A surge protector malfunctioned in a peaceful area of Fairfax County, Virginia, last July. It was located in an area known colloquially as Data Center Alley, on a 230-kilovolt transmission line that most people will never hear of. In a matter of seconds, sixty data centers immediately disconnected from the grid and switched to backup generators. Before it could turn into something much worse, grid operators at PJM and Dominion Energy hurried to rebalance an unexpected excess of electricity.
Nothing blew up. The following morning, there were no headlines. However, federal regulators now characterize it as a near-miss that revealed something that the nation’s electrical system was never intended to handle.
That is the unsettling context of the AI explosion. Chips, model size, and the company that shipped the smartest assistant this quarter are all topics of conversation. The physical reality that lies beneath it all—racks of servers pulling massive, constant loads from a grid designed for a different time period, climate, and type of demand—is rarely discussed.
The grid isn’t built to handle losing a 1,500 megawatt data center all at once, according to John Moura, who is in charge of reliability assessments at the North American Electric Reliability Corporation. That is no longer a hypothetical. Regulators are now actively preparing for it, just as they have long prepared for an unplanned power plant outage.
The industry-integrated safety system is what distinguishes data centers. Facilities are designed to automatically disconnect rather than run the risk of damaging hardware in order to shield delicate chips from voltage fluctuations. At the level of a single building, it’s a sensible design decision. It becomes more akin to a coordinated shock to the system at scale when dozens of facilities are grouped together, but no one planned it. Because each facility responds to the same signal in the same way, it simply occurs all at once.
Variations of this have already occurred in Texas. Since 2020, ERCOT has recorded over thirty near-miss incidents related to data centers and sudden disconnections of cryptocurrency miners. In one instance in 2022, a failed transformer close to a substation in the state’s western region caused nearly 400 facilities to go offline at the same time, forcing operators to shut down 112 megawatts of generation in order to maintain stability.
Sitting with the scale here is worthwhile. Today, data centers in Northern Virginia alone use about the same amount of electricity as the city of Boston. Additionally, the Lawrence Berkeley National Laboratory predicts that by 2028, total data center power consumption—which has already tripled over the previous ten years—may triple once more. Grid planners believe that facilities that can swing thousands of megawatts in either direction in a matter of seconds simply do not fit into the old playbook, which was based on predictable industrial and residential demand.
Additionally, this is not just a domestic tale. Drone attacks on Amazon Web Services locations in Bahrain and the United Arab Emirates in March were described by World Economic Forum analysts as the first instance of commercial hyperscale data centers being specifically targeted by the military. It turns out that cloud infrastructure was designed to withstand hardware malfunctions and outages rather than missiles. Even a year ago, that distinction was less significant than it is now.
Observing this develop from the outside, it’s remarkable how little public discussion there has been about it in comparison to the discussion surrounding AI capability. Forecasting has turned into a silent crisis in and of itself, underestimating demand and risking blackouts, overestimating it, and burdening ratepayers with unnecessary infrastructure. The 1983 Washington Public Power System bond default, in which bondholders held nearly worthless paper due to overconfident demand forecasts for nuclear plants, has been compared by some economists.
The industry, which places a higher priority on hardware protection than grid stability, has opposed proposals for stricter disconnection standards, so it’s still unclear whether regulators and data center operators will agree on solutions. It is evident that the discussion of AI’s future cannot remain limited to model performance indefinitely. It eventually runs directly into a transmission line outside of Fairfax or a transformer in rural Virginia, supporting far more than anyone intended.
