A single answer from the smartest model on earth costs a fraction of a cent. The megawatt that runs a million of them, the water that cools it, and the town that decides whether it gets built do not come cheap.
For three years the AI conversation inside institutions was a conversation about models. Which one is smartest, which is cheapest, which to standardize on. That part got easy, because the model stopped being scarce. Capability falls in price every quarter. Open weights sit a few months behind the frontier. A strong model runs on hardware a mid-sized firm can own. The thing everyone spent three years choosing became a cheap and swappable input.
The scarce thing moved. It moved off the chip and into the physical world: power, water, land, transmission, permits, and the consent of the people who live next to where the machines go.
The numbers are hard to argue with. In 2025, data centres consumed 23 percent of Ireland's metered electricity, close to what every home in the country used combined, according to its Central Statistics Office. In July 2026, New York became the first state to freeze new large data centre construction, pausing projects at or above 50 megawatts while it writes rules that do not yet exist. The same week, opponents staged 142 protests across 42 states in a single coordinated day. A June poll found only about a third of Americans approve of the current pace of data centre construction. None of that moves with a benchmark score.
There is an old pattern in economics for why this arrived so fast. When a resource gets cheaper to use, total consumption tends to rise rather than fall. Cheaper steam engines burned more coal, not less. Cheap intelligence behaves the same way in aggregate. The cheaper a model gets, the more of it the world runs, and all of that running lands somewhere physical: on a grid that takes a decade to expand, in a county that votes, next to a river with a fixed flow.
So the cost did not vanish when the model got cheap. It changed address. It left the vendor's invoice and reappeared as an interconnection queue, a zoning fight, and a capacity limit your supplier has not mentioned yet.
The smartest model in the building is worth nothing if the building cannot get power.
Here is the fair objection, and a good CTO raises it first. That is my cloud provider's problem, not mine. I do not sit in a county hearing. I buy inference by the token, and someone else pays the power bill and fights the zoning board.
For most institutions, most of the time, that is correct, and it is a rational division of labour. The specialists who own the physical layer are better and cheaper at it than you will ever be. If that describes you, the lesson is not to go build a substation.
But the constraint does not disappear when you rent past it. It turns into a different exposure, one you do not control. A moratorium in one state and a grid cap in one region tighten supply and lift price for everyone downstream, including you. Your roadmap starts to depend on a bottleneck you cannot see and cannot influence: where your compute physically sits, who else is competing for the same power, and how concentrated your supplier's footprint is in places that are turning against it. The physical layer became a first-order risk in your AI plan whether or not you ever touch it. Renting is not the mistake. The mistake is not noticing that the thing gating your roadmap moved from the model to the map.
For one set of institutions the constraint binds directly, with no supplier in between. If your data cannot leave the jurisdiction, if a regulator needs to see where every inference ran, or if you run enough inference that the supplier's margin dwarfs the physical cost, then the physical layer is your problem by law, by mandate, or by arithmetic. For a sovereign, a regulated bank, or an operator at real scale, owning into the constraint is a clear-eyed read of your own exposure rather than an ideological pose. It is the segment we build for, and for that segment a system that ignores where it will run is not finished. It is a demo with a deadline.
So the reframe for anyone allocating capital this year is narrower than own everything, and more useful for it. Capability is now the commodity input, and commodities do not decide strategy. The decisions that decide strategy have a physical address: where the compute lives, what powers it, who else is fighting for that power, and how exposed you are to a bottleneck you may not own.
Two years ago the question was which model. The question now is where it runs, what feeds it, and who lets you. The institutions that keep answering the first question will keep being surprised by the second.
The model was never the hard part. We just could not see the hard part until the model got cheap enough to stop blocking the view.