AI's next decade will be settled behind a transformer

The next decade of AI will not be won or lost on chips. It will be settled on a smaller, duller question: how many clean, firm megawatts a utility can actually deliver to a given substation by a given year. Total capacity on a slide is not the same thing as power that runs twenty-four hours a day without a gas turbine behind it. The two numbers travel under one name and bind nothing alike.

AI Tool Discovery's reading of the IEA makes the gap visible. Global data centers drew 415 TWh in 2024, roughly 1.5 percent of world electricity. The same agency's central scenario puts that figure at 945 TWh by 2030, exceeding everything Japan uses today. The headline reads as a demand story. The structural story is a supply story, and the supply in question is not megawatts but firm, clean, time-matched megawatt-hours.

Treat any hyperscaler announcement as a request, not a contract. Read the utility's integrated resource plan before the press release. Check the interconnection queue, the transformer lead time, and the share of new load that actually clears without a behind-the-meter gas plant. The hyperscalers with early PPAs and the small set of former crypto sites already wired to the grid are the firms sitting inside the constraint. Everyone else is renting an assumption.