Months Before OpenAI Announced Its $20B Georgia AI Campus, Georgia Power Was Already Planning for It

Months before OpenAI unveiled Project Camellia, its $20 billion, 3.2-gigawatt AI campus in Effingham County, Georgia, the state's largest utility had already begun planning its grid around an unnamed project of almost identical size. Georgia Power's Q1 2026 Large Load Economic Development Report logged a 3,200-megawatt new-customer commitment in April 2026. In the same window, a separate Public Service Commission planning attachment listed an anonymized 3,210 MW project in "Technical Review," with an initial in-service date of Q2 2028 and a ramp to roughly 3.2 GW by 2031. OpenAI announced a 3.2 GW campus in Effingham County in July 2026, with phased energization starting in 2028 and a full ramp to about 3.2 GW by 2031 — matching the profile of the project already in Georgia Power's planning documents. The match is circumstantial, but the numbers line up almost exactly.

The match is not, however, a confirmation. PSC customer identities are confidential by rule. The 3,210 MW project in the docket is not explicitly identified as OpenAI. What the documents show is a project of the same scale, the same in-service year, and the same ramp curve as the campus OpenAI later made public, which is itself a story about how AI megacampuses now enter utility planning.

Georgia Power's announcement of the agreement, and OpenAI's own Effingham County page, confirm the headline terms: a 25-year service deal under which Georgia Power supplies the campus and OpenAI hands the utility up to 1,000 MW of flexible demand response the utility can pull during grid-stress events. That flexibility is unusual at gigawatt scale. Most large customers buy power and expect it to be there; OpenAI is selling the option to be told no.

Both the Q1 Large Load Economic Development Report and the docket 226607 attachment were public filings months before the Data Center Knowledge review walked through them. A reader who knew where to look could have seen the 3,210 MW entry and the 2028 in-service date, and waited.

"A single 3.2-gigawatt load is extraordinary," said Neil Osnato, founder of Persistence Analytics Group, in commentary on the planning significance. Osnato is not confirming the customer; he is commenting on the scale, and on what a single-site load of that size does to a regional planning cycle. In context, 3.2 GW is more than the peak demand of several major U.S. cities combined.

These anonymized filings decide how much new generation and transmission gets built, who pays for it in rates, and which communities host the new lines. When a 3.2 GW campus enters that pipeline under a confidential project name, the planning is already bending around it before ratepayers or local officials can weigh in. The records are already public. The question is whether anyone is reading them.

The next AI megacampus filing will probably look the same way: a name, a megawatt figure, a 2028 or 2029 in-service date, a 2031 or 2032 ramp. The Q1-to-press-release gap is no longer a one-off. It is the new normal for how the largest AI loads enter the grid.