You begin to notice them if you drive through some parts of northern Virginia at dusk. Fenced off, humming, long, windowless buildings sit low against the tree line. There are no logos, no signs, and occasionally not even a clear address.
Only gravel lots, chain-link, and the gentle mechanical sound of cooling systems operating nonstop. These areas of land were soybean fields ten years ago. These days, they are the engines of a technology that most people engage with via a chat window without ever considering the grumbling grid behind it.
| Subject Profile | Details |
|---|---|
| Topic | AI data center expansion and its pressure on U.S. electricity grids |
| Primary Region Affected | PJM Interconnection territory — 13 states, roughly 67 million customers |
| Estimated Demand (2023) | 176 TWh, about 4.4% of U.S. electricity consumption |
| Projected Demand by 2028 | Between 325 and 580 TWh, or 6.7%–12% of national use |
| Notable Incident | July 2024 voltage fluctuation in northern Virginia disconnected 60 data centers, creating a 1,500 MW surplus |
| Policy Flashpoint | Texas Senate Bill 6, plus PJM’s contested capacity auctions |
| Bill Increase This Summer | Over 20% in parts of PJM territory |
| Major Players | Microsoft, Amazon, Alphabet, Meta, OpenAI |
| Core Tension | Hyperscaler demand vs. residential affordability and grid reliability |
The numbers are beginning to match the landscape. The U.S. data center electricity demand is predicted by Lawrence Berkeley National Laboratory to increase from approximately 176 terawatt hours in 2023 to between 325 and 580 terawatt hours by 2028. With that higher number, data centers would account for more than 12% of the country’s consumption. This is the type of jump that utilities typically plan for over several generations rather than just three or four years. Furthermore, it’s possible that the actual figure is higher because it’s unclear how quickly OpenAI, Anthropic, Meta, and the others will continue to grow.
The country has essentially run out of headroom, according to Wood Mackenzie’s Ben Hertz-Shargel. He claimed that there is a land grab going on, with hyperscalers believing that the next ten years of AI will go to the person who locks down the most compute capacity. When OpenAI warned of a “electron gap” and declared that electrons are the new oil in their pitch to the White House last year, you could hear that anxiety. It sounds like marketing until you look at the PJM auction results.
The biggest grid in the nation is operated by PJM Interconnection, which extends from Illinois to Tennessee and New Jersey. It is essential to about 67 million people. Prices increased by over 800% at its annual capacity auction last year. Customers in some areas of that territory may see electricity bills this summer that are more than 20% higher than they were last year. Pennsylvania’s governor has threatened to withdraw the state from PJM completely if the operator doesn’t connect new plants more quickly. The CEO is leaving. There was a vote to remove two board members. It feels like a slow institutional disintegration, the kind that occurs when a system built for predictable load unexpectedly encounters an unforeseen circumstance.
Something from July of last year keeps coming back to me. Sixty data centers disconnected from the grid at the same time due to a voltage fluctuation in northern Virginia. The outcome was a 1,500 megawatt surplus, an uncommon issue that required operators to work quickly to prevent a cascade in the surrounding network. It was the opposite of a blackout and exposed something unsettling: the actions of a small number of server farms owned by a small number of businesses now shape the grid in real time. Areas tremble when they sneeze.
The fact that all of this has happened so quietly is startling. After years of luring data centers with tax breaks and reduced tariffs, state and local governments are now witnessing an increase in their citizens’ expenses. In an effort to regain some control, Texas passed Senate Bill 6. Other states are keeping an eye on it. Investors continue to claim that massive new generation projects will solve the problem, but ratepayers may be left footing the bill for plants that no one ends up needing if AI demand even slightly declines. If a calculation goes wrong, it’s still unclear who will be responsible.
You get the impression that the nation is taking a big risk without fully acknowledging it as you watch this play out. The demand is new, the grid is outdated, and the individuals who write checks for both typically don’t have the same zip code. The decisions being made in rooms most of us will never see will determine whether this becomes a managed transition or a slow-motion battle between hyperscalers and homeowners. Behind the fence, the hum continues in both directions.
