Almost every debate about powering data centers starts from the same assumption: a new load needs new generation, sized to its peak, and new transmission to deliver it. A report published on February 11 by Duke University's Nicholas Institute for Energy, Environment and Sustainability challenges that assumption with a simple question. How much new load could the existing US power system serve if that load agreed to reduce consumption during the small number of hours when the system is under the most stress?
The answer is large. Across 22 of the largest balancing authorities serving 95% of US peak load, the authors estimate that 76 GW of new load, about 10% of current aggregate peak demand, could be integrated with an average annual curtailment rate of 0.25%. That means the new load would be curtailed for 0.25% of its maximum possible uptime, roughly 22 hours' worth of full output a year. At a 0.5% curtailment rate, the figure rises to 98 GW, and at 1%, to 126 GW.
Where the headroom comes from
The logic rests on how power systems are planned. Utilities and grid operators build enough capacity to meet the highest demand of the year plus a reserve margin, even though that peak occurs only for a handful of hours, usually on the hottest summer afternoons or the coldest winter mornings. For the rest of the year, a substantial share of capacity sits idle or runs at partial output. The authors call that spare capacity headroom.
A constant new load, such as a data center that runs flat out around the clock, can use that headroom most of the time. The problem arises only in the peak hours, when there is no spare capacity left. If the new load can step down in those hours, it does not add to the peak, and the system does not need new capacity to serve it. The report finds that a system's potential to serve new demand without capacity expansion depends primarily on its load factor: the more peaky the existing demand, the more headroom there is in off-peak hours.
Short and partial curtailments
The details matter as much as the totals. The average curtailment event would be short: 1.7 hours at a 0.25% annual limit, 2.1 hours at 0.5% and 2.5 hours at 1%. And most events would not require shutting the new load down. In nearly 90% of the hours in which curtailment is needed, at least half of the new load could keep running. The number of curtailment hours per year is comparable to those of existing US demand response programs.
That changes the practical question for data center operators. A facility asked to shed half its load for two hours a few times a year can shift some computing work to other sites, defer batch jobs, or run on-site backup generation or batteries for that period. That is a very different proposition from being offline for days.
Which grids have the most room
The report ranks the balancing authorities with the largest potential at a 0.5% curtailment rate: PJM at 18 GW, MISO at 15 GW, ERCOT at 10 GW, SPP at 10 GW and Southern Company at 8 GW. Those are also the regions where data center demand is growing fastest. In PJM, where data center load is driving capacity prices sharply higher, 18 GW of flexible headroom is a large number relative to the forecast growth over the next five years.
The report also frames the context. Aggregated US winter peak load is forecast to grow by 21.5% over the next decade, from about 694 GW in 2024 to 843 GW in 2034, according to NERC's 2024 Long-Term Reliability Assessment. FERC's latest five-year outlook forecasts 128 GW of peak load growth as early as 2029. Against those numbers, flexible load is not a complete answer, but it could cover a meaningful share of near-term growth while new generation is built.
The caveats
The authors are explicit about what the study does not do. It treats the new load as constant and estimates headroom at the level of each balancing authority, which means it does not model transmission constraints within that area. A data center connected to a congested part of the network may not be able to use headroom that exists elsewhere in the same system. Nor does the study model generator operating limits, fuel constraints or the reliability standards that govern operating reserves in detail. It is a screening analysis, a measure of how much room exists, not a plan for using it.
There is also a commercial question. AI training and inference are extremely capital-intensive, and operators are reluctant to idle expensive chips even for short periods. The incentive to accept curtailment will depend on what they get in return. If flexibility earns faster interconnection, data centers that would otherwise wait years for firm service may find curtailment a bargain.
How it compares with building
The alternative to flexibility is new capacity sized for the peak. A gas turbine built only to cover a few dozen hours a year of data center demand is an expensive asset to carry, and in PJM and elsewhere the wait for turbines and interconnection now runs to several years. Paying a large load to step back for those hours can be far cheaper than building a plant to serve them, provided the curtailment is dependable.
Why it matters
The study's main contribution is to reframe the policy debate. If the choice is between waiting years for new generation and connecting now with modest curtailment, many large loads may accept the latter. That makes flexibility a resource that grid operators and regulators can design around, through interconnection queues that move flexible loads faster, tariffs that price curtailment, and market rules that let large loads offer demand response. Several of those ideas are already being tested. The Duke numbers give them a quantitative basis.
The report also carries a warning. Headroom is finite, and the first flexible loads will use the cheapest of it. The progression in the report makes the point: doubling the curtailment rate from 0.25% to 0.5% adds 22 GW of capacity, and doubling it again to 1% adds 28 GW more. Each additional block of flexible load requires deeper and more frequent curtailment than the last. Flexibility buys time, which is valuable, but it does not replace the need to build.