U.S. electricity demand is growing at its fastest pace in decades, driven overwhelmingly by data centers, and grid operators can’t build generation and transmission fast enough to keep up. PJM Interconnection, which manages the grid across 13 states and Washington, D.C., projects its capacity surplus will flip to a nearly 27 GW deficit by 2035—with data centers responsible for about 94 percent of the region’s projected demand growth. A new paper by graduate student Chris Dunlap, a 2025-2026 Bartlett Fellow, argues that the fix doesn’t have to wait on new supply, because much of the load causing the problem is more flexible than grid planners currently assume.
Using four years of day-ahead electricity pricing data across 20 destination zones in five U.S. grid regions, Dunlap tested whether data centers could reliably shift computing load away from a stressed grid to one operating normally elsewhere in the country. In 198 of 200 stress hours observed in Northern Illinois, one of the country’s most power-constrained regions, at least one other grid region was operating normally and available to absorb shifted load. The two exceptions both occurred during a single continental cold snap in January 2024. Combining that geographic flexibility with on-site power scaling, the paper estimates inference-dominant data centers can credibly commit about 40.0 percent of their electricity demand as dial-up/dial-down power.
“Data centers already shift load across regions for their own commercial reasons. The missing piece is a policy framework that makes that flexibility visible to grid operators and lets them plan around it, instead of treating every megawatt of data center demand as fixed,” says Dunlap, a student at the University of Chicago Booth School of Business and Harris School of Public Policy.
The paper proposes two policy mechanisms to put that flexibility to work. Under a flexible interconnection agreement, a data center could energize sooner, before years-long network upgrades are complete, in exchange for a binding commitment to curtail power during grid emergencies. This is similar to a model already used in Texas. Later, as its infrastructure matures, a facility could shift to offering that same flexibility as demand response in PJM’s capacity market. Dunlap argues the approach shifts reliability risk away from ratepayers, who currently bear the cost when the grid falls short, and onto the data center operators best equipped to manage it.
“With hyper-scalers already planning to spend more than $600 billion on infrastructure in 2026, the industry has the capital and the operational sophistication to manage this risk itself,” the paper concludes. “What’s been missing is a policy framework to make that flexibility count.”