OpenAI offers to turn down 1,000 megawatts when Georgia's grid is strained

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 2 November OpenAI announced Grid Protectors. Its first agreement, with Georgia Power for a campus in Effingham County, includes up to 1,000 MW of flexible demand response. OpenAI compares that to the average demand of about 700,000 households. It has committed to pay the full infrastructure and electric-service costs of the campus. Georgia Power projects about $950 million a year in customer savings from 2029, or at least $15 a month for a household using 1,000 kWh. OpenAI cites projections that US data-centre capacity will pass 100 GW by 2030, and adds a caveat of its own: "Flexibility cannot replace investment in new energy supply and transmission."

Why it matters

Paying a large customer to cut its load at peak times is an old utility tool. What is new is applying it to AI compute, where cutting load means pausing training or inference. The value of the promise depends on terms OpenAI has not published: how many megawatts, at what notice, for how long, and with what penalty if the campus does not respond.

A promise to switch off is only as good as the hours, the notice and the penalty written beside it.

In Issue 01 I argued that tokens per megawatt is the metric behind the AI business. In my view this is the next step: megawatts you can give back become part of the price of getting connected. For lenders and investors, power terms are now part of the credit story, alongside the chip and financing commitments in this week's other pieces.

What I'd watch

• Whether the Georgia agreement's response terms (notice, duration and penalties) are published in regulatory filings.
• Whether other AI labs and cloud providers offer comparable flexibility commitments.
• Whether the projected household savings from 2029 hold up once the campus is running.