AI Data Centers and the U.S. Grid: What the Electricity Numbers Actually Show, and Who Pays

A Watchdog and a Fact-Checker Say Different Things

PJM, the grid operator that keeps the lights on for 65 million people from Chicago to Washington, has an independent market monitor whose job is to call things straight. In 2025 it said data centers were the "primary reason" the grid's capacity auction hit record prices three years running ^[5][6][26]. Around the same time, PolitiFact rated a viral claim about data centers and electric bills "Mostly False" ^[9].

Both statements are backed by real numbers. Both come from careful sources. And they seem to say opposite things.

The gap is not that one side is lying. It's that "capacity price" and "your electric bill" are not the same number, even though both get called "what data centers cost you." Untangling that gap is most of this story.

The Tripling Nobody Disputes

Start with what almost nobody argues about. In 2023, data centers in the United States used about 176 terawatt-hours of electricity, roughly 4.4% of everything the country consumed ^[1]. A terawatt-hour is a trillion watt-hours — enough to power a few hundred thousand homes for a year. A decade earlier, in 2014, data centers used about 58 terawatt-hours ^[1]. That's a tripling in ten years.

The federal Energy Information Administration puts 2024 usage even higher, near 183 terawatt-hours ^[2][4]. This growth is landing on a grid that spent roughly 20 years barely growing at all. U.S. electricity demand was close to flat from the early 2000s through the early 2020s ^[2][4].

Where this goes next is the first real fork in the story. Lawrence Berkeley National Laboratory, the Department of Energy's own research arm, projects data centers could reach anywhere from 6.7% to 12% of all U.S. electricity by 2028 ^[1]. That's not a typo — it's one lab's own range, and it's wide enough that the low end and high end would lead to very different policy responses. The International Energy Agency separately projects U.S. data-center demand growing 133% between 2024 and 2030 ^[3].

None of this is spread evenly across the country. It clusters hard in a handful of places: Northern Virginia, the PJM grid across the mid-Atlantic, Texas's ERCOT grid, Georgia, Ohio, and Arizona ^[1][4]. PJM alone expects peak demand to grow by 32 gigawatts from 2024 to 2030, and all but 2 of those gigawatts are expected to come from data centers ^[4][5]. A national average completely hides this. If you don't live near one of these grids, the numbers above may barely touch your bill at all. If you do, they might already be reshaping it.

What a Capacity Auction Actually Buys

Here's the mechanism the PJM monitor's "primary reason" comment was about, and it's worth explaining because the term "capacity auction" does a lot of work in this debate. PJM doesn't just buy electricity — it also pays power plants, years in advance, simply to promise they'll be available when demand spikes. That advance payment is the "capacity price," and it gets set at an auction.

For three straight years, that auction cleared at record levels ^[5]. PJM's market monitor pinned about 40% of the most recent auction's $16.4 billion cost, roughly $6.5 billion, on data-center demand — and most of that demand is for facilities that haven't even been built yet ^[26]. In an earlier accounting, the monitor estimated data centers added about $9.3 billion, a 174% jump, to capacity costs for the 2025-26 delivery year compared with what costs would have been without them ^[6].

That capacity cost gets spread across all of PJM's customers, which is the basis for the "primary reason" claim. But it is not the same thing as your monthly electric bill. Supply and capacity costs typically make up only 30% to 50% of what you actually pay; the rest is delivery, maintenance, and other charges ^[9]. That distinction is exactly why PolitiFact could call the 267% figure "Mostly False" without disputing that PJM's own numbers are real: the 267% referred to wholesale prices at specific grid locations, not what shows up on a household's bill ^[8][9].

There's a second wrinkle in the demand side of this story: not all of the demand behind those numbers is real yet. Developers sometimes file the same data-center project's request to connect to the grid with several utilities at once, so the same project gets counted multiple times. Analysts call this "phantom load," and it means the pipeline of requested demand can run 5 to 10 times larger than what actually gets built ^[12]. The utility Exelon estimated only about 22% of its 65-gigawatt pipeline through 2040 is likely to materialize ^[12]. So even the size of the underlying problem carries a real asterisk.

Four Honest Ways to Read the Same Bill

Strip away the noise and the debate over "who pays" comes down to four distinct, defensible arguments, each leaning on real evidence and each with a real blind spot.

The first says data centers are shifting costs onto ordinary households. This is the view of ratepayer advocates, Senator Elizabeth Warren's Senate investigation, and outlets like Fortune and Bloomberg ^[6][7][8][9]. Their case: new grid infrastructure built for data centers gets billed to everyone under standard cost-sharing rules, and the grid operator's own monitor, not an advocacy group, is the one calling data centers the "primary reason" for price records ^[6]. A Bloomberg analysis found that 73% of grid locations where wholesale prices rose sit within 50 miles of major data-center activity ^[8]. The critique of this view: it can blend wholesale and capacity-market costs, which are a fraction of a bill, with the full retail bill, and it can undercount how much of the recent bill increase is really about gas prices, storms, and aging infrastructure ^[9][23].

The second says data centers are one driver among several, not the driver. PolitiFact, a Rutgers University policy lab, and Berkeley Lab's own report take this line ^[1][9][23]. Their case: electricity bills have been climbing for reasons that predate the AI boom entirely — natural gas prices, extreme weather, wildfire and storm hardening, and simply replacing an aging grid ^[1][9][23]. The Rutgers lab's blunt conclusion was that data centers are "mostly not yet" the cause for typical households ^[23]. The critique here: this view can understate how concentrated and immediate the effect already is in PJM specifically, where the market monitor has directly quantified it — and "not yet" could turn into "yes" quickly as more planned data centers come online ^[6].

The third says this is a supply opportunity: build more power and make big customers pay their own way. This is the view of energy-abundance conservatives, utilities, and nuclear advocates ^[14][15][16]. Their case: demand growth after 20 flat years is a chance to build gas, nuclear, and small modular reactors, and new "large-load tariffs" — special rate structures that require big customers to cover their own capacity costs — can even lower bills for everyone else by spreading fixed grid costs across more customers ^[14][15]. Ohio and Georgia both approved large-load tariffs in 2025 that require data centers to pay for the large majority of their contracted capacity, sometimes over contracts running as long as 12 years ^[14][15]. The critique: new power plants take years to build while price spikes are happening now, and co-located generation has already hit a wall — federal regulators rejected the high-profile Amazon-Talen deal to power a data center directly from a nuclear plant ^[16].

The fourth says the whole forecast is inflated. Grid-forecasting analysts point to that phantom-load problem described above and argue utilities have a long history of overestimating future demand ^[12][22]. Their case: if planners build generation and transmission for demand that never shows up, ratepayers could be left paying for stranded infrastructure, when instead, flexibility could handle most of the real load. Duke University's Nicholas Institute found the existing grid could absorb roughly 100 gigawatts of new demand if data centers accepted being briefly powered down during less than 1% of hours each year ^[13]. The critique: even after subtracting the phantom requests, what's left is still historically large and is already moving auction prices, and if flexibility commitments aren't legally binding, this argument could become an excuse to under-build and then face real reliability problems ^[5][20].

A Grid Built for a World That No Longer Exists

Underneath all four arguments sit the same structural facts, and they explain why this debate is so hard to settle cleanly. The most basic one: American electricity demand sat nearly flat for about two decades, so utilities and regulators built almost no muscle for handling fast growth ^[2][4]. Data centers, factory reshoring, and the shift to electric vehicles and heating are all arriving into that same unprepared system at once.

Whether any single data center ends up raising your bill or lowering it is decided in dry, obscure rate cases and tariff filings, not in headlines ^[14][15]. The exact same physical building can be a bill-raiser in one state's rules and a bill-lowerer in another's.

There's also a basic speed mismatch. A data center can be built in one to two years. A new transmission line, gas plant, or nuclear reactor takes five to fifteen ^[5][16]. That gap is a big part of why prices spike before new supply ever arrives to answer them.

And the incentives on both sides of the forecasting fight point the same direction: up. Developers gain by filing more requests than they'll ever build, since it costs them little. Utilities earn a regulated return on the capital projects they get approved to build. Both of those facts push reported demand, and planned spending, higher than what may actually be needed ^[12][22].

What's Real, What's a Press Release, and What Nobody Knows

Some fixes are already law, not just talk. Ohio's utility regulator approved AEP Ohio's large-load tariff in July 2025, requiring data centers using 25 megawatts or more to pay for at least 85% of their contracted capacity, whether they use it or not ^[14]. Georgia adopted similar rules the same year ^[15]. Texas signed a law, SB6, in June 2025 that lets the state's grid operator disconnect or curtail data centers during emergencies ^[17]. Those are operating rules today, not proposals.

Other fixes are announced but shakier. Co-located nuclear power and small modular reactors have generated plenty of press releases from tech companies, but the legal path is still being written — the federal rejection of the Amazon-Talen nuclear arrangement is under appeal right now ^[16]. Whether the new Ohio and Georgia tariffs will actually hold up over their full multi-year contracts, without renegotiation, is untested; the rules are only months old ^[14][15].

And some of the biggest questions are genuinely open. Nobody can say with confidence whether U.S. data centers will land near 6.7% or near 12% of the nation's electricity by 2028 — Berkeley Lab's own range is that wide ^[1]. Nobody has cleanly isolated how much of a typical household's bill increase, if any, actually traces back to data centers rather than gas prices or storm damage ^[9][23]. And if a meaningful share of today's forecast demand turns out to be phantom, it isn't yet decided who absorbs the cost of the power plants and transmission lines built to meet it — ratepayers, utility shareholders, or the data-center developers themselves ^[12][22].

That last question is the one worth watching closest. It's being decided, case by case, in state regulatory dockets most people will never read — and the outcome there, more than any national statistic, will determine whether this decade's AI boom shows up on your bill at all.