Data center curtailment and flexible load: faster grid access

Flexible load means a customer agrees to use less power, or switch to its own generation, during the few hours when the grid is stretched. In return it may connect sooner. A Duke study found 76 GW of new load could fit if curtailed 0.25 percent of the time, though it did not model transmission limits.

This post explains what curtailment means for data centers, the evidence for and against the idea, and how federal, Texas and PJM proposals are turning it into rules. Facts are as of 19 October 2026.

Key numbers at a glance

• 76 GW of new load could be integrated at an average annual curtailment rate of 0.25 percent, about 10 percent of current aggregate US peak demand (Duke University Nicholas Institute, "Rethinking Load Growth," February 2025, as reported by POWER magazine).
• 98 GW at 0.5 percent curtailment and 126 GW at 1.0 percent (same source).
• Average curtailment event length of about 1.7 hours at 0.25 percent, 2.1 hours at 0.5 percent and 2.5 hours at 1.0 percent (same source).
• About 85, 177 and 366 hours a year of curtailment at the 0.25, 0.5 and 1.0 percent rates (Utility Dive report on the Duke study, republished by Facilities Dive, 11 February 2025).
• The study covered 22 of the largest US balancing authorities, serving about 95 percent of national peak load (same source).
• At 0.5 percent, the largest potential gains were PJM 18 GW, MISO 15 GW, ERCOT 10 GW, SPP 10 GW and Southern Company 8 GW (same source).
• NERC's 2024 assessment projected winter peak load growing 21.5 percent, from 694 GW in 2024 to 843 GW in 2034 (cited by POWER magazine).
• 60 days: the expedited study time on which DOE sought comment for curtailable large loads (DOE letter, 23 October 2025).
• Loads of 75 MW or more: the size at which Texas must procure demand reductions, with curtailed loads staying off for the full emergency event (Texas Senate Bill 6).
• At least 50 percent of on site demand: the backup generation capability that Texas SB 6 ties to ERCOT's ability to direct deployment (same bill text, as summarized from its provision on backup generation).
• About 70 GW: PJM's forecast of new large load demand by 2038 (PJM Board decisional letter of 27 July 2026, as summarized by Babst Calland).
• 1 June 2027: the date PJM's proposed Interim Resource Adequacy Service would apply to new large loads lacking enough capacity (same summary).

Two meanings of curtailment

The word is used two ways, and mixing them up causes confusion.

Generation curtailment is when a grid operator tells a solar or wind plant to produce less because the lines are full or supply exceeds demand. This is not what this post is about.

Load curtailment is when a customer reduces its electricity use when asked, or when the grid is short of supply. Factories have done it for decades through demand response programs, in which they are paid to cut use on request. Data centers are now being asked to do the same.

For a data center, curtailment can mean several things. The facility might delay flexible computing jobs such as model training runs, move work to another site, run on its own batteries or generators, or reduce cooling and non critical load. Which of these is possible depends on the customer's service promises. A hospital's cloud provider cannot accept the same outages as a research cluster.

The Duke study: how much room is there

In February 2025 the Nicholas Institute at Duke University published "Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems." The figures in this section come from two press accounts of the report: POWER magazine's article of the same month and a Utility Dive report by Ethan Howland, republished by Facilities Dive on 11 February 2025. The study asks a question: how much new load could the US grid host if the new load accepts being curtailed for a small part of the year?

The answer depends on how small. For 22 of the largest balancing authorities, serving about 95 percent of peak load, the study found 76 GW at an average annual curtailment of 0.25 percent, 98 GW at 0.5 percent and 126 GW at 1.0 percent. The 76 GW figure is about 10 percent of current aggregate peak demand. Typical events were short, about 1.7 hours at the lowest rate and 2.5 hours at the highest, and half of the new load would stay on during about 90 percent of curtailment events. POWER said the curtailment hours were comparable to existing US demand response programs.

The reasoning is that the grid is built for the hottest day, so most of the year has unused room. If new customers will step aside on the rare peak hours, they can be added without building for those hours. That is the headroom the study counts.

There are important limits, and POWER lists them:

• The study does not account for transmission constraints. An analyst quoted in the article argued that congestion could wipe out much of the potential.
• It assumes the loads can be curtailed safely for a few hours a year, using on site generation, storage or shifted operations.
• It proposes factoring load flexibility into interconnection and planning studies early, since existing demand response usually treats large loads as firm for planning.

Facilities Dive adds two more limits from the report: the calculations were based on peak demand levels, and they did not count reserve margin capacity, which the authors note could increase the headroom. So the caveats cut both ways. Transmission limits could shrink the number, while unused reserve margin could raise it.

So the study is best read as a measure of generation capacity headroom, not as a promise that any given site can connect. A local substation or line can still be the limit.

What a data center would have to accept

The study's percentages are an average annual curtailment rate, and they should not be read as a simple share of the 8,760 hours in a year. That distinction matters. Facilities Dive reports that the curtailment rates translate into about 85 hours a year at 0.25 percent, 177 hours at 0.5 percent and 366 hours at 1.0 percent. The hours are higher than the percentages alone suggest because most events cut only part of the load: according to the same report, during 88 percent of curtailment time at least half of the new load keeps running.

So a data center accepting the middle case would need a plan for roughly a week of partial reductions spread across a year, mostly in short events of about two hours. A facility that promises customers very high availability needs batteries, on site generation or the ability to move work elsewhere for those hours, which is why backup equipment comes up so often in these proposals.

The trade is speed for certainty. A firm connection, one that is never curtailed, may need major grid upgrades and a long wait. A flexible one may be offered sooner or without some upgrades, but the operator can cut it. Contracts spell out how many hours, how much notice, and what compensation applies.

What federal rules propose

DOE's 23 October 2025 letter to FERC includes a principle that the interconnection study of large loads that agree to be curtailable, and of hybrid facilities that agree to be curtailable and dispatchable, should be expedited. The operator's ability to control them must be enough to fold them into operations and planning. DOE asked whether such studies could be completed in 60 days, and whether to do it serially or another way.

FERC's June 2026 show cause orders took up the theme. According to Vinson and Elkins, each grid operator should consider interim network service and firm or non firm contract demand services for flexible loads, using automated demand response, real time telemetry and communication protocols, and load management plans covering demand response and curtailment protocols. Troutman Pepper Locke lists flexible large loads as the first of the tariff areas in the orders. Those orders asked for responses by 17 August 2026. A consolidated account of what each region filed had not been published in the sources cited here as of 19 October 2026. The companion FERC explainer covers the orders in detail.

Compute Law Blog also reports that SPP's conditional high impact large load service, called CHILLS, offers seven years of non firm transmission, under a FERC order dated 5 June 2026. That is a bridge: connect early on a non firm basis while upgrades are built. This is the blog's account of the order, and the order's own text is the place to check its exact terms.

What Texas requires

Texas turned flexibility into law with Senate Bill 6, passed in the 2025 regular session of the 89th Legislature. According to the enrolled text, for transmission voltage loads connected after 31 December 2025, the load must have a protocol and equipment allowing curtailment during firm load shed, with exemptions for critical industrial loads and designated critical natural gas facilities. Firm load shed is the emergency step in which the operator orders power cut to protect the whole grid.

The bill also directs ERCOT to competitively procure demand reductions from large loads of at least 75 MW for anticipated emergencies. Curtailed loads must stay curtailed for at least the notice period and the full event, and price responsive loads are excluded from this service. And interconnected large loads must disclose on site backup generation that cannot export to the grid, which ERCOT may direct to deploy, after market services are used, subject to a threshold set in advance. The bill states that emissions limits are not overridden.

The Keentel Engineering newsletter of July 2026 adds that ERCOT has filed rules for controllable load resources and for bring your own generation, though it did not say they were final. The companion Texas explainer covers the statute and ERCOT's batch process.

What PJM is proposing

PJM serves a large part of the eastern US and has seen capacity prices climb. Per Babst Calland's summary, the PJM Board issued a decisional letter on 27 July 2026 with a package that includes:

• A mandatory Large Load Registry of customers, including data centers, whose electricity may be curtailed during system stress.
• A Reliability Backstop Procurement, an additional capacity auction, to start in September 2026.
• An Interim Resource Adequacy Service for new large loads that, as of 1 June 2027, lack sufficient capacity and cannot be served at a specified reliability standard.
• Compensation for large loads when directed to reduce consumption, similar to demand response.

The same summary reports about 70 GW of new large load demand forecast by 2038, about 15 GW of generation retired since 2022, and a backstop aimed at a shortfall of about 60 GW. It also says the Board rejected a stakeholder proposal that had received a supermajority because it did not adequately ensure the near term shortfall would be procured, and that state action would be needed to implement the program. A headline on RTO Insider's PJM page reported that FERC conditionally approved the backstop procurement but delayed its implementation, around 30 September 2026. Only the headline was available, so the conditions FERC attached are not described here; the FERC order itself is the authoritative record.

In plain terms, PJM's direction is that data centers that do not bring their own supply could be curtailed in grid stress. This is a proposal in progress, and its final terms are not settled in the sources cited here.

What one company has already agreed

Flexibility is not only a regulatory idea. In August 2025 Google announced what Data Center Frontier, in a 15 September 2025 article, described as the company's first formal demand response agreements with utilities: one with Indiana Michigan Power, a unit of American Electric Power, and one with the Tennessee Valley Authority. Under them, Google can reduce the power used by machine learning workloads at some of its data centers when the utility asks during peak demand.

The same article reports an earlier test with Omaha Public Power District, where Google cut machine learning power demand during three grid events in the prior year. It also records the limits Google set. Demand flexibility "will only be implemented at certain locations, and only where service reliability won't be compromised," and services such as Search, Maps and those used by healthcare or cloud customers are not interrupted. Training runs can sometimes be paused or rescheduled; latency sensitive work cannot.

That is a useful real example of the split the proposals assume. Some AI work, such as batch training, can move in time. Work that answers live users usually cannot. The agreements did not disclose financial terms in the sources cited here.

The arguments on each side

For flexibility: it uses capacity that already exists, it can bring new load online years sooner than new supply, and it keeps costs from landing on other customers. It also rewards data centers that invest in batteries, on site generation and workload shifting.

Against, or at least cautions: transmission and local limits can still block a connection. Service promises to cloud customers constrain how much can be curtailed. Operators need reliable telemetry and control, and enforcing curtailment on a customer that refuses is a legal and technical question. And some load, such as AI inference serving live users, is harder to shift than training, a distinction Google's own agreements draw. These cautions follow from how the proposals are written rather than from a single study that measures them.

What is firm and what is uncertain

Firm:

• Duke's study found 76, 98 and 126 GW at 0.25, 0.5 and 1.0 percent curtailment, with a stated transmission caveat.
• DOE proposed expedited study for curtailable loads and asked about 60 days.
• Texas SB 6 requires curtailment capability for new transmission loads and a demand reduction service for loads of 75 MW or more.
• PJM's Board issued a decisional letter on 27 July 2026 proposing a registry, a backstop and compensation for curtailment.

Uncertain:

• How much of the Duke headroom survives transmission limits.
• What FERC will require nationally.
• PJM's final terms and timing, and the status of the backstop after FERC's reported conditional approval.
• How much flexibility data centers can truly offer without breaking service promises.

What to watch next

• FERC's rulings on the regional filings for flexible service.
• ERCOT's rules for controllable load resources and the first demand reduction procurement.
• PJM's registry, and whether curtailment compensation is set at a meaningful level.
• Whether more data center operators follow Google with published demand response agreements, and whether those agreements disclose hours and compensation.

For related context, see what a data center needs to connect to the grid and the generation interconnection queue.

Safeguard scans dependencies and builds SBOMs for the software behind infrastructure projects, and the free scan is at https://gold.safeguard.sh/scan.

Quick check

Sources

• POWER magazine, "Duke researchers: grid flexibility key to accommodate load growth," February 2025: https://www.powermag.com/duke-researchers-grid-flexibility-key-to-accommodate-load-growth
• US Department of Energy, Secretary of Energy letter and proposed rulemaking, 23 October 2025: https://www.energy.gov/sites/default/files/2025-10/403%20Large%20Loads%20Letter.pdf
• Vinson and Elkins on FERC's June 2026 show cause orders: https://www.velaw.com/insights/ferc-institutes-six-simultaneous-section-206-proceedings-targeting-large-load-interconnection-across-all-rto-iso-markets/
• Troutman Pepper Locke on the June 2026 orders: https://www.troutmanenergyreport.com/2026/07/ferc-orders-regional-grid-operators-to-justify-or-reform-tariff-provisions-for-interconnection-of-large-loads/
• Texas Legislature, Senate Bill 6 enrolled text: https://capitol.texas.gov/tlodocs/89R/billtext/html/SB00006F.htm
• Babst Calland, "PJM's proposal to FERC targets data centers and other large load customers": https://www.babstcalland.com/news-article/pjms-proposal-to-ferc-targets-data-centers-and-other-large-load-customers/
• RTO Insider, PJM large load page (headline only): https://www.rtoinsider.com/?p=127412
• Compute Law Blog, "How AI data centers connect to the grid": https://computelaw.blog/power/how-ai-data-centers-connect-to-the-grid-ferc-power-queue/
• Facilities Dive, Utility Dive report by Ethan Howland on the Duke study, 11 February 2025: https://www.facilitiesdive.com/news/us-grid-headroom-flexible-load-data-center-ai-ev-duke-report/739938/
• Data Center Frontier, "Google partners with utilities to ease AI data center grid strain," 15 September 2025: https://www.datacenterfrontier.com/energy/article/55316600/google-partners-with-utilities-to-ease-ai-data-center-grid-strain
• Keentel Engineering, ERCOT Energy Market Update, July 2026: https://keentelengineering.com/newsletter/ercot-energy-market-update-july-2026

Last reviewed: 19 October 2026.