Data center demand response: How AI campuses can cut power use when the grid is strained

Data center demand response is an agreement in which a large computing campus promises to use less electricity, or use it at different times, when the power grid is under strain. In return, the campus can often connect sooner and the utility can avoid building as much new capacity. The idea is to make a giant, steady customer behave more like a flexible one.

A seasonal restaurant and a hot afternoon

Picture a hot summer evening. Millions of air conditioners run at once, and the grid operator is close to its limit. Normally it would fire up an expensive backup power plant. Under demand response, it can instead ask a big customer to turn something down for a few hours.

Factories and aluminum smelters have done this for decades. A data center is a newer candidate, because AI campuses can draw hundreds of megawatts (a megawatt, MW, is a million watts) and utilities worry about serving them.

That worry is behind a recent announcement. In OpenAI’s Grid Protectors Program, the company says its planned Effingham County campus in Georgia includes up to 1,000 MW of flexible demand response, and that further sites will follow the same approach.

Four ways to shrink a data center’s draw

A data center cannot simply be unplugged. Its power use depends on the work it is doing, so there are several levers, which can be combined.

• Shift the work in time. Jobs that nobody is waiting on, such as batch processing or model training, can be paused or rescheduled to a calmer hour.
• Shift the work in place. A company with several campuses can move some tasks to a region where power is plentiful.
• Use stored or on-site power. Batteries or generators can supply the campus while it draws less from the grid. OpenAI says it may also coordinate with grid operators to feed power back from generators or storage, and may use batteries and control systems to soften swings in its load.
• Slow the chips down. Processors can be set to run at a lower power level, which stretches a job out but lowers the draw.

Which levers a campus gets depends on its equipment and its local grid. OpenAI says exactly this: the size, speed and duration of each demand cut will differ by location.

From video processing to machine learning jobs

Google has been the most public about the practice. In an August 2025 post, Michael Terrell, the company’s head of advanced energy, described an early approach that moved non-urgent tasks, such as video processing, to periods when the grid was not stressed. Partners named in that post include Centrica Energy, Elia in Belgium and Taiwan Power Company.

The same post said Google had cut power demand from machine learning workloads during three grid events in 2024, in a pilot with Omaha Public Power District. It then announced agreements with Indiana Michigan Power, covering a new Fort Wayne, Indiana data center, and with the Tennessee Valley Authority. Google described these as the first to target machine learning workloads.

The post does not explain the technical method in detail, so how Google schedules or throttles those jobs is not public from that source.

Why grid planners care about a few hours a year

The research case rests on a simple observation: grids are built for rare peaks. For most of the year, much of that capacity sits unused. If a new customer agrees to step aside during the few tightest hours, the grid may be able to serve it without new plants.

A February 2025 study from Duke University’s Nicholas Institute, “Rethinking Load Growth” by Norris, Profeta, Patino-Echeverri and Cowie-Haskell, put numbers on this. It analyzed the 22 largest US balancing areas, which cover about 95% of US load. It estimated that about 76 gigawatts (GW) of new load, roughly 10% of US peak demand, could be added if that load accepted curtailment for about 0.25% of the time. Allowing 0.5% and 1% raised the figures to about 98 GW and 126 GW.

These are planning estimates for the whole system, not a guarantee that any given site can connect. A commentator on Medium argued the study is being misused in interconnection debates, and noted that cutting service for 0.5% of the year still falls well short of the uptime many commercial cloud contracts promise. That is an opinion piece, but it points to a real tension.

Who pays, and what the Georgia pledge does not cover

Flexibility is also a cost argument. OpenAI says families and small businesses should not subsidize a compute campus’s electricity needs, and that it has committed to pay the full infrastructure and electric-service costs for the Effingham project. It cites Georgia Power figures: about $950 million in yearly customer savings from 2029 across the utility’s large-load agreements, and monthly relief of $15 or more for an average home using 1,000 kWh. These are the utility’s projections, not independent findings.

The numbers also have limits. The 1,000 MW cut is under a third of the 3.2 GW power deal reported by DatacenterDynamics for the site, so the campus would still draw most of its contracted power in an emergency unless the final terms say otherwise. OpenAI also calls Effingham a pilot and names no other sites, dates or standard terms.

Limits worth watching

Not every AI task can wait. Google’s own post says flexibility is constrained because services like Search, Maps and Cloud for essential sectors such as healthcare need high reliability. Real-time chatbot answers fall in the same category, so flexible capacity is likely to come from training and other deferrable jobs.

Other open questions:

• Contract terms. How many hours, how much notice, and what penalty applies if a campus fails to cut.
• Verification. Whether regulators and utilities can confirm that promised cuts actually happen during a real emergency.
• Hardware wear and cost. Idle or throttled chips still cost money, so operators weigh the lost compute against the benefit of a faster connection.
• Standardization. OpenAI says further commitments will follow once terms are settled; whether common templates emerge will show how widely the model spreads.

For now, demand response is a promising tool that has been shown at pilot scale. The test is whether large campuses can deliver cuts on demand, repeatedly, without the customers they serve noticing.