Artificial intelligence is being built faster than the power to run it. The race to pour concrete for data centres has collided with a grid that was never designed for this — and something has to give.
A single new AI data centre can ask its local utility for as much electricity as a mid-sized city. Developers are now requesting that much in dozens of places at once — and the grid operators, quietly, are starting to say no.
The boom that everyone can see — the models, the chatbots, the trillion-dollar valuations — rests on an unglamorous foundation that almost no one is watching: whether there is enough power, in the right places, to keep the machines fed. Increasingly, there is not.
For twenty years, electricity demand in most developed economies was flat. Efficiency gains cancelled out growth; utilities planned for a future that looked much like the present. Then, almost overnight, the forecasts broke.
The cause is concentrated and enormous. A frontier AI training cluster can draw more power than a steel mill. A single hyperscale campus can rival the peak demand of a small city — and the companies building them want not one but many, connected as fast as the concrete can cure.
Utilities that spent a generation managing gentle decline are now confronting requests that would, in some regions, double the load on the local grid within a few years. The wires, the transformers, and the generation simply are not there.
Almost all the electricity a data centre consumes ends up as heat that has to be moved somewhere else. Cooling can add a third again to the power bill — and, increasingly, a thirst for water that pits the campus against the towns around it.
A developer can pour a data hall in under a year. The substation that feeds it, and the transmission lines behind that, can take five. The mismatch is the whole story: capital moves at software speed, infrastructure at concrete speed.
In the most contested regions, the waiting list to connect new load has swollen to years. Some operators have begun turning projects away outright — the first time in living memory that the grid, not the market, is the binding constraint on growth.
Power is only the first constraint. Many of the fastest, cheapest cooling designs drink water — millions of litres a day at a large campus — and a striking number of the sites being chased sit in exactly the dry, sunny regions where water is already fought over.
The result is a quiet collision between two of the decade's defining stories: the AI build-out and the deepening scarcity of water in the American West and beyond. The data centre and the reservoir are, increasingly, drawing on the same shrinking account.
“We spent twenty years planning for a grid that shrank. Now we have to build one that doubles — and we have to do it before the demand arrives, not after.— A regional grid planner
The companies driving the demand are among the richest on earth, and they are moving to secure their own power — signing deals for nuclear output, funding new generation, promising to bring their own electrons. Some of that is real. Some of it is a press release.
The unglamorous truth is that the grid is a commons, and the AI boom is the largest new claim on it in a generation. Whether that claim is met with new capacity or simply borrowed from everyone else's supply is now one of the most consequential questions in technology — and it will be decided not in a lab, but in utility hearings almost no one attends.
This entire story — the cinematic photography, the scroll-driven graphics, the reader login and paywall, the AI features behind it — runs on Cloudflare's Developer Platform. No origin servers. No separate CDN. One deploy, live in 300+ cities. This is what a modern media product looks like when the platform is the newsroom's advantage.
Every page, this interactive, the paywall, and the API render at the edge — closest to each reader — with nothing to provision or scale.
Responsive images and adaptive video delivery for immersive stories — served fast on any device, any connection.
Zero-egress object storage for media, and low-latency KV for subscriber preferences — read in single-digit milliseconds worldwide.
Semantic search, article-to-podcast narration, and daily-brief summaries — inference runs on the same platform as the site.
OAuth login, a real paywall, and an MCP endpoint that lets a reader's AI agent access paid content — with the publisher always in control.
Stateful sessions for the AI agent and a scheduled Worker that emails every subscriber their personalised brief — no servers, no cron box.
Immersive storytelling, AI features, subscriptions, and global performance — on one platform your engineers will actually enjoy shipping on.
Explore the Developer Platform →