America's AI power problem isn't generation. It's the substation.

The United States has spent two years arguing about how much electricity the AI buildout will consume. The binding constraint by 2030 is not how much total power the country can generate; it is how much power a specific patch of land can actually receive, and the buildout will be won or lost at substations, not at power plants.

RAND's new research brief formalizes the reframe. Researchers applied a five-part framework, covering power availability, infrastructure readiness, environmental risk, water, and governance, to 22 candidate sites: 17 identified by the US Department of Energy, two private-sector proposals, and three retired or retiring power plants. No site scores well on every dimension; the work is in trading them off. The framework's quiet claim is that legacy industrial sites, with existing substations and transmission rights-of-way, are the realistic near-term path because new high-voltage lines cannot be permitted and built before 2030.

The cleanest single number in RAND's analysis makes the point. A retired Indiana plant, Rockport, may be able to deliver about 4.2 gigawatts of grid-connected power to a single site on the 2030 timeline, and RAND frames that as a practical transmission ceiling rather than a target. The buildout will be decided one site at a time, on the grid's existing seams.