AI's next constraint is electric, and it's the grid

The next phase of the AI buildout will be decided in utility interconnection queues, where a project's fate turns on something more prosaic than model performance—whether it can get firm, round-the-clock power—and the demand numbers make the stakes plain.

On the demand side, BloombergNEF expects US data centers to more than quadruple their 25GW draw in 2024 to 106GW by 2035, while McKinsey's energy tally climbs from 147TWh in 2023 to 606TWh by 2030 — nearly 12 percent of the country's electricity. The Department of Energy and Lawrence Berkeley National Laboratory put data centers at about 50GW of the new peak capacity the grid must add by 2030 and say they could already reach 12 percent of US electricity by 2028; other analysts see the number rising as much as 130 percent by 2030. Globally, the International Energy Agency expects data-center electricity use to more than double to roughly 945TWh by 2030, with the United States responsible for the largest single increase.

The economics are pitiless: a GPU that isn't computing is a depreciating asset, and large training and inference workloads must run nearly continuously to justify their capital cost, so intermittency or curtailment directly erodes the return on billions of dollars of compute. A single 100MW IT campus running flat out consumes on the order of 876GWh a year. When interconnection queues stretch for years and high-voltage lines take most of a decade to permit and build, hyperscalers are increasingly building generation behind the meter, placing it directly alongside facilities to secure supply on their own timeline—a workaround that seemed exotic two years ago and is now a design criterion.

For investors, the lesson lands in underwriting. The interconnection queue, as PWD has argued, is an underwriting variable rather than an administrative wait, set in practice by customer class and what a developer will pay for access. The X2M case makes that concrete: its first GPU facility contract arrived with no named customer and no capacity figure, leaving the blanks to mark where the real risk sits. Across the sector, a data center that cannot name firm, round-the-clock power behind it remains a speculative position on someone else's grid problems.

The projections will not land with precision, but their direction is consistent, and with natural gas still leading the US generation mix at near 43 percent, the sustainability side of the problem gets harder before it gets easier. The useful benchmark for the AI trade is shifting from model performance to megawatt firmness: contracted power separates the projects that get built from the ones that remain press releases.