In the early months of 2024, a small team at Georgia Power began noticing something unusual in the utility's interconnection queue: requests for new large-load customers were arriving faster than the planners could process them. Most of the requests came with no name attached — only a square-footage estimate, a target energization date, and a load profile that, at the high end, would draw more power than a small city.

Eighteen months later the queue has grown sixfold. By the utility's own most recent integrated resource plan, the data-center pipeline now represents more than 8.5 gigawatts of committed and prospective new load by the early 2030s — roughly a fifth of the state's current peak demand. Almost none of it existed on paper two years ago.

The math that broke

For most of the last decade, American utilities planned around a roughly flat demand curve. Efficiency gains in lighting and appliances offset the electrification of vehicles and home heating. The 2030 forecast looked a great deal like the 2020 forecast, only greener.

That assumption has collapsed across the Sun Belt. The collapse has been most visible in Northern Virginia, where Loudoun County's data-center alley now consumes more power than several mid-Atlantic states. But the second wave — the wave that matters for the politics of the next four years — is unfolding in Georgia, the Carolinas, and central Texas, where land is cheaper, tax incentives are richer, and water for cooling, in most of the year, is still abundant.

"We used to plan around what people would do. Now we plan around what models will do."

— A senior planning engineer at a Southeastern utility, speaking on the condition of anonymity

Why nuclear, and why now

The arithmetic of AI compute is what makes nuclear suddenly viable again. Hyperscale operators want power that is both carbon-free — to honor public commitments to their largest customers — and dispatchable, meaning available at 3 a.m. when wind isn't blowing and the sun isn't up. Solar plus storage can meet a portion of the load, but at the scale a single training campus now requires, intermittent generation alone does not pencil.

Vogtle Units 3 and 4, the long-delayed reactors east of Augusta, came online in 2023 and 2024 at a combined cost north of $30 billion. Two years ago they were widely cited as the cautionary tale that would end new American nuclear construction. Today they are cited as proof that the next round can be built.

The political weather

The Georgia Public Service Commission is now reviewing four separate filings related to the data-center load — a special tariff class, a generation procurement plan, a transmission cost-allocation rule, and a contested case on whether the cost of new generation built to serve the hyperscalers should be borne by their customers alone or socialized across the ratepayer base.

That last question is the one that will shape American energy policy through the end of the decade. The utilities, broadly, want the costs socialized. The hyperscalers, broadly, are willing to pay — but they want the certainty of long-term contracts in exchange. Consumer advocates want neither: they want the hyperscalers to build their own generation.

Whichever way the commission moves, the answer will be copied in Raleigh, in Austin, in Columbus, and in Phoenix within a year. What looks like a Georgia regulatory matter is, in fact, the first national test of who pays for the AI build-out.

About the author

Elena Vance

Tech Editor based in San Francisco. Covers AI infrastructure and the people building it.