Could the answer to America’s ever-escalating electricity bills be hiding in plain sight? For decades, the U.S. grid has been engineered to withstand rare but extreme peaks in demand-those sweltering summer afternoons or frigid winter nights when power use surges and outages could be life-threatening. This conservative design philosophy means that for most of the year, vast amounts of transmission and generation capacity sit idle. In some rural areas, utilization can be as low as 30%, while even in dense urban zones it rarely exceeds 70%.
The challenge now is that peak demand is projected to grow by nearly 24% in the next five years, driven in large part by explosive expansion of AI-focused hyperscale data centers. These facilities, consuming upwards of more electricity than entire cities, strain local grids and are prompting utilities to invest billions in new infrastructure often fossil-fueled costs passed directly to ratepayers. In Virginia’s PJM market, for example, growth in the data center sector has contributed to a $9.3 billion capacity price increase, raising monthly residential bills by $16–$18 in some areas.
Quantifying the untapped potential of this “curtailment-enabled headroom.” has been a key project for researchers at Duke University’s Nicholas Institute. Their modeling work indicates that if new large loads can accept being curtailed for just 0.5% of annual uptime, or roughly 177 hours a year, the existing grid could absorb close to 100 GW of such additional consumption, equal to that of dozens of major cities. Data centers, EV charging hubs, and industrial facilities can be integrated with no need for expensive capacity additions by using flexible load strategies like shifting computations to off-peak periods of the day and using on-site generation.
Some utilities and tech companies are already experimenting with this model. Portland General Electric (PGE) tapped startup GridCARE to deploy AI-driven modeling tools that detect spare capacity and optimize flexible resources like batteries and microgrids. That enabled PGE to approve 80 MW of new data center interconnections in Hillsboro, Ore. years sooner than traditional grid upgrades would allow. “If they’re willing to be flexible, we’ll put them at the top of the queue,” said Larry Bekkedahl, PGE’s senior vice president for advanced energy delivery.
Google has also started incorporating machine learning workloads into its demand response programs. In deals with Indiana Michigan Power and Tennessee Valley Authority, the company’s data centers will cut power consumption related to ML during times of grid stress, extending earlier pilots that shifted non-critical compute loads. That evolution is important: AI-optimized servers can consume two to four times more power than traditional hardware, and cooling systems can account for up to 30% of a facility’s load. By orchestrating clusters of GPUs in real time, pilots like the Oracle data center in Phoenix have reduced peak demand by 25%.
The solutions underneath are sophisticated engineering. Flexible interconnection models require granular, hourly load forecasting, scenario analysis across thousands of contingencies, and integration of distributed energy resources. GridCARE’s platform, for example, runs over 200,000 permutations of load ramping, storage dispatch, and nodal constraints in search of the lowest-cost pathway to add capacity. This computational rigor is required when balancing multiple 50-500 MW data center projects against regional transmission limits.
Demand response technology is evolving to meet these needs: traditional programs enrolled industrial customers to shed load during emergencies; now, hyperscalers are experimenting with “batchable workloads” and hybridizing renewable generation with battery storage to create dispatchable capacity. The Nicholas Institute report points out that in 88% of the curtailment hours, half of the flexible load could stay online to minimize operational disruption.
The policy framework will be the determinant to the scaling of such innovations. Separate rate classes for data centers set by state utility commissions should ensure that they bear full costs of infrastructure built to serve them. Incentives towards flexibility, such as expedited interconnection or tax benefits subject to curtailment commitments, could accelerate the pace. International precedents, like Germany’s requirement for 100% renewable sourcing by 2030 for all data centers, show how regulatory levers can align economic growth with grid stability.
For utility planners and tech infrastructure managers alike, the message is: the fastest, cheapest capacity is the capacity that already exists. By unleashing the grid’s hidden half through engineered flexibility, the United States can meet surging AI and electrification demands without sacrificing one ounce of affordability or reliability.
