% in Pennsylvania.
The temporary price caps that shielded PJM customers — trimming costs by an estimated $13.1 billion across two auctions — are expiring, which means more of the true cost of tight supply is about to reach bills. That has turned an engineering problem into a political one, with legislatures in Maryland and neighboring states pushing to add generation and soften rate increases before voters feel the full impact.
Can supply catch up in time?
The industry's answer is to build, aggressively. Utilities have outlined roughly $1.1 trillion in investment over five years, and Exelon itself is advancing projects such as a 765-kilovolt transmission line running about 220 miles across Pennsylvania and West Virginia. The obstacle is time. Interconnection queues — the waiting line new projects join to plug into the grid — can stretch seven years, and orders for the large gas turbines needed to add firm capacity are backlogged well into the next decade.
That leaves today's grid leaning on yesterday's plants. In PJM's latest auction, natural gas supplied 43% of cleared capacity, nuclear 21% and coal 20%, while wind and solar together contributed only low single digits. Regulators see the same gauges flashing. The North American Electric Reliability Corporation reported that peak winter demand across its regions jumped about 20 gigawatts — 2.5% — in a single year, and flagged elevated risk of shortfalls under extreme conditions in New England, the Southeast, Texas and parts of the West, along with tighter summer margins across all three U.S. interconnections.
Butler's 2027 date is a warning shot, not a settled forecast, and the American grid has absorbed shocks before. But the gap he keeps describing — between how fast AI is arriving and how slowly power plants and high-voltage lines get built — is real, measurable and, for now, still widening. The next two winters will reveal whether the warning lights he points to persuade anyone to act before something actually breaks.