Will AI Really Take Down the Power Grid?

There is a story going around that artificial intelligence is about to switch off the lights. Data centers are multiplying, the reasoning goes, each one drinking power like a small city, and sooner or later the grid buckles and the rest of us sit in the dark. It is a tidy fear, and like most tidy fears it contains a real problem wrapped in a wrong conclusion.

Start with the real part, because it is substantial. A Department of Energy study from Lawrence Berkeley National Laboratory projects that US data center electricity use could nearly triple by 2028 and reach as much as 12 percent of the country's total. The International Energy Agency reports that data center demand rose about 17 percent in 2025, with the AI-heavy facilities climbing faster still. A single large AI data center can draw as much power as 100,000 homes. None of that is hype, and none of it is free.

Where the strain actually lands

The question is what that demand breaks. The blackout version imagines the grid simply running out of electricity and failing everywhere at once. That is not how modern grids tend to fail, and it is not what the warning signs are showing.

PJM Interconnection, the largest US grid operator, serving more than 65 million people across 13 states, projects it could fall roughly six gigawatts short of its reliability targets in 2027. That is a serious gap. But the first casualty of a gap like that is usually not your refrigerator. It is the next data center in the queue. The research firm Gartner estimates that power shortages will delay about 40 percent of planned AI data centers by 2027. The constraint bites hardest on the industry causing it, which struggles to get new sites connected rather than knocking existing customers offline.

Grid operators also have tools that a blackout narrative tends to skip. New gas and nuclear capacity is being contracted, often paid for by the hyperscalers themselves. Demand-response agreements let operators throttle data-center load during the worst peaks, something a server farm can tolerate far more gracefully than a hospital can. Several large AI campuses are being built with their own on-site generation so they lean on the public grid less, not more.

The cost that is real, and the one that is overstated

So the honest reading splits in two. The apocalyptic version, in which AI plunges cities into darkness, has little support in how these systems are actually stressed and managed. The quieter problem is genuine and already here: building this much new demand this fast raises wholesale prices, and some of that cost flows through to household bills and to the air near fossil-fueled plants kept running to meet the load. That is a distribution-of-costs story, not an end-of-the-grid story, and it deserves the attention the blackout headline is stealing.

It belongs in the same drawer as the other resource panics around AI, such as the fear that the models are draining the world's water, and the broader worry that the whole build-out is a bubble about to burst. In each case the figures are big enough to take seriously and the doom is bigger than the figures justify. The grid is under real pressure. It is not about to fail, and treating a billing problem as a blackout only makes the actual fix harder to see.