Why Data Centres Are Exploding — and Where They're Being Built

For most of their history, data centres were invisible infrastructure. They sat in business parks, drew a steady and unremarkable amount of power, and quietly ran the websites, video streams and cloud services that the rest of us took for granted. Then artificial intelligence arrived at scale, and the quiet utility became the single fastest-growing source of electricity demand on the planet.

What a data centre actually is

Strip away the mystique and a data centre is a building full of computers that never sleep. Rows of servers run continuously, and nearly all the electricity they consume ends up as heat — which then has to be carried away by cooling systems that draw still more power. Because the machines run around the clock, a data centre's appetite is measured the way a power station's output is: in megawatts. A modest facility might draw a few megawatts; a large one, tens.

That was the normal range for decades. The internet grew, more servers came online, and the total crept up — but steadily, predictably, in step with the slow improvement in how much work each chip could do per watt.

Why the line suddenly bent

The thing that changed is the kind of computing. Training a large AI model is not like serving a billion web pages, which is bursty work spread thinly across many machines. It is one colossal calculation, run across tens of thousands of specialised chips — graphics processors, or GPUs — that are wired tightly together and pushed to nearly full power continuously, for weeks on end.

Two facts about those chips matter. First, each modern AI accelerator draws far more power than an ordinary server processor. Second, an AI training run wants as many of them as possible in one place, talking to each other at high speed. Spread the chips across many small buildings and the calculation slows to a crawl. So the work concentrates — and the power concentrates with it.

The result is a new category of facility. A single frontier AI campus can be designed to draw hundreds of megawatts on one site — comparable to the demand of a small city, and an order of magnitude beyond a conventional data centre. And because several companies are racing to build the most capable models at once, dozens of these campuses are being planned, built and energised in parallel.

What "frontier" means on the map

On the live map, the small charge-blue server-rack marks are the baseline: data centres of every kind, mapped from OpenStreetMap. They show where the world's data quietly lives — dense across the United States and Europe, thinner elsewhere, reflecting both where the industry has historically clustered and how thoroughly each region is mapped in open data.

Picked out on top, in a hotter colour and sized by megawatts, are the largest disclosed AI campuses — the "frontier" sites tracked by Epoch AI. These are the ones driving the load surge: the new builds whose names appear in the news, each tied to a company training large models. Tap one and you can read its operator, its estimated power, the scale of its compute, and what it cost.

A few of these campuses appear as a distinct hollow rack rather than a filled one. Those are announced or under construction — real projects, but with their power not yet online. Drawing them as empty outlines rather than full hot racks keeps the map honest: it marks intent without pretending the capacity is already running.

Where they cluster, and why

The concentration of the biggest campuses in the United States is striking, and it is real. It reflects where the companies building frontier AI are based, and — just as importantly — where large blocks of electricity and grid connections can be secured quickly. Land is part of the equation, but power is the binding constraint. Sites have gone up near cheap energy and spare grid capacity: parts of Texas, the Ohio Valley, Wisconsin, the Southeast. Major projects are also under way in the Middle East, where energy is abundant, and in Europe and East Asia.

This is the story the map tells that a list of numbers cannot: not just that AI is hungry for power, but where that hunger is reshaping the grid — and where the next nuclear plants, gas turbines and transmission lines are likely to follow the load.

Reading it honestly

Two caveats travel with this layer, and both are surfaced on the map itself. The baseline is coverage, not a census — OpenStreetMap is community-mapped, so absence of a dot doesn't prove absence of a data centre, especially outside the well-mapped regions. And the frontier set is curated and currently US-heavy: it tracks the largest disclosed campuses, with power figures that are estimates. That US concentration is a fact about the world right now, not a flaw in the data — and it is exactly the kind of thing a live map is good at showing.