Data center electricity consumption in the United States is projected to reach 292 terawatt hours in 2026, equal to 6.5 percent of total national power demand, according to McKinsey analysis. The figure places a single industrial category on par with the electricity consumption of several mid sized states combined.

The cost structure behind that load is shifting the economics of new capacity. Power procurement has moved from a line item in site selection to a gating factor, and developers now negotiate generation and transmission access before committing to land. Utilities in constrained regions have begun requiring large load customers to fund transmission upgrades directly or to bring their own generation.

The demand curve for AI training and inference workloads differs from traditional enterprise computing in shape as well as size. Training runs draw near constant power for extended periods, and inference load follows usage patterns that concentrate during business hours in each region served. Both profiles are less flexible than the interruptible industrial loads utilities have historically used to manage peaks.

Capital allocation reflects the constraint. Power supply under construction to serve data centers has reached multiple gigawatts in the United States, and a growing share of new nuclear, gas, and solar capacity announcements now cite data center offtake as the anchor commitment rather than general system growth.

Source: McKinsey - https://www.mckinsey.com/featured-insights/week-in-charts/ais-power-binge