The Missing Fine Print in OpenAI’s Pledge to Power Down

On November 2, OpenAI announced a national program under which its future computing campuses would cut the power they draw from the grid when the grid is under strain. The company said it would work out agreements region by region with utilities and regulators, that the amount, speed and duration of each reduction would be settled locally, and that it would disclose further agreements once they were final. It also described on-site generation, storage and batteries as part of the toolkit, depending on local needs.

The example at the center of the announcement was not new. Georgia Power said on July 22 that OpenAI had agreed “to provide up to 1,000 MW of flexible demand response” at a planned campus in Effingham County, under a 25-year agreement to serve a facility expected to need about 3,200 megawatts, and that “Georgia Power will have the ability to reduce energy delivered to the OpenAI facility during certain times” to support grid stability. Commission staff approved the contract in August, and the utility said on August 26 that the arrangement would let it reduce energy to the site “during periods of high demand” and build fewer new power plants.

The 1,000 megawatts is the size of the commitment as the utility announced it: the most the campus can be asked to give up. It is not a record of reductions made.

The underlying bargain is familiar to electricity regulators. A Federal Energy Regulatory Commission staff report published in February 2011 found that 248 utilities and other entities offered “interruptible/curtailable rates” in its 2008 survey, and 265 in its 2010 survey. Such rates, it explained, give customers a price incentive, and “the customer, rather than the sponsor, curtails the load during peak times when directed.” Typical participants were nonresidential customers—factories and other businesses, not households.

What Texas did in 2025 was to put comparable terms in a statute. Senate Bill 6, which covers the grid run by the Electric Reliability Council of Texas, sets a default threshold of 75 megawatts for new or expanded large loads. For most loads connecting at transmission voltage after December 31, 2025, it requires utilities to develop a protocol, “including the installation of any necessary equipment or technology,” so that the load can be cut off during firm load shed—the forced outages ordered in a grid emergency. Separately, it directs the grid operator to buy voluntary reductions from customers of at least 75 megawatts, with “at least a 24-hour notice,” and to keep them curtailed for the duration of the emergency alert or until they can safely return. The two mechanisms work differently, but in Texas both are written where anyone can read them.

Against that background, OpenAI’s program is an attempt to carry an established form of demand response to a new kind of customer, at the size of a single AI campus. Done as described, it could help: a site able to shed nearly a third of its expected demand when the grid is tight could reduce the need for new plants. Georgia Power projects that revenue from its new large-load customers as a group will ease rates for others; that is a forecast, not a result.

What determines how much the promise is worth, though, is missing from the announcements reviewed here. Georgia Power’s two announcements say when the utility may reduce power—“during certain times,” “during periods of high demand”—without saying how many times a year, for how many hours, with how much notice, or what happens if the campus falls short. When commission staff approved the deal in August, the contract was not public; under safeguards agreed with staff, the utility promised a public summary and semiannual reports on its large data-center customers, according to The Current, a Georgia news outlet. I could not find that summary. A local advocacy group, Georgia Conservation Voters, welcomed the safeguards with a condition: “But only if we have a governing majority on the Commission willing to enforce its provisions.”

Without those terms, a 1,000-megawatt ceiling is hard to value. A commitment used for a few hours on the hottest afternoons and one that can be called for days during a long cold snap would do very different work for the people who share the grid, and the public descriptions do not say which this is.

There is also a question that an AI company is better placed to answer than any utility filing. Drawing less from the grid need not mean computing less: OpenAI points to on-site power and storage. But how it would combine its own generation with adjustments to its workload, and, if work has to slow, which work slows first—a training run that can resume tomorrow, or the responses people and businesses are waiting for—the announcement does not say.

The semiannual reports Georgia Power has promised are the place to look. If they show how often the utility called on the campus, for how long and whether the campus delivered, the Georgia agreement will be the kind of bargain ratepayers and regulators can check. If they do not, the most important numbers in OpenAI’s announcement will remain the ones it did not give.