Flexible Data Center Loads and Demand Response

Most data centers are designed to draw steady power around the clock. A flexible data center is one that can reduce its draw from the grid on request, for a defined period, without breaking its service commitments. The reduction can come from three places: the IT load itself, the facility systems that support it, or on-site resources that replace grid power for a few hours.

Duke’s researchers describe two kinds of IT flexibility. Temporal flexibility means rescheduling tasks such as AI model training to ride through a short curtailment. Spatial flexibility means shifting computation to data centers in regions that are not under stress.¹ Inference that serves live users is harder to move than a training job that can checkpoint and pause.

Flexibility differs from firm versus non-firm service, which is a contract term about what the grid owes you. Flexibility is a capability. A site needs the capability before it can sign the contract; see firm vs. interruptible power for the service side and the glossary entry on curtailment.

Published in February 2025 by Duke University’s Nicholas Institute for Energy, Environment and Sustainability, “Rethinking Load Growth” asked how much new load the existing U.S. system could absorb if that load were willing to curtail briefly during the hours when the system is tightest.¹⁴ It introduced the term “curtailment-enabled headroom” for that quantity.¹

The study looked at the 22 largest balancing authority areas, which together serve about 95% of U.S. load.¹ At an average annual curtailment rate of 0.25% of maximum uptime, they could accommodate up to 76 GW of new load, about 10% of the nation’s aggregate peak demand; at 0.5% the figure was 98 GW, and at 1% it was 126 GW.¹

Two details make the result more practical than it first sounds. The average curtailment event lasted about two hours, and nearly 90% of the hours in which curtailment was needed retained at least half of the new load, so the typical event is a partial reduction, not a shutdown.⁴ The authors note the implied curtailment time is comparable to existing demand response programs.⁴

The first commercial agreements tie flexibility to specific facilities. In 2025 Google announced agreements with Indiana Michigan Power and the Tennessee Valley Authority under which it would reschedule or pause non-urgent machine learning workloads when the grid is strained.⁵⁶ It described these as its first demand response arrangements aimed specifically at machine learning work, building on an earlier pilot with Omaha Public Power District in which it reduced ML-related demand during three grid events.⁵⁶

Google said the benefit is that large loads can be interconnected more quickly and the need for new transmission and power plants reduced.⁵ That is the core bargain: the utility gets a load it can count on to drop at peak, and the customer gets an earlier or larger connection than a fully firm study would allow.

Grid operators are building the same bargain into tariffs. SPP’s non-firm CHILLS service lets large loads connect on available capacity for up to seven years while firm service is secured, and PJM’s Board has directed a “connect and manage” framework in which new large loads that do not bring their own generation are curtailed before pre-emergency demand response.³⁹ Pacific Gas and Electric’s Flex Connect program offers faster distribution connections in exchange for curtailment agreements while upgrades are built.¹⁰

Utilities need to trust that a load will drop when called. The Electric Power Research Institute launched its DCFlex initiative in October 2024 to test this, with founding members including Google, Meta, NVIDIA, ERCOT, PJM, Duke Energy, Southern Company and several data center developers.¹¹ By February 2026 DCFlex counted nine field demonstration sites, ranging from compute flexibility tests in Chicago to geographic load shifting between Ashburn, Virginia and Chicago, and a Manassas, Virginia campus under construction with flexibility designed in.¹²

The best-documented result so far comes from Phoenix, where Emerald AI, Oracle Cloud Infrastructure, NVIDIA and Salt River Project ran a test through DCFlex. The AI cluster ramped down over 15 minutes, held a 25% power reduction for three hours and ramped back without exceeding its original baseline, using workload orchestration and techniques such as GPU frequency scaling rather than batteries or new hardware.⁷

A single test on one cluster is not a fleet-wide commitment. Developers selling flexibility to a utility should expect to show telemetry, a control interface and a track record, and to accept penalties if the load does not respond.

Texas has the most developed market path for flexible loads. In ERCOT, a Controllable Load Resource (CLR) is a load resource capable of controllably reducing or increasing consumption under ERCOT dispatch, receiving new base points through security-constrained economic dispatch every five minutes.²⁸ Qualifying large loads can participate in ancillary services as CLRs or non-controllable load resources, and ERCOT has observed large-load participation in Responsive Reserve, Non-Spin and ERCOT Contingency Reserve Service.⁸

Registration takes work. A CLR registers with ERCOT, acts through a qualified scheduling entity, follows dispatch instructions and must submit primary frequency response test results.⁸ Loads that can drop almost instantly, such as cryptocurrency mining, fit this model most easily; AI and cloud data centers with strict uptime commitments find it harder.

ERCOT also has an economic signal. Transmission charges for many large customers are based on their demand during the four 15-minute intervals of system peak in June through September (the “4CP”), so reducing load during those intervals cuts the next year’s transmission costs.¹³ In March 2026 the PUCT sought comments on a draft report on transmission cost recovery in ERCOT, and SB 6 requires it to amend its rules by December 31, 2026, so this signal may change.¹⁴

Separately, Texas Senate Bill 6 requires new non-critical large loads of 75 MW or more to support remote disconnection in emergencies and directs new demand-response services.¹⁵ See Texas Senate Bill 6 and ERCOT site selection.

Flexibility does not create transmission. What it changes is the planning question. A firm load must be served at its full peak through the worst studied contingencies, so the planner sizes upgrades to that hour. A load that agrees to curtail in those few hours can be served from capacity that already exists in the other hours, and connected before upgrades are complete.¹³

• The flexible portion still needs a physical connection: a tap, substation and transformer capacity on site.
• Curtailment terms bind operations for years and may affect customer contracts and financing.
• Running backup engines to cover curtailment is limited under federal air rules to 100 non-emergency hours a year, with up to 50 for qualifying demand response.¹⁶
• If many loads near one substation are flexible, their combined behavior has to be studied, not each in isolation.

Our guides to on-site generation and bridge power and microgrids and battery storage cover the equipment that makes flexibility possible without touching compute.

1. 01Decide early how much of the planned load could curtail, for how long and with what notice, and put it in the ramp schedule.
2. 02Ask the utility or RTO whether it offers conditional, non-firm or flexible connection, and how flexibility changes the study result.
3. 03Size on-site batteries or permitted generation to the curtailment you are offering, and check the air permit.
4. 04Reserve land for that equipment and any fuel supply in the site layout.
5. 05Plan telemetry and controls that the utility can verify.

A parcel that looks too constrained for firm service may still work with flexibility, and flexibility may be the difference between a five-year wait and an earlier start. If you want to know how a specific parcel stands, get a site reviewed or start with our tools.