Background data — the American grid
Everyone is digging for a villain, and this year it's the data center. But the record isn't hidden — it's buried, and we did the burying. Take the core sample down. See what fifty years of looking away put in the ground.
AI is eating the grid — that's the headline. It's tidy, and it's lazy — and, frankly, it's dangerous.
Because AI isn't new, and neither are the data centers behind it. Machine learning has quietly run the digital world for years: our spam filters, our banks' fraud detection, the search box, the logistics that put food on the shelf, even the software utilities use to balance the grid itself. Every time we send an email, stream a show, or save a photo to the cloud, we're leaning on a data center.1 The compute we're now blaming is the compute modern life already runs on. It didn't ambush us. We invited it in years ago, and we would never give it back.
Data centers and AI didn't cause America's energy crisis. They're the stress test that finally exposed it — and we've been failing that test for half a century. Everything below this line was already in the ground.
For nearly two decades, U.S. electricity consumption was essentially flat.2 Utilities got comfortable. Then load came roaring back — data centers, reshored factories, electrification — and Grid Strategies now projects peak demand growing 166 GW by 2030, more than half of it from data centers.3 FERC pegged the gap at 50–130 GW beyond what anyone had planned for.4
None of it snuck up on us. Utilities are not blind. Large customers file interconnection requests and sign power contracts years in advance; every utility runs a resource plan and watches the pipeline fill. Data center load has tripled over the last decade.5 The utilities' own five-year demand forecasts leapt from 38 GW to 128 GW in a single year6 — the signal was sitting on their desks the whole time. It was a forecast, and we filed it away.
Look at the wires. High-voltage transmission construction has cratered — from 1,700 miles a year in the early 2010s, to 350 miles between 2020 and 2023, to a bare 55 miles in 2023.7 We build a fifth of what we did a decade ago, even as spending hit an all-time high. And the bones are ancient. Much of the grid was strung in the 1960s and 1970s, and a 2015 Department of Energy review found roughly 70% of transmission lines were already past 25 years old.8
More than 2,000 GW of generation and storage sit stranded in interconnection queues — nearly twice the entire existing U.S. fleet, most of it clean.9 Only about one in five projects ever gets built, and the ones that do wait more than four years.9 The supply is there, and so is the demand. We let the machine that connects them rust.
Here the drill turns green — and here politics ran the show instead of engineering. It did it two ways.
Indian Point — a nuclear plant on the Hudson, thirty-odd miles north of New York City — did not just fail. A governor, a state attorney general, and an advocacy group campaigned it shut, and natural gas filled the hole — roughly 2,000 MW of carbon-free baseload, retired on principle.10
San Onofre — a twin-reactor station on the Southern California coast between Los Angeles and San Diego — is, unfortunately, even worse. Its replacement steam generators failed, and rather than do the hard, expensive work of fixing and restarting a plant we already had, we let it go: decapitated by the battle to restart it, the cost, and the ease of a quick fix in cheap gas. It didn't have to close. We decided it wasn't worth saving. Complex problems demand complex solutions, and we took the easy exit — 2,150 MW of clean, firm power, walked away from.11
One was buried on principle; the other because fixing it was too much trouble. 4,150 MW of firm, carbon-free baseload at these two plants alone — gone by choice either way. Vermont Yankee and the near-death of Diablo Canyon are versions of the same story: not that any of it was "dirty" energy per se, but that it was not the right kind of clean, or the right kind of convenient.
Meanwhile the transmission that could carry new power dies in NIMBY fights and state-versus-state turf wars,12 and the permitting reform that might let us build has been a partisan nuclear football for years — neither side willing to do what is good for our children's futures. Energy policy became a proxy war, fought for this cycle's applause, with no one minding the next decade's demand.
The one large firm plant we did manage to finish, Vogtle, came in seven years late at $35 billion — 2x its estimate.13 We struggle to build new solutions, and we fold when it comes to fixing the old — choosing not to spend the roughly $680 million a repair of San Onofre's steam generators would have run, against the $14 billion Vogtle was originally supposed to cost, and letting 2,150 MW of clean power go instead.
So no — AI didn't do this. We chose it. Fifty years of vetoing what wasn't pure enough while permitting nothing in its place.
The data center villain on the evening news is the same one running our hospitals, our paychecks, our feeds. Go read the interconnection reports, the transmission numbers, the shutdown decisions with names attached. The record's been there the whole time. The monster was more appetizing.
And that isn't even the hardest part to sit with. The hardest part is watching us line up to do it all over again.
Think of the last time AI came up in casual conversation. I bet it took seconds to cascade into some version of concern for the energy demand. Positives are often tossed in as well — but caveated, of course, with dismay over this drastic situation. It is a reflex, and a reflex is just that: sensation, response. We were handed a half-truth and swallowed it whole, and now that half-truth has been fed back into the same machine that hollowed out the grid to begin with: another wedge for the next campaign.
So where is the accountability? Not the kind that trends for a weekend and evaporates — the kind that asks who decided and what it cost the rest of us. It doesn't come. Good government — the ideal that country and community come before self-gain, that we collectively build toward a brighter future even if we do not get credit for it — was supposed to be the point. It gets tossed aside almost the moment power is tasted.
And the people whose job is to hold the line — the press, the feeds, the ones with the reach — are not asking the right questions. They ask for a villain. A villain is far more beneficial in the short term for all interested parties — understanding that being a decent human is clearly not an interested party. It's easier, and the clicks travel faster.
The answer is what it always is.
Every reading above is sourced. Dig for yourself.