The gigawatt

Compute is an electricity problem wearing a software costume. The country that can connect firm power to racks will train and infer. The country that cannot will ration. China built generation first. The United States is asking the grid to absorb a new heavy industry after twenty years of flat demand.

The load

U.S. data centers used on the order of 313 TWh in 2025 — nearly 40 percent of world data-center electricity. Forecasts put U.S. data-center demand on a path to roughly 10–15 percent of national electricity by 2030 in government-adjacent estimates, with interconnection capacity measured in the low hundreds of gigawatts if the buildout lands. Goldman’s 2026 commodities desk put U.S. data-center power demand rising from about 31 GW in 2025 toward 41 GW in 2026 and 66 GW in 2027. Those are planning numbers. Interconnection queues, transformers, and gas turbines decide what actually energizes.

China’s installed generating capacity was reported at 4,080 GW by July 2026, up 11 percent year over year. That is more than twice the U.S. system. Beijing adds the output of a mid-size country in a single year and still has room for compute in the west under the Eastern Data, Western Compute layout. Industrial power is cheaper there. Cooling water is cheaper there.

The county problem

This is where the register meets the series. Data-center fights show up as hearings, water permits, and base-land proposals. Communities resist because a campus can draw like a city. Lawmakers have already asked whether Pentagon land should host civilian racks that strain base power and local aquifers. That is not an abstract AI race. It is a notice on a calendar.

Firm power is the scarce input. Solar and batteries are being added at record clips in the United States. They do not, by themselves, give a training cluster the 24-hour bus it wants. Gas peakers and delayed nuclear do not arrive on the same clock as a GPU purchase order.

Opportunity cost, measured in watts

A megawatt that serves a training run is a megawatt that is not serving a munitions plant, a yard, or a radar site that has to stay up in the dark. Dual-use power is fine in peacetime. In a crisis the grid is a target and a rationing board. Concentrated data-center alleys are both.

Bottom line

The AI race is already an energy-industrial race. China is running a marathon of capacity. The United States is sprinting gas and permits to keep racks from waiting on the queue. Racks that never interconnect do not train. Models that cannot infer at the edge do not fight.