Let’s talk about why the cheapest AI on the market sits on top of the most expensive thing we have ever built.
There is a man in Manassas, Virginia. John Steinbach. Nearly forty years in the same house.
In December, his electricity bill was about a hundred dollars. In January, two hundred and eighty-one.
He did not change anything. The county around him did. Virginia runs close to six hundred data centres now — the densest concentration on earth — and a single large one can draw as much power as a hundred thousand homes. Steinbach’s bill did not move because he used more electricity. It moved because the building next door did.
Hold that man in your head. Because this is the part of the AI story almost nobody is putting on the invoice.
For two years we have been told the story of AI through price. Tokens got cheap. Then cheaper. DeepSeek broke the floor, the big labs piled in behind it, and every launch email promised you more intelligence for less money. It read like the best kind of competition — a commodity war where the customer wins. And that part is true. Inference is cheap now, and getting cheaper.
But cheap on top of what?
Every cheap token is produced by a machine that eats electricity at a scale we have never had to feed before. A typical hyperscale data centre draws around a hundred megawatts, which the IEA translates, carefully, as the consumption of a hundred thousand households. Meta is building a campus in Louisiana — Hyperion — across three thousand six hundred acres, and it will need at least five gigawatts. That is three times the power of the entire city of New Orleans. For one company’s building.
US data centres went from roughly eighty gigawatts of demand in 2025 toward a projected hundred and fifty by 2028. Lawrence Berkeley reckons that by 2028 they could swallow twelve percent of everything America generates. And the hyperscalers are committing something in the region of seven hundred and twenty-five billion dollars to this build-out this year alone.
Now watch where the smart money stopped buying, and started owning.
Google didn’t just sign a power contract. It bought the company. Intersect Power — four point seven five billion dollars. Microsoft is pushing for fusion by 2028 and experimenting with hydrogen on-site. In Memphis, xAI is running gas turbines to keep its own lights on. The largest technology companies in the world have quietly become energy companies, because they did the maths: you cannot win a race in cheap intelligence if someone else controls the thing that makes it.
This isn’t a new pattern. It’s the oldest pattern in infrastructure. When the visible layer commoditises, the value retreats downhill to the chokepoint. Railways owned the tracks. Telecoms owned the line. Cloud ran the same play — give away the easy part, own the thing nobody can route around.
The thing nobody can route around now is the grid. And the land. And the water.
So when I say the grid is becoming the new app store, that is not a metaphor for effect. It is the actual structure. A toll booth is being built between human intention and machine output, and the toll is not thirty percent of your app revenue. It’s megawatts, priced into somebody’s monthly bill.
Here is the part that should stop you. The bill is already arriving.
One Bloomberg analysis found wholesale power in the densest pockets running as much as two hundred and sixty-seven percent above where it was five years earlier. The Dallas Fed put a more careful number on it: existing data centres have already pushed wholesale prices two to six percent higher nationwide, with sharper spikes where they cluster, and the middle forecast is twenty to thirty percent higher generating costs by 2028. Almost three quarters of Virginia voters, asked who they blamed for rising electricity costs, pointed at the data centres. Consumer research found seventy-eight percent of Americans worried the build-out will raise their bills. Lawmakers in more than thirty states have filed over three hundred bills about it this year — moratoriums, rate classes, disclosure rules, tax fights.
And the companies know. In January, Microsoft and Anthropic signed a White House ratepayer pledge, promising to cover the electricity price increases their data centres cause. Sounds like the fix. It isn’t. As Harvard’s electricity law man Ari Peskoe pointed out, the companies don’t hold the pen. Utilities and utility regulators decide who pays for grid upgrades — that’s how the utility business model has always worked. A pledge is a press release. A rate case is a fact.
So here’s what I want you to see clearly, because it’s what people are missing.
The subsidy is invisible. It doesn’t arrive as a line item called “AI.” It arrives as a slightly higher bill you cannot trace to anything you did. That is the most dangerous kind of cost there is — the kind with an author you never see. The cheapest intelligence in history is being financed, in part, by household budgets in Manassas and Memphis and every place a campus lands.
And that reframes something we keep getting wrong. We treat AI dependency as a software problem. Which model, which vendor, which API. Keep the login portable, keep the budget uncommitted, don’t get captured by one ecosystem. All of that is right. It is also the shallow layer.
Because you can switch vendors in an afternoon. You cannot switch grids. Ever. The grid is the one dependency with no export button, no migration path, no competitor standing by. When your unit economics quietly rest on cheap inference, you are not buying a commodity. You are renting somebody’s loss-leader — and the loss is being carried by a household that never clicked agree.
The “sovereign AI” crowd has this backwards too. Running the model on your own machine does nothing if the power contract is signed in someone else’s name. Sovereignty was never a location. It’s a set of keys — and the most important key in the building is the one that turns on the lights.
So the operator lesson, in plain terms. Price your stack as if inference costs real money, because eventually it will. Don’t build a margin on a subsidy you don’t control. Watch the energy line the way you watch the API bill, because it’s about to become the same line. And when a platform offers you cheap compute, ask the only question that matters: who is actually paying for this, and what do they get from me later?
Steinbach’s bill has an author. He just can’t see the signature on it. And that’s the thing about this whole build-out people keep missing — abundance and subsidy look identical from the outside. The only way you ever tell them apart is by asking who picks up the tab when the music stops.
You’ve got this. Just keep an eye on the meter.