OpenAI's Georgia grid-flexibility test runs into Ember's load-shifting maths

OpenAI tests grid flexibility in Georgia

The program is designed to cut a data center's power draw as local demand rises. OpenAI says the program keeps the lights on for everyone else. It also lets its campuses connect to the grid sooner instead of waiting for new power sources.

“Whether it’s heating homes during a northern cold snap or running air conditioning through a southern heat wave, families need reliable electricity,” OpenAI said in a blog post. “Our goal is to make campus flexibility a resource that utilities can call on when communities need it most”. The company added that it wants flexibility to become “a core part of how we build and operate future campuses”. That means designing facilities to cut electricity use on request. It also means agreeing with utilities on how much demand can be reduced, how quickly and for how long.

Georgia is the test bed. The flexible block in the Georgia Power agreement is enough to power around 800,000 households, nearly 20% of the state's homes, according to Tom's Hardware.

Demand response itself is decades old. Grid operators once phoned industrial customers such as steel mills and asked them to cut consumption during peaks. Automation now lets data centers trim load in under a minute. Nvidia has been running a similar experiment in the UK.

Why load shifting cannot carry the grid alone

An Ember analysis published in October 2026, and reported by pv magazine on 3 November, puts numbers on the limits. It covers nine markets across the United States, Europe and Southeast Asia.

The constraint is the workload mix. Cloud services, enterprise workloads and AI inference need low latency. Ember estimates they will account for about 87% of global data center capacity by 2030. AI training can be interrupted or moved with fewer consequences. Under its assumptions, about 6.5% of global capacity could theoretically shift. That assumes 35% flexibility for training, 5% for inference and none for traditional data centers. An average utilization rate of roughly 48% then cuts the capacity actually available during grid stress to around 3%.

Ireland shows the scale of the result. Load shifting could trim peak demand there by about 100 MW by 2030, or 1.5% of the projected peak. The other markets Ember studied land between 0.1% and 0.6%.

Keeping workloads running while swapping grid power for on-site generation or storage does better. Ember calls this bring-your-own-capacity. Ember finds it could deliver nearly five times as much grid relief as load shifting. For two large Texas data centers with a combined projected load of 5.6 GW, shifting would cut about 0.55 GW at peak. On-site generation could cut up to 2.7 GW.

What utilities and regulators will watch

Ireland is the stress case. Data centers already took about 19% of peak demand in January 2025, when system demand peaked at 6 GW and they used an estimated 1.1 GW. Ember projects that share at 24% by 2030, with data center demand at 1.6 GW against a 6.5 GW system peak.

Timing makes the gap harder. A large data center needs 200 MW to 500 MW of capacity and takes 18 to 24 months to build. Conventional grid expansion takes four to 10 years.

For regulators, the open question is verification. A demand-response promise is only worth the metering and controls behind it. OpenAI has not said how the 1,000-megawatt block will be measured or checked.

What is still unknown

OpenAI has committed to one campus. The company said arrangements will differ by location. It will share additional commitments as it finalizes them.

Ember's study is a model rather than a track record, and its nine markets include Ireland, the UK and Germany.

None of that settles the commercial question. A data center that promises to shed load gets a faster connection. The value of that promise depends on how much capacity is actually flexible, and on who checks.