Let’s talk about the nine dollars that showed up on a power bill in Maryland for a building sixty miles away that has no windows and doesn’t send anyone a bill.
Picture a woman in a rowhouse in Baltimore. Same house she’s been in for nineteen years. Same appliances. Same two kids, one of whom has finally moved out. Her bill goes up. Not because she used more. Because a grid operator ran an auction and the price of capacity cleared at the ceiling. She has no line item that says “inference.” There’s no box on the form for “someone else’s model.” The number just moves, and she pays it, because the alternative is the lights going off.
That’s the thing I want to be precise about. The AI build-out has a bill. Everyone has spent two years arguing about who captures the upside. Almost nobody has asked who signs for the downside.
Here’s what happened. PJM runs the largest wholesale electricity market in the United States — thirteen states plus DC, sixty-five million people. In its 2026/27 capacity auction, the price cleared at the cap. Three hundred and twenty-nine dollars and seventeen cents per megawatt-day, across the entire footprint. Sixteen point one billion dollars. The year before cleared at two hundred and seventy. Two years before that, roughly a tenth of that.
PJM’s own filings say data centres make up more than ninety percent of the new demand it expects by the end of the decade. The New York Times reported in July that since 2024 those auctions have added about twenty-nine billion dollars of cost to utility customers across the region, because of data centres. ICF, the consultancy, put the residential retail rate impact at thirty to sixty percent by 2030 if nothing changes. PJM expects the first pass-through — one and a half to five percent — to hit bills from June this year.
Ireland is the finished version of the story. Data centres went from five percent of metered electricity in 2015 to twenty-three percent in 2025, according to the Central Statistics Office. Friends of the Earth Ireland estimates the average household paid about three hundred and sixty euro extra between 2015 and 2023, with another three hundred to six hundred and forty coming by 2034. Same country, same wires, no vote.
Now, the honest counterweight, because I’m not going to sell you a panic. Rutgers ran the numbers on New Jersey and found homes in ZIP codes with a data centre were paying about two dollars and nine cents a month more than homes without one. Two bucks. Their own conclusion was “mostly not yet.” That’s the accurate state of play in most of America. Which is exactly why this is the moment to look at it, because “not yet” is a schedule, not a verdict.
So what’s the pattern?
A regulated utility earns a return on capital. Not on how well it serves you — on how much it builds. Growth in demand is not a problem for a utility. It’s the business model. When a hyperscaler walks in and asks for two gigawatts, the utility gets to build substations, lines, generation, and earn a guaranteed return on all of it. The cost of that capital gets spread across every customer on the system, because that is how regulated cost-of-service ratemaking works. It’s designed that way on purpose, and the purpose was good: the grid got built because everybody paid a little so that everybody could have light.
Except the payer and the beneficiary used to be roughly the same people.
Now they’re not. The revenue from that two-gigawatt campus goes to a company in California. The capacity cost goes to a pensioner in Ohio. And there is no meter on earth that tells her which kilowatt-hour of her bill was hers and which was the model’s. That’s the structural problem, and it’s not a conspiracy — it’s a tariff that hasn’t been rewritten.
Which is where people are missing the actual fight. The interesting question in 2026 isn’t “is AI too power-hungry.” It’s “who is in which rate class.”
Pennsylvania’s Public Utility Commission advanced a model large-load tariff in April — a separate class for customers pulling twenty-five megawatts or more, with its own terms. Columbia’s climate law blog has been walking local governments through the mechanics. AEP Ohio settled a case that put minimum-take obligations and exit fees on big loads. This is unglamorous municipal stuff, and it decides more about the next decade of AI than any model release. Where the large load has to bring its own generation, guarantee its demand for fifteen years, and pay to leave, the household stops moving. Where it doesn’t, the household is the lender of last resort. It’s that simple.
And notice the shape of it, because you’ve seen this shape before.
Free inference, paid for with your context. Free storage, paid for with your switching costs later. Free pilots, paid for with a margin nobody’s measured. The bill always exists. It always lands somewhere. And it always lands on the party with the least ability to say no — because that’s the entire point of picking where to send it.
So if you run a business, here’s the operator lesson, and it’s not abstract.
Put a power line in your AI unit economics. Not “cloud.” Power. If you’re on a colo contract, read the pass-through clause — capacity charges, demand charges, and the escalator are where the surprises live. If your provider’s price looks too cheap for the compute it’s delivering, ask what it’s assuming about energy costs and when that assumption expires. Because a price built on subsidised electricity is a price with a political clock on it, and the clock is being wound right now in state capitals.
In South Africa this isn’t theoretical at all. Any business here already prices energy risk — solar, batteries, generators, the cost of a bad week. The data centre build landing on this grid is competing for the same megawatts, and the tariff question arrives here the same way it arrives in Pennsylvania: who gets the cheap bulk deal, and who absorbs the difference.
The principle worth keeping is smaller than it sounds. Nobody should be billed for a decision they were never party to. Not a model’s training run, not a boardroom’s capacity reservation, not somebody else’s growth story. If a machine’s cost lands on your household, you should be able to see it, name it, and argue about it in a room with a regulator in it.
Separate the meter. Name the payer.
It’s a boring answer. Boring answers are the ones that actually hold.
You’ve got this — but only if you check whose name is on the bill.