Powering AI: Why Overbuilding Data Center Energy May Not Be the Disaster Some Fear

A new McKinsey report tackles a question that has been circulating in boardrooms and policy circles alike: is the United States at risk of overbuilding power capacity for artificial‑intelligence (AI) data centers? The study, based on a blend of scenario modeling and historical demand patterns, concludes that the risk of stranded electricity assets is lower than many alarmists predict.

The analysis begins by quantifying the scale of power projects already under way or announced for the next decade. According to the report, roughly 30 gigawatts of new generation capacity—primarily natural‑gas and renewable projects—are earmarked to serve the expanding compute footprint. This represents a 15% increase over current grid capacity dedicated to data‑center loads.

Crucially, McKinsey points to three operational levers that mitigate the overbuilding risk. First, demand elasticity: AI workloads can be shifted temporally, using demand‑response programs to align compute peaks with periods of abundant renewable generation. Second, modular scaling: modern data‑center designs allow operators to add compute racks incrementally, avoiding the need for monolithic power contracts. Third, cross‑industry load sharing: excess capacity can be redirected to other high‑intensity sectors, such as electric‑vehicle charging or industrial processes, thereby improving overall asset utilization.

From a metrics perspective, the report estimates that even in a “low‑adoption” scenario—where AI compute growth stalls at 5% annual CAGR—utilization rates for the newly built power assets would still hover around 70% of capacity within five years. By contrast, a “high‑adoption” trajectory (12% CAGR) would push utilization above 90%, delivering a better return on investment for utilities and reducing the per‑kilowatt‑hour cost of AI services.

The implications for the AI ecosystem are twofold. Operationally, firms can justify continued investment in high‑density compute without fearing immediate regulatory backlash over energy waste. Strategically, the finding encourages tighter integration between AI developers and grid operators, fostering collaborative planning that aligns compute cycles with renewable supply curves. However, the report also warns that complacency could erode cost advantages if providers ignore demand‑response opportunities or over‑commit to single‑fuel generation.

In practice, the takeaway for enterprise leaders is clear: focus on measurable efficiency gains—such as workload scheduling, cooling optimization, and modular expansion—rather than speculative concerns about stranded power. By doing so, they can harness AI’s productivity upside while keeping the electric bill, and the broader sustainability narrative, in check.