The cloud has a smokestack

We picked a soft, weightless word for it: the cloud. It makes AI feel like it happens nowhere — a query floats up, an answer floats down. The reality is the opposite. Every prompt you send runs through a warehouse the size of a stadium, packed with chips that draw enormous power and need constant cooling. And in 2026, those buildings have gotten numerous enough that the physical cost is no longer invisible. It's showing up on power grids, in water tables, and on other people's electricity bills.

This is worth thinking about even if you only ever touch AI through a keyboard, because the most physical thing about the industry is quietly becoming the thing that limits it.

The numbers are not small anymore

Start with scale. Global data-center electricity use is on track to roughly double between 2025 and 2028 — from about 80 to 150 gigawatts, the equivalent of adding a country the size of Spain in three years. By 2028, data centers could draw around 12% of all US electricity. If the world's data centers were a country, they'd already rank among the top few electricity consumers on the planet.

Per query it sounds tiny — somewhere around 0.3 to 3 watt-hours, with the heavy end roughly ten times a Google search — plus water for cooling; one estimate puts a 100-word prompt at around half a liter in the worst regions. Tiny times billions is not tiny. And the part people miss: much of this new demand is being met by fossil fuels, with some big sites bypassing the public grid entirely to get power fast. The "cloud" has a smokestack, and right now it's often burning gas.

The externality is turning into a constraint

For a while the energy and water cost of AI was an externality — a real cost, but one the industry didn't have to feel. That's ending, and it's ending in the most concrete way possible: local people are paying for it and pushing back.

Where data centers cluster, grids are buckling. Ireland's data centers already eat more than 20% of the country's electricity; Northern Virginia, Texas, and Phoenix are facing serious grid-expansion costs that land on ratepayers. The backlash is now political, not just grumbling: in 2026 so far, lawmakers across more than 30 states introduced over 300 bills on data centers — moratoriums, tax-incentive rollbacks, mandatory energy-and-water reporting. Recall from the IPO piece that public comfort with a local data center collapsed from 69% to 35% in three years. That's a social licence eroding in real time.

When the limiting factor stops being "can we build a better model" and becomes "can we get the power, the water, and the permit," you're not in a software constraint anymore. You're in a physical one — and physical constraints don't get solved by a smarter algorithm next quarter.

What this means if you build

You're not running a data center. But the physical reality reaches your keyboard in a few ways worth sitting with:

• Stop treating compute as weightless and infinite. The mental model of "AI intelligence is basically free and unlimited" is wrong at the level of physics, not just budget. There's a real ceiling — megawatts — and it's getting closer, which is part of why prices and availability won't simply fall forever.
• Efficiency is now an ethical and strategic choice, not just a cost line. The same moves that cut your bill — routing easy work to small models, caching, not sending a frontier model to do trivial work — are also the moves that use less power and water. Wasteful AI design now has a footprint someone else pays for, and increasingly a regulatory one too.
• Expect the physical layer to shape the product layer. Power scarcity, reporting rules, and moratoriums will influence where compute is available, what it costs, and how the public feels about AI generally. Building as if none of that exists is building on an assumption that's already wobbling.

The bottom line

"The cloud" was always a marketing word for "someone else's computers." In the AI era it's a marketing word for "someone else's computers, drawing a power-plant's worth of electricity and a river's worth of water, near someone else's town." None of that makes AI bad or means you should stop building. It means the weightless mental model is false, and the people who internalize that early will design leaner, defend their social licence better, and be less surprised when power — not cleverness — turns out to be the binding constraint.

So the next time a query floats up and an answer floats down, remember it didn't happen in a cloud. It happened in a building with a smokestack, and the bill for it is increasingly landing on people who never typed the prompt. Build like that's true, because it is.