One Fallen Power Line Exposed a Growing AI Data Center Problem

On a Wednesday morning this week, a single transmission line fell near Washington, DC — the kind of mundane grid fault that normally takes seconds to correct. Instead, the eastern United States saw voltage spikes ripple from Northern Virginia all the way to Chicago, a disturbance that took more than ten minutes to resolve and stretched nearly 1,000 miles. The cause wasn't a storm, a cyberattack, or a cascading equipment failure. It was more than 3 gigawatts of AI data centers doing exactly what their protection systems were designed to do.

Background You Need

Northern Virginia is home to the densest concentration of data centers anywhere on Earth. Loudoun County alone — sometimes called "Data Center Alley" — hosts a sprawling campus ecosystem that feeds the cloud infrastructure underlying AI chatbots, streaming platforms, and enterprise software for hundreds of millions of users. All of that infrastructure sits within the PJM Interconnection, the grid operator managing electricity for 67 million customers across a region stretching from the mid-Atlantic through Illinois.

For years, grid planners have tracked data centers as a predictable, stable class of load. Unlike residential demand — which spikes on hot afternoons and dips overnight — data centers historically ran at near-constant draw, 24 hours a day. That made them, in grid terms, easy to plan around. What planners did not adequately anticipate was the behavioral change that comes with AI workloads, which are far more volatile than traditional cloud computing, and with the sheer density of interconnected facilities responding to the same signal at the same millisecond.

The problem has been building for at least two years. A similar event occurred in 2024, when a comparable voltage fault triggered a coordinated disconnection — but that event involved roughly 1.5 gigawatts of load. The 2026 incident was more than twice as large at 3.1 gigawatts. The direction of travel here is not subtle: data centers currently represent approximately 6% of PJM's total demand. Projections cited by TechCrunch put that figure at 24% by 2040. The arithmetic of what happens to grid stability when a quarter of regional load can vanish in thirty seconds is not reassuring.

What's New

This week's event has now put a face on an abstract engineering concern. According to data collected by Ting Labs — a startup that embeds IoT sensors into residential electrical outlets to monitor grid behavior — voltage spiked across the entire PJM footprint within seconds of the line fault, with the effects detectable well into the Midwest. Dominion Energy, the utility serving much of the Northern Virginia corridor, confirmed to Reuters that the data centers' internal protection systems responded as intended: they sensed the voltage anomaly and automatically switched to on-site backup generators, dropping their grid draw nearly simultaneously.

That synchronized response is the crux of the problem. Each data center made a locally rational decision. Their protection systems detected an abnormal condition and disconnected to protect equipment. No single facility did anything wrong. But when hundreds of facilities execute the same algorithm within the same thirty-second window, the collective behavior looks, from the grid's perspective, like a sudden loss of demand on a massive scale. Supply surges. Voltage spikes. Stabilization equipment across a thousand-mile corridor scrambles to compensate.

Two remedies are now being discussed with genuine urgency. The first is behavioral: grid operators want neighboring data centers to stagger their disconnection and reconnection sequences rather than acting simultaneously. If facilities with similar protection thresholds could be persuaded — or required — to wait even a few seconds before switching over, the aggregate demand drop would spread across time rather than collapsing into a single pulse. The second approach is architectural. A startup called ON.Energy is commercializing what amounts to a grid buffer: large battery installations that wrap an entire data center campus, presenting the grid with a single flat, stable load regardless of what the compute equipment inside is doing. The battery absorbs the volatility. When an AI training job ramps from idle to full draw, or when protection systems want to disconnect, the grid sees none of it. ON.Energy currently has 3 gigawatts of this buffering capacity under installation across four campuses.

The regulatory picture is uneven. ERCOT, the Texas grid operator, has moved to require large loads to "ride through" minor grid disturbances rather than disconnecting — essentially mandating that data centers tolerate brief voltage anomalies rather than switching to backup power at the first sign of trouble. PJM, which governs the Northern Virginia facilities at the center of this incident, has not yet issued equivalent requirements. That gap matters because PJM territory is where the problem is most acute.

Our read is that the 2× escalation from 2024 to 2026 is the most important number in this story. It suggests the synchronized-disconnect problem isn't a quirk of a particular facility configuration — it's a structural feature of dense AI data center clusters, and it gets worse as those clusters grow. A 24% grid share by 2040 isn't a distant scenario. At the current pace of data center construction in Northern Virginia, meaningful portions of that trajectory will be locked in within the next few years of permitting and build decisions.

The Pushback

The staggered-disconnect proposal faces an obvious practical obstacle: data center operators compete with each other, and coordinating their protection systems requires either voluntary agreement or regulatory compulsion. Voluntary coordination among competitors on shared infrastructure is historically slow. Regulatory compulsion from PJM would require rulemaking that takes years, while the data center build-out continues monthly.

The battery-buffering approach from ON.Energy is technically cleaner but raises its own questions. Three gigawatts across four campuses is significant, but it represents a fraction of the Northern Virginia footprint. Scaling campus-level battery buffers to cover the full density of the region would require substantial capital investment from operators who have historically pushed grid-stability costs onto utilities and ratepayers rather than absorbing them directly. One analysis from the Yahoo News coverage disputes the implicit framing that this is primarily a technology problem — the argument being that the grid's fragility to coordinated demand swings is as much a function of regulatory inaction as engineering, and that without PJM mandates comparable to ERCOT's ride-through rules, the voluntary adoption of buffering hardware will lag far behind the growth of the problem it's meant to solve.

Sources