Updated:
U.S. technology companies and utilities are exploring ways to shift or temporarily reduce data center electricity use during periods of peak demand, a strategy that could defer costly power plants and grid upgrades as artificial-intelligence workloads drive electricity needs higher. The approach, known as demand response, is moving from pilots and individual agreements toward broader policy and industry efforts, but widespread use will depend on both data centers and grid operators making operational and market changes.
A February 2026 report by Duke University’s Nicholas Institute estimated that greater data center flexibility could save $40 billion to $150 billion in capital investments over the next decade. The estimate concerns avoided or deferred investment, not guaranteed savings on electricity bills. Meanwhile, forecasts cited by the Electric Power Research Institute (EPRI) put U.S. data center electricity consumption at roughly 177 to 192 terawatt-hours in 2024, with a range of 383 to 793 TWh by 2030.
Shifting demand could ease pressure at peak times
Demand response means reducing or moving electricity consumption when the grid is under stress, rather than requiring facilities to operate at their maximum draw at every moment. Some computing work, particularly certain training tasks, may be delayed or shifted between sites. Operators may also draw on facility resources such as batteries, subject to technical limits and service requirements.
EPRI has reported that surveyed data centers identified potential peak-power reductions of 10% to 30%, depending on facility type, with some hyperscale operators indicating greater potential. Separate EPRI field experiments, described in a 2026 paper, found that AI data center clusters could reduce power use by 10% to 20% during peak grid events with minimal operational disruption. These results show potential in tested settings; they do not establish that every facility can reduce demand by the same amount or for the same duration.
The underlying grid-planning issue is concentrated demand at particular times and locations. If flexible loads can reduce their draw during those periods, utilities and grid operators may be able to rely on existing infrastructure more effectively and defer some additions. The scale of any benefit depends on when and where the flexibility is available, and whether operators can count on it when needed.
Early agreements and industry coordination
So far, data center demand-response efforts have largely taken the form of pilot projects or one-off arrangements. Reuters reported that OpenAI agreed to reduce grid withdrawals by as much as 1 gigawatt during periods of grid stress at a planned 3.2-gigawatt facility in Georgia. The reported commitment concerns curtailment during specified conditions, rather than a permanent reduction in the facility’s planned capacity.
In September, Alphabet’s Google, Nvidia and Emerald AI launched the AI Energy Management Alliance to advance flexible data center operations. The initiative adds industry coordination to a field where individual agreements have so far been more common than standardized arrangements across regions and facilities.
The policy landscape is also evolving. On June 18, 2026, the Federal Energy Regulatory Commission directed all six regional grid operators under its jurisdiction to justify or reform tariffs governing how data centers and other large electricity users connect to the grid. FERC said its orders addressed faster integration alongside consumer safeguards. They opened a process for changes; they did not, by themselves, establish a single nationwide flexible-load program.
Scaling requires workable incentives and safeguards
Broader adoption would require data center operators to control electricity demand without unacceptably disrupting customers or computing services. The ability to shift workload varies by task, facility design and operating requirements, while the grid may need reductions at short notice and in specific locations. Those differences make it difficult to treat all data center demand as equally flexible.
Utilities and grid operators, for their part, need tariffs and market rules that define how curtailments are requested, measured and compensated. They also need interconnection procedures that recognize credible flexibility commitments. EPRI’s January 2026 review of large-load tariffs found that demand-response credits can vary with the terms of participation, including advance notice and the number of interruption hours allowed.
Alexander Kheder, an analyst with BMI, a unit of Fitch Solutions, told Reuters that expanding curtailment agreements across hundreds of new facilities would require significant capital expenditure and coordinated policy frameworks. The effort therefore involves more than installing controls: operators, utilities and regulators must agree on dependable commitments and how their value is reflected in grid planning.
What remains unsettled
The $40 billion to $150 billion Duke estimate points to a possible range of avoided capital investment, not a forecast that the full amount will be realized. The outcome would depend on how much load can be shifted, how reliably facilities respond, and whether grid investments can actually be deferred rather than merely postponed or relocated.
FERC’s June action puts tariff reform and large-load interconnection rules into focus, while the industry alliance and individual agreements indicate growing interest in operational flexibility. The available reporting does not specify a nationwide timetable for implementing new demand-response standards, nor does it establish how many facilities will participate. For now, the central question is whether the promising reductions demonstrated in surveys and field experiments can be turned into reliable, repeatable commitments at the scale required by a rapidly expanding data center fleet.
