OpenAI is touting a national program to reduce electricity consumption at future data center campuses during periods of grid stress, extending an approach already covered by a 25-year agreement with Georgia Power.
The agreement gives Georgia Power the ability to curtail up to 1,000 megawatts of electricity delivered to OpenAI's planned Effingham County, Georgia, campus. The project is expected to require approximately 3,200 MW, making the potential reduction about 31% of its projected demand.
OpenAI's Grid Protectors Program calls for the company to work with utilities, regulators and policymakers on arrangements specifying how much electricity a campus can reduce, how quickly it can respond and how long it can sustain a reduction. The terms will vary by location, and OpenAI said it will announce additional commitments as agreements are finalized. This news comes as community pushback on AI data centers and their effects on local power grids has become a national political hot topic.
The Georgia agreement, announced in July and approved through the Georgia Public Service Commission's regulatory process in August, predates Monday's national announcement. The campus remains under development, with power delivery expected in phases between 2028 and 2032.
Thate agreement gives Georgia Power the ability to curtail up to 1,000 megawatts of electricity delivered to OpenAI's planned Effingham County, Georgia, campus. The project is expected to require approximately 3,200 MW, making the potential reduction about 31% of its projected demand.
A Georgia Power spokesperson told TechTarget that the agreement gives the utility the ability to reduce electricity delivered to the OpenAI facility during certain periods to support grid reliability. The demand-response resource will also allow the utility to invest in fewer new generation resources to serve future growth, the spokesperson said.
The utility declined to disclose further contract specifics, including the conditions under which it can call on the resource and how the 1,000-MW commitment will be measured and verified.
Contracted flexibility versus dependable capacity
The 1,000-MW commitment is substantial, but its value to the grid depends on how reliably the campus can deliver the reduction, said Neil Osnato, founder of Persistence Analytics Group.
Utilities need to evaluate response time, duration, availability, repeatability and independent verification before treating demand response as a dependable grid resource, Osnato said. They also need to know whether the full 1,000 MW can be curtailed simultaneously or whether the capability depends on the campus's operating conditions and equipment configuration.
“Capacity that can be called upon under defined conditions is a resource,” Osnato said. “Capacity that is merely stated in a development agreement is an assumption until its operating performance is demonstrated.”
The type of computing work affects how much electricity a data center can reduce. AI training jobs can offer flexibility when operators can pause, checkpoint or reschedule them. Inference workloads are harder to interrupt because they support active applications with latency and availability commitments, Osnato said.
Cooling, networking, storage and power-conversion equipment also have their own electricity requirements. Reducing server activity does not necessarily produce an equivalent reduction in demand at the utility meter.
Osnato also pointed to the potential for a rebound in electricity demand when operators resume workloads deferred during a curtailment event. He said utilities should measure reductions at the facility's point of interconnection and account for the recovery period.
The Georgia agreement runs for 25 years. Osnato said changes in computing equipment, workload mix and operating economics over that period make periodic performance testing, independent metering, event reporting and revalidation important.
“A data center should receive recognition for the megawatts it can demonstrably release to the grid – not simply the megawatts it promises it can release,” he said.
Communities want a say in data center development
Grid flexibility addresses some concerns about large data center electricity demand, but local debates also center on who pays for infrastructure and who has a voice in development decisions, said Tricia Martinez, founder and CEO of Dapple and a former White House Fellow at the Department of Energy.
“The conversation usually starts with electricity prices, but the underlying issue is much bigger,” Martinez said. “Communities want to understand who benefits, who pays, and who gets a say in the decisions being made.”
Martinez said developers often treat community engagement as a permitting exercise rather than part of infrastructure planning. She also questioned the industry's focus on gigawatt-scale campuses, arguing that facility size should reflect customer demand, available energy and local conditions.
Dapple advocates a distributed network of purpose-built facilities, Martinez said. That approach could involve using existing capacity, developing dedicated generation or working with utilities on infrastructure to support growth.
On OpenAI's Grid Protectors program, Martinez called the pledge constructive but said its value depends on the mechanisms behind it.
“Communities deserve to know who pays for infrastructure upgrades, how electricity costs are allocated, what happens during periods of grid stress, and whether those commitments can be independently verified,” she said.
She also cautioned that demand flexibility will not work equally well for every AI workload, particularly when facilities serve enterprise customers with long-term service commitments.
OpenAI outlines cost protections
OpenAI said it will pay the infrastructure and electric-service costs required to serve its Effingham County project. Future arrangements will include long-term contracts and financial assurances intended to protect other customers if a project is delayed, reduced or canceled, the company said.
OpenAI argues that large data centers can help spread fixed grid costs by consuming electricity throughout the year while reducing demand when the system is under pressure. More consistent use of grid infrastructure could help utilities recover fixed costs across a larger volume of electricity sales, according to the company.
OpenAI cited Georgia Power's estimate of approximately $950 million in annual customer savings beginning in 2029. That figure reflects the utility's broader portfolio of large-load agreements and expected growth, including the OpenAI project. It is not a projected saving attributable to the OpenAI contract alone.
The company also compared the project's maximum 1,000-MW demand response with the average electricity demand of roughly 700,000 households. The figure represents electricity the utility could avoid delivering to the campus under the agreement, rather than power generated by the facility and sent to the grid.
OpenAI acknowledged that demand flexibility cannot replace new electricity supply and transmission infrastructure. “Flexibility cannot replace investment in new energy supply and transmission,” the company said.
The company said the amount and duration of demand reductions at future campuses will be negotiated separately with local utilities. It has not announced additional agreements under the program.
Contracted flexibility versus dependable capacity
The 1,000-MW commitment is substantial, but its value to the grid depends on how reliably the campus can deliver the reduction, said Neil Osnato, founder of Persistence Analytics Group.
Utilities need to evaluate response time, duration, availability, repeatability and independent verification before treating demand response as a dependable grid resource, Osnato said. They also need to know whether the full 1,000 MW can be curtailed simultaneously or whether the capability depends on the campus's operating conditions and equipment configuration.
“Capacity that can be called upon under defined conditions is a resource,” Osnato said. “Capacity that is merely stated in a development agreement is an assumption until its operating performance is demonstrated.”
The type of computing work affects how much electricity a data center can reduce. AI training jobs can offer flexibility when operators can pause, checkpoint or reschedule them. Inference workloads are harder to interrupt because they support active applications with latency and availability commitments, Osnato said.
Cooling, networking, storage and power-conversion equipment also have their own electricity requirements. Reducing server activity does not necessarily produce an equivalent reduction in demand at the utility meter.
Osnato also pointed to the potential for a rebound in electricity demand when operators resume workloads deferred during a curtailment event. He said utilities should measure reductions at the facility's point of interconnection and account for the recovery period.
The Georgia agreement runs for 25 years. Osnato said changes in computing equipment, workload mix and operating economics over that period make periodic performance testing, independent metering, event reporting and revalidation important.
“A data center should receive recognition for the megawatts it can demonstrably release to the grid – not simply the megawatts it promises it can release,” he said.
