Demand response and curtailment for large AI data centers

In short

Large AI data centers can reduce their power consumption when the grid is stressed, earning revenue or faster interconnection in return. Federal rules from the Federal Energy Regulatory Commission (FERC) require regional grid operators to accept demand response bids and pay for load reductions. Actual participation by big data centers is still low. Texas now mandates curtailment for new large loads. PJM, the largest grid operator, proposes a new class of loads that can be curtailed without compensation. The key numbers include a 20 megawatt threshold in a pending federal rule, 50 MW in PJM’s proposal, 75 MW in Texas, and 1 GW of contracted demand response already in place by Google.

What is demand response and why does it matter for data centers?

Demand response means an electricity customer cuts its power use during times of high grid stress. It can be voluntary and compensated, or it can be mandatory.

Curtailment is a reduction in load ordered by a grid operator to prevent blackouts. It is usually mandatory.

AI data centers consume enormous power. In 2023 they used about 176 terawatt hours, roughly 4.4 percent of US electricity. AI data centers’ share of U.S. electricity could grow to between 9 and 17 percent by 2030. EPRI Powering Intelligence A single hyperscale campus can draw hundreds of megawatts, rivaling a small city.

Grid operators must match supply and demand every second. When a large load comes online, the grid can be strained. Allowing that load to be flexible, to reduce or shift its consumption, helps keep the grid stable. It can also avoid the need for expensive new power plants and transmission lines.

An AI data center can be flexible in several ways. It can shift computing workloads to a different time or location. It can run on battery backup instead of grid power for short periods. It can power down nonessential equipment. Some workloads, like model training, can be deferred. Others, like real-time inference, are harder to shift. The exact split between inference and training flexibility is not publicly quantified.

The payoff for a developer is two things. First, being able to curtail can speed up the interconnection process, the approval to connect to the grid. Second, offering demand response can generate revenue from wholesale electricity markets.

Key thresholds at a glance

Several rules set a megawatt size above which special curtailment or interconnection terms apply. The table below captures the main ones.

Who sets the federal rules for demand response?

FERC has authority over wholesale power sales and transmission under the Federal Power Act. 16 U.S.C. § 824(b) That authority lets FERC order regional grid operators, called RTOs or ISOs, to integrate demand response into their markets.

Three FERC orders make up the core.

First, FERC Order No. 719 in 2008 required RTOs to accept demand response bids in ancillary services markets, which are markets for quick-response power that helps keep the grid frequency stable. It also allowed third party aggregators to bundle small customer loads and bid them into wholesale markets. However, Order 719 included an opt out. A state could prohibit aggregators from bidding retail customer demand response into wholesale markets. Many states used that opt out, creating a patchwork. FERC Order No. 719, LBNL state regulatory report

Second, FERC Order No. 745 in 2011 required RTOs to pay demand response resources the same wholesale price, called the locational marginal price, that generators receive, when the demand response is cost-effective. That means reducing load is worth the same as generating power, dollar for dollar, when the numbers work. The order was upheld by the U.S. Supreme Court in 2016. FERC Order No. 745

Third, FERC Order No. 2222 in 2020 went further. It requires all RTOs to allow distributed energy resource aggregations, including demand response, to compete in wholesale energy, capacity, and ancillary services markets. The minimum aggregation size is just 100 kilowatts. Even small loads can be bundled, but a large single load can also participate through an aggregator. FERC Order No. 2222

For AI data centers, Order 2222-A eliminated the state opt-out for demand response resources that participate in heterogeneous DER aggregations, but Order 2222-B subsequently reversed that decision and reinstated the opt-out, leaving the issue to a pending Notice of Inquiry. That means a state can still block a data center from joining a heterogeneous aggregation that includes demand response, and in April 2026 FERC withdrew its inquiry on removing the opt-out, leaving it in place. FERC Order No. 2222-B, FERC Order Withdrawing Demand Response Opt-Out NOI

One critical note is that ERCOT, the grid operator for most of Texas, is not under FERC jurisdiction. Federal orders do not apply there. ERCOT’s rules come from Texas state law and the Public Utility Commission of Texas. Tex. Util. Code §39.151, ERCOT compliance

Compliance with Order 2222 has been slow. As of late 2023, only CAISO (California) had a fully approved compliance framework while ISO-NE (New England) had its third compliance filing accepted in November 2023 but its implementation remained suspended. MISO (Midcontinent) proposed compliance by late 2029, which FERC rejected as too late. SPP (Southwest) targeted Q3 2025, NYISO (New York) targeted late 2026. NARUC-NASEO overview The result is that in many regions, AI data centers cannot yet fully tap into wholesale demand response markets under the Order 2222 framework.

How do the biggest grid regions handle demand response and curtailment?

The rules differ sharply between PJM (the grid spanning the Mid-Atlantic and Midwest) and ERCOT. These two grids are where the most AI data center load is projected to grow.

PJM’s capacity market and the proposed non-capacity-backed load

PJM operates a capacity market. Generators and demand response resources commit to being available years in advance. They earn a fixed capacity payment per megawatt-day in exchange for that commitment. If they fail to deliver during a grid emergency, they face stiff penalties.

AI data center load in PJM is surging. PJM expects peak load to rise by 32 gigawatts by 2030, with AI data centers driving 94 percent of that growth. Modo Energy analysis The three most recent capacity auctions show the strain. The 2026/2027 auction cleared at $329.17 per MW-day. The 2027/2028 auction cleared at $333.44, near the price cap, and fell short of the reliability requirement by about 6,517 megawatts. That was the first auction in which the entire RTO fell short of the reliability requirement. PJM 2026/2027 BRA Report, PJM 2027/2028 BRA Report The PJM Independent Market Monitor calculated that data center load growth alone added $23.1 billion in combined capacity revenues across three recent auctions, as higher demand pushes prices toward the cap. PJM IMM 2025 State of the Market Report

But large AI data centers have rarely participated directly in PJM’s demand response programs. Hyperscale operators told PJM that the risk of interruptions far exceeds any economic value they could get under current incentives. PJM Large Load Additions Workshop PJM was seeing many new loads interconnect without any contractual obligation to curtail.

To address this, PJM proposed a new category in August 2025, Non-Capacity-Backed Load (NCBL). A facility of 50 megawatts or more that does not bring its own generation would be designated NCBL. In a grid emergency, NCBL loads would be cut off before capacity-backed demand response resources and before regular firm loads. Critically, NCBL loads would not receive capacity payments. They would be first in line to be curtailed. Modo Energy analysis The proposal excludes critical infrastructure like hospitals and 911 centers. Loads that arrange behind-the-meter generation or accredited demand response are exempt up to the amount they can self-supply. Large tech companies including Amazon, Google, and Microsoft opposed the proposal, as did some generation developers. Modo Energy analysis, Joint stakeholder alternative Whether the proposal will be filed with FERC in time for the June 2026 capacity auction remains uncertain.

Separately, the PJM Board of Managers adopted principles in January 2026 to handle large loads. They included an expedited track for loads that bring their own generation (Bring Your Own Generation, or BYOG) and better load forecasting. A large load was defined as 50 MW or more at a single interconnection point. PJM Board Decisional Letter In December 2025, FERC also found PJM’s existing tariff unjust and unreasonable for lacking standard rules for co located loads, AI data centers built next to power plants,and ordered PJM to fix it. PJM filed a compliance proposal in February 2026. FERC PJM Co-Located Load Order, Federal Register

The direction in PJM is clear. Large AI data centers should either bring their own generation or accept being curtailable without compensation, rather than being paid as traditional demand response resources.

ERCOT’s large flexible load and mandatory curtailment

ERCOT is not regulated by FERC. Its rules come from the Public Utility Commission of Texas and the Texas legislature. ERCOT already has a voluntary program for Large Flexible Loads (LFLs). Any facility drawing 75 MW or more from the grid qualifies. LFLs can bid into ERCOT’s energy and ancillary service markets. They typically curtail when wholesale electricity prices spike above about $100 per megawatt-hour, a threshold used in EIA modeling. EIA Today in Energy In a high-growth scenario, LFL facilities were curtailed for only about 10 hours in a year. EIA Today in Energy Historically, LFLs have used only about 65 percent of their approved capacity on average.

At the end of 2024, roughly 6,500 MW of LFL capacity was operational. An additional 26,500 MW had applied to become operational by the end of 2025, though not all of that may materialize soon. EIA Today in Energy

The voluntary framework changed dramatically on June 20, 2025, when Texas Senate Bill 6 was signed into law. SB 6 mandates that new transmission-voltage loads connecting to ERCOT after December 31, 2025 must be capable of curtailment during firm load shed events. Tex. S.B. 6, § 39.170(a), Utility Dive, State data center regulation survey This is not a market payment program. It is a reliability requirement. The load must reduce or go offline when told, or face penalties.

So in ERCOT, a developer can still participate in voluntary demand response markets and earn revenue. But by statute, they now face mandatory curtailment during emergencies. There is no opt out.

What is the DOE’s large-load interconnection proposal?

In October 2025, the Secretary of Energy directed FERC to start a rulemaking on interconnecting large loads to the interstate transmission system. The direction came under the Department of Energy Organization Act. DOE ANOPR The advanced notice of proposed rulemaking (ANOPR) proposed a key tradeoff. Large loads that agree to be curtailable would qualify for an expedited interconnection study process. It seeks comment on whether such studies could be completed within 60 days. FERC Docket No. RM26-4

The ANOPR defines a large load preliminarily as greater than 20 megawatts. Law firm analysis That matches FERC’s existing pro forma Large Generator Interconnection Procedures threshold, which many developers already know.

The DOE directive set an April 30, 2026 deadline for FERC final action, but on April 16, 2026 FERC announced it would instead act by the end of June 2026. FERC Whether the 60-day study timeline is realistic remains an open question. ISO New England expressed skepticism in comments, noting that interconnection studies require complex engineering analysis that may not be compressible to 60 days. CAISO deferred to its transmission owners on whether such studies can be completed in 60 days. CAISO ANOPR Comments

If finalized, this rule would offer a powerful incentive for AI data center developers to accept curtailment. Instead of waiting years in an interconnection queue, a project willing to be flexible could get through in months. OpenAI filed comments supporting the ANOPR, describing its own Grid Flexibility Initiative where a planned hybrid hub would pair AI data centers with generation and load flexibility, and would invest heavily in grid upgrades. OpenAI FERC filing

Which states impose their own curtailment mandates?

As of May 2026, the only state with a statewide law requiring mandatory curtailment for large loads is Texas, through SB 6. Utility Dive, Law firm analysis

Several states are, however, moving toward treating AI data centers as a distinct customer class with special tariff conditions. Utilities in at least five states, Ohio (AEP Ohio), Arkansas, Idaho, Montana, and Wyoming, have received regulatory approval to create separate rate classes for AI data centers. EPRI DCFlex newsletter These new rate designs can include demand charges that make it expensive to draw power during peak hours. That creates a strong financial reason to curtail, even without a mandate.

Beyond explicit mandates, many utilities are negotiating individual contracts with large AI data center customers that include load flexibility provisions. Google’s 1 GW portfolio of demand response contracts is an example, where the obligation to reduce load is bilateral, not imposed by a state law. Google Blog

How are AI data centers actually delivering demand response today?

Despite the slow uptake in formal wholesale markets, some leading operators are building practical demand response capacity.

Google announced in March 2026 that it had integrated 1 gigawatt of demand response capability into long-term contracts with five U.S. utilities, including Indiana Michigan Power, Tennessee Valley Authority, Entergy Arkansas, Minnesota Power, and DTE Energy. Google Blog The approach shifts machine learning workloads to times and locations where the grid has spare capacity, or temporarily reduces non-critical operations. Google often commits to reduce load when the grid is stressed, in exchange for utility infrastructure upgrades or preferential rates.

The DCFlex initiative, launched by the Electric Power Research Institute (EPRI) in 2024, now includes over 45 members and nine active demonstration sites. EPRI DCFlex, EPRI press release, EPRI press release Founding members include major tech companies, grid operators, and utilities. In an Arizona field test, an AI compute cluster at an Emerald AI site reduced power consumption by 25 percent over a three-hour peak event without compromising essential AI data center services. EPRI Journal

Another approach uses the AI data center’s own uninterruptible power supply (UPS) batteries. A hyperscale facility’s UPS battery array can provide 50 to 100 megawatts of short-term capacity for one to four hours. ITIF Report Meta has piloted dispatching its UPS fleets into demand response markets, demonstrating 10 to 15 megawatts of dispatchable capacity in collaboration with PJM’s DCFlex project. ITIF Report Microsoft piloted a grid interactive UPS at its Dublin AI data center, providing fast frequency response that helps stabilize the grid in under one second. Eaton-Microsoft whitepaper, Microsoft news article

There are also cautionary examples of what happens when AI data centers disconnect without coordination. In July 2024, a voltage fluctuation in Northern Virginia caused 60 data centers to disconnect simultaneously. That sudden loss of load created a 1,500 megawatt power surplus on the grid, forcing emergency actions to prevent cascading outages. Belfer Center analysis The event showed both the risk and the value of controlled, planned curtailment.

What are the main tensions and open questions?

The push for AI data center demand response intersects several big debates.

Should flexible loads be paid or just curtailed? The PJM Independent Market Monitor argues that AI data centers that do not bring their own generation should be curtailable but should not be paid as if they are providing demand response. PJM IMM 2025 State of the Market Report Others say that appropriate compensation is necessary to build the software and systems that enable safe, rapid load shedding.

Are costs shifting to other ratepayers? In seven PJM states, more than $4.3 billion in transmission costs for upgrades to serve AI data centers was passed on to consumers in 2024. Sustainable FERC Project comments If AI data centers are required to curtail before firm customers, those cost shifts may be more politically acceptable. If not, the backlash could slow development.

Does demand response help or hurt carbon emissions? A study by Knittel, Senga, and Wang (NBER Working Paper 34065, also published through MIT CEEPR, 2025) found that in carbon-intensive regions, shifting loads to off-peak hours can increase emissions because off-peak generation may be coal-heavy. NBER Working Paper 34065 The net emissions effect of AI data center flexibility is still an open question, and federal policy has not resolved it.

Can inference workloads be made flexible? Training a large AI model can be scheduled over days or weeks, making it highly deferrable. But a growing share of AI data center load is inference, responding to real time queries, which is harder to shift. The split between training and inference is not publicly known for most operators, and estimates vary.

Will FERC act on schedule? FERC has already pushed its action date from the DOE’s April 30, 2026 deadline to the end of June 2026. Further delay would postpone the incentive to accept curtailment for expedited study.

How many of the proposed loads are real? Interconnection queues across the U.S. are packed with speculative projects, making it hard to plan for actual flexible load potential. Some requests may be duplicates or placeholders, a problem FERC Order No. 2023 sought to address with higher deposits and readiness requirements.

Jurisdictional battles. The DOE ANOPR asserts FERC authority over large load interconnection under the Federal Power Act. Some states may challenge this, arguing that aspects of interconnection and retail service fall under state jurisdiction. A legal fight could create uncertainty.

Key takeaways

• AI data centers that can curtail load on demand may get faster interconnection under the pending DOE rule, which targets a 60-day study for curtailable loads.
• In PJM, a developer should expect to either bring its own generation or accept being designated as Non-Capacity-Backed Load, meaning curtailment before firm load without capacity payments.
• In Texas (ERCOT), under SB 6 (signed and effective June 20, 2025), new transmission-voltage loads over 75 MW connecting after December 31, 2025 must be capable of mandatory curtailment during grid emergencies.
• Voluntary demand response markets are open under FERC Order No. 2222, but RTO compliance is still incomplete. Only CAISO had a fully approved compliance framework as of late 2023.
• Large operators like Google have bypassed wholesale markets and contracted directly with utilities for demand response, achieving 1 GW of flexible capacity across five utilities.
• UPS batteries can provide 50 to 100 MW of short-term flexibility per site. Early DCFlex field tests show 25 percent load reduction in real world operation.
• The PJM Independent Market Monitor recommends no payment for AI data centers that do not bring generation, a stance that could shape future market rules.
• Transmission cost allocation for AI data center upgrades is a growing problem. A flexible, curtailable load can help justify the investment but may not shield a developer from local opposition.

Frequently asked questions

Q:What is the difference between demand response and curtailment?

A:Demand response is usually voluntary and compensated. A customer agrees to reduce load when asked and receives a payment or reduced rates. Curtailment is a mandatory reduction ordered by a grid operator to keep the system stable, typically without a direct payment, although it may avoid penalties.

Q:Do all AI data centers have to participate in demand response?

A:No. Participation is generally voluntary. In Texas, new loads over 75 MW must be able to curtail on command during grid emergencies, which is mandatory. In other regions, a developer can choose not to participate, but doing so may mean longer interconnection queues or less favorable utility tariffs.

Q:What are the key FERC orders to know?

A:Order No. 719 (2008) opened ancillary services markets to demand response. Order No. 745 (2011) required paying demand response the locational marginal price. Order No. 2222 (2020) required RTOs to allow DER aggregations, including demand response, with a 100 kW minimum size. The state opt-out for demand response remains in place following subsequent FERC orders.

Q:What is PJM’s NCBL proposal and why does it matter?

A:PJM proposed a Non-Capacity-Backed Load class for loads 50 MW or larger that do not bring their own generation. In a grid emergency, these loads would be curtailed before firm customers and before capacity-backed demand response resources. NCBL loads would not be paid for their curtailment. This could change the economics of an AI data center that relies solely on the grid.

Q:How does ERCOT’s mandatory curtailment work?

A:Texas Senate Bill 6, signed and effective June 20, 2025, requires new transmission-voltage loads over 75 MW connecting to ERCOT after December 31, 2025 to be able to curtail when directed during a declared grid emergency. The grid operator, ERCOT, can order the reduction. This is not a market payment program. It is a reliability obligation.

Q:Can an AI data center use its UPS batteries for demand response?

A:Yes. A typical hyperscale AI data center’s UPS battery array can supply 50 to 100 megawatts for one to four hours. Operators like Meta and Microsoft have piloted using UPS batteries for grid services, including frequency response and peak shaving. The DCFlex initiative is testing this at multiple sites.

Q:How fast can an AI data center reduce load?

A:Workload shifting can happen in minutes to hours. UPS batteries can respond in milliseconds to seconds. The DOE ANOPR considers a target of 60 days for interconnection study if the load is curtailable, but that refers to the study timeline, not the physical ramp rate. The actual ramp rate for load reduction would be defined in an interconnection agreement.

Q:What happens if an AI data center fails to curtail when ordered?

A:Consequences vary by region and contract. In ERCOT under SB 6, failure to curtail during a grid emergency could lead to penalties, disconnection, or both. In PJM, under the capacity market, failure to deliver committed demand response results in financial penalties. In bilateral contracts like Google’s, there may be financial penalties or forfeiture of grid service upgrades.

Q:Has FERC finalized the large load interconnection rule?

A:No. The rulemaking is at the ANOPR step. FERC extended the DOE’s original April 30, 2026 deadline and now expects to act by the end of June 2026. Key details, including the 60-day study timeline and the definition of curtailable, could change.

Q:Are there any tax incentives for AI data center demand response?

A:Demand response itself is not a federal tax credit. But battery storage used for grid services can qualify for the Section 48E clean electricity investment credit, which covers standalone storage placed in service after 2024. A data center that pairs demand response with on site batteries may capture that credit. Confirm the specifics with a tax professional. 26 U.S.C. § 48E

Subscribe to The Compute Law Brief

The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.

Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.