OpenAI's voluntary grid plan and the limits of shifting data center demand

OpenAI's non-binding grid promise

OpenAI launched the Grid Protectors Program, DataCenterDynamics reported. The plan is voluntary and applies to future data center campuses.

OpenAI framed the stakes with a projection: US data center capacity will exceed 100GW by 2030. The company said that growth creates an opening to build campuses that can adjust their power use when communities need it most.

"With US data center capacity projected to exceed 100GW by 2030, we have a once-in-a-generation opportunity to strengthen the electric grid by building campuses that can adjust their power use when communities need it most."

OpenAI described the effort as an "ambition" rather than a set policy. Arrangements will differ by location, the company said. It will share further commitments as it finalizes them.

The program could include helping grid operators supply power from on-site generation or storage. It could also use batteries and power controls to smooth demand swings. OpenAI said it will provide up to 1GW of flexible demand response and will work with utilities, regulators and policymakers.

OpenAI pointed to a Georgia Power deal. In that deal, OpenAI committed to pay the full cost of the infrastructure needed to serve a data center. Kim Greene, chairman, president and CEO of Georgia Power, said at the time that "data centers are paying more so families and small businesses can pay less". In January, OpenAI pledged to "pay its own way" to power its Stargate data centers.

Why load shifting delivers little

Ember's October 2026 analysis covered nine markets across the US, Europe and Southeast Asia. It separates two routes to grid relief. One is moving computing work to another time or place. The other is replacing grid electricity with on-site generation or storage.

Ireland shows the scale of the problem. Data centers there already account for about 19% of peak demand. In January 2025 they drew an estimated 1.1GW when system demand peaked at 6GW. Ember projects that share could reach 24% by 2030. Data center demand would then be 1.6GW against a 6.5GW system peak. The timing gap makes it worse. A large data center can need 200MW to 500MW and take 18 to 24 months to build. Grid expansions take four to 10 years.

Load shifting cannot close much of that gap. Ember estimates cloud services, enterprise workloads and AI inference will be about 87% of global data center capacity by 2030.

Those workloads need low latency. Pausing or relocating them is hard. Under Ember's assumptions, about 6.5% of global data center capacity could theoretically shift. After a 48% utilization rate, the capacity available to respond during grid stress falls to around 3%. In Ireland, load shifting could cut peak demand by about 100MW by 2030, or 1.5% of the projected peak. Across the other markets in the analysis, estimated reductions range from 0.1% to 0.6%.

The investment response: equipment and wires

Ember found that behind-the-meter generation and storage could deliver nearly five times as much grid relief as load shifting.

Two large Texas data centers have a combined projected load of 5.6GW. For them, load shifting could cut grid demand by about 0.55GW during periods of stress. On-site generation could cut it by up to 2.7GW.

That demand is changing valuations for equipment suppliers. GE Vernova agreed to acquire Trench Group for US$9 billion, expanding capacity in its Electrification segment. The deal brings high-voltage grid equipment manufacturing in house and adds an operational base in Europe and Asia. GE Vernova expects to close the deal in late 2027 and will fund the purchase with cash and debt.

E.ON is running a similar program in Hungary. It plans nearly HUF 230 billion of grid investment, adding more than 2,300MW of connection capacity by 2030.

EU funds cover only around a quarter of those planned grid investments, E.ON says. For every forint it receives from the EU, it adds three of its own. The company is also spending HUF 13.5 billion on 237,000 additional smart meters by the end of 2028. That brings its total to nearly 600,000. E.ON's Guntram Würzberg told the Budapest Business Journal that the challenge is now "when, where and at what cost we can use it".

What is still unquantified

OpenAI has published no binding targets. It said it will share additional commitments as it finalizes them.

Ember's figures rest on assumptions. They include 35% flexibility for AI training workloads and 5% for inference, with none for traditional data centers. Move those numbers and the relief shrinks or grows.

On the equipment side, the timelines are long. GE Vernova does not expect to close the Trench purchase until late 2027. E.ON says its financing depends on fair, reliable regulatory rules and tariffs.

Politics is already part of the picture. Data center energy use and its effect on local electricity prices have drawn opposition across the US. The issue featured in the midterm elections under way when Ember published the analysis.