Fallen Power Line Exposes a New AI Data Center Risk for the Grid

A fallen power line near Washington, D.C., this week triggered a chain reaction across the PJM grid, briefly knocking more than 3 gigawatts of data center load offline and causing lights to flicker across a wide region. The incident matters because it exposed how rapidly clustered AI and cloud facilities can destabilize a power system when many of them respond to the same disturbance at the same time.

What should have been a brief grid disturbance lasted more than 10 minutes as data centers in Northern Virginia automatically switched to backup power and then disconnected almost in unison. The episode did not produce a blackout, but it showed utilities, data center operators and regulators that the electricity system is now contending with a new kind of concentrated, highly automated demand.

For years, the electric grid was designed around predictable industrial loads and conventional generation. Today, it must also manage sprawling campuses packed with servers, chillers and power electronics that can move from full demand to backup systems in a matter of seconds. As AI infrastructure expands, that challenge is becoming more urgent.

What happened on PJM’s grid?

When the transmission line failed this week, grid conditions changed quickly enough to trigger protective systems at a large number of nearby data centers. Rather than staying online and waiting out the disturbance, the facilities switched to backup power and stopped drawing electricity from the grid nearly at once.

According to data cited from PJM, about 3.1 gigawatts of load disappeared in roughly 30 seconds, with additional demand dropping shortly afterward. At the peak of the event, the grid was left carrying an extra 3.49 gigawatts of electricity before it rebalanced. Reuters reported that the disconnected data centers represented about 3% of total PJM demand at the time.

The result was not a blackout, but it was enough to make lights flicker from Northern Virginia through parts of the wider PJM territory, which stretches from New Jersey to Illinois and serves 67 million customers. The system eventually stabilized, but not before the disturbance lingered for more than 11 minutes.

“It’s the canary in the coal mine,” said Ricardo de Azevedo, chief technology officer at ON.Energy, describing the event as part of a broader pattern of increasingly frequent disturbances involving very large electric loads.

Why do data centers create such a big grid problem?

Data centers create a big grid problem because they are unusually large, unusually concentrated and usually fast to react. A cluster of facilities that each sees the same voltage dip can behave like a single giant load, all deciding within seconds to leave the grid.

The electric system depends on near-perfect balance. Supply and demand must stay aligned almost continuously, and even relatively small imbalances can create voltage swings. The grid can absorb modest fluctuations, but if many facilities disconnect at once, the sudden loss of load can create a sharp oversupply that makes voltage rise and forces more equipment to react.

That is exactly what appeared to happen in Northern Virginia. Once the voltage disturbance reached a critical threshold, multiple facilities made the same decision almost simultaneously. Instead of helping the system recover, their collective response amplified the initial problem.

How the feedback loop develops

The sequence is simple in concept, but dangerous in practice:

1. A transmission fault or local power disturbance causes a brief voltage dip.
2. Nearby data centers detect the anomaly and transfer to backup systems.
3. As one campus disconnects, the overall demand on the grid falls.
4. More generation suddenly exceeds load, pushing voltage upward.
5. Additional systems react, and the disturbance spreads.

That chain reaction can occur in seconds. Because modern facilities rely on automated controls, they do not wait for human operators to assess the event. They respond immediately, which is usually a feature, but in tightly clustered regions it can become a vulnerability.

Why Northern Virginia matters so much

Northern Virginia matters because it is home to the world’s highest concentration of data centers and sits inside PJM’s footprint. The region has become a crucial digital infrastructure hub, serving cloud computing, enterprise workloads and the fast-growing AI buildout.

That concentration creates scale benefits for developers and customers, but it also means many facilities are exposed to the same grid conditions. If one local event affects dozens of campuses at once, the electrical consequences can spread far beyond the immediate fault location.

This week’s episode also echoed a similar event on PJM’s grid two years ago. That earlier disturbance involved far less load, but it pointed in the same direction: as data centers multiply, the grid increasingly has to deal with large blocks of demand that can vanish or reappear almost instantaneously.

How big is the risk for future AI growth?

The risk is growing because data center demand is rising faster than many grid operators can adapt. PJM already manages the largest electric system in the United States, and data centers are becoming a bigger share of that load every year.

Synapse Energy Economics has estimated that data centers accounted for about 6% of PJM load in 2024. By 2040, that share could climb to 24% if current trends continue. If that forecast proves anywhere close to accurate, events like this week’s will no longer be unusual edge cases. They will be a regular operating challenge.

The scale of the problem is not only about total electricity use. It is also about geography and behavior. A large share of new demand is appearing in a relatively small number of regions, and much of it is controlled by sophisticated power management systems that can all react in the same direction at the same time.

How this week compares with the 2024 incident

This week’s disruption was roughly twice as large as the 2024 event that saw 60 data centers disconnect at once. That earlier incident removed about 1.5 gigawatts from the grid. The new event suggests the system is already seeing bigger bursts of concentrated load shedding, and the trend could worsen as AI clusters get larger.

There is an important lesson in the comparison: the grid is not just dealing with more electricity demand. It is dealing with more synchronized demand, where dozens of facilities may respond to the same signal almost simultaneously.

What can grid operators do about it?

Grid operators can require large loads to behave more like grid participants and less like passive consumers. One proposed path is to make data centers “ride through” disturbances instead of disconnecting immediately.

That means facilities would stay online through short-lived fluctuations, just as many generators are expected to remain connected during certain disturbances. A ride-through requirement could reduce the chance that a single fault becomes a cascading demand collapse.

Ali Zain Banatwala, a senior market models specialist at the Independent Electricity System Operator, said the key is to find ways for neighboring data centers to disconnect or reconnect in sequence rather than all at once, so operators have time to manage the system more deliberately.

A more orderly response would give utilities a chance to prepare specific procedures for large data center clusters, rather than reacting to a single massive drop in load after the fact.

Why sequencing matters

Sequencing matters because a few seconds can determine whether a voltage disturbance passes harmlessly or turns into a system-wide problem. If one campus rides through while another transfers to backup, the grid sees a smaller change. If all of them move together, the disturbance is magnified.

• Staggered disconnections reduce sudden demand loss.
• Sequential reconnections prevent a new surge from overloading the grid.
• Coordinated controls help operators maintain voltage stability.
• Planning for local clustering becomes more important as campuses grow.

How ON.Energy says it can help

ON.Energy is trying to solve the problem from the facility side rather than the utility side. The company says it is building an uninterruptible power supply at the scale of an entire data center campus, not just for servers but also for chillers and other supporting equipment.

In practical terms, the system places batteries and power conversion equipment between the campus and the grid. That makes the data center look like a steady, predictable load instead of a collection of volatile subsystems. The grid sees one smooth electrical profile rather than a series of sudden spikes and drops.

The design has two major benefits. First, it helps data centers ramp workloads up or down without creating trouble for the grid. Second, it allows the campus to absorb disturbances rather than instantly abandoning the grid connection.

When the grid has extra power, ON.Energy says its batteries can charge. When the grid sags, the system can send power back to the facility and keep operations running. The company says its equipment can respond within milliseconds, which is fast enough to counter the kinds of voltage changes that triggered this week’s event.

De Azevedo said the goal is to make data centers “well-behaved” from the grid’s perspective, so the utility sees a stable customer rather than a volatile one.

What is the company building now?

ON.Energy says it is currently installing systems totaling 3 gigawatts across four data center campuses. If completed as planned, that would make it one of the more ambitious attempts to wrap large-scale digital infrastructure in utility-style buffering technology.

That figure is significant not only because of the total size, but because it suggests battery-backed grid smoothing is moving from theory to deployment. The question now is whether enough operators will adopt similar systems quickly enough to match the pace of AI buildout.

Why regulators are paying attention

Regulators are paying attention because the PJM event made a technical issue visible to a much wider audience. The flickering lights were a public sign of a deeper infrastructure problem: the grid is now heavily exposed to the behavior of massive digital campuses clustered in the same geography.

ERCOT, the Texas grid operator, is already moving in this direction. According to de Azevedo, ERCOT plans to require large loads such as data centers to ride through disruptions rather than disconnecting immediately. That approach reflects a broader realization that the grid cannot be allowed to treat huge AI campuses like ordinary commercial customers.

The industry is likely to face more rulemaking, more interconnection scrutiny and more pressure to demonstrate how new loads will behave during faults. The days of simply asking whether a grid has enough megawatts may be ending. Operators now also want to know how those megawatts will act in a crisis.

What this means for AI infrastructure

This week’s incident is a reminder that the AI boom is not only a computing story. It is also a power systems story. Every new training cluster, inference campus and cloud expansion plan depends on a grid that can serve immense demand while remaining stable through faults and weather events.

As AI models grow larger and data center campuses become more concentrated, the burden on local utilities and regional grid operators will rise. The problem is not simply supplying energy; it is integrating huge loads in a way that does not destabilize the network when something goes wrong.

That makes electrical design part of the AI infrastructure conversation. Battery banks, power conditioning systems, ride-through controls and sequencing logic are becoming as important as chips, cooling systems and fiber links. In the near future, the reliability of AI may depend as much on grid engineering as on model architecture.

Why the clock is ticking

The clock is ticking because the next incident could be bigger. This week’s mass disconnection involved about twice as much load as the comparable event in 2024, which suggests the scale of grid stress is already increasing. As more campuses come online, the margin for error shrinks.

Without changes in how data centers connect, disconnect and absorb faults, a routine line problem could one day create a far more serious regional disturbance. Utilities know it. Operators know it. The question is whether construction, equipment upgrades and regulations can keep pace.

Timeline of the PJM disturbance

The bottom line

The PJM event was not a blackout, but it was a warning. A single fallen power line revealed how quickly clustered AI and cloud data centers can amplify a routine grid disturbance into a regional stability problem.

As data centers keep growing in number and size, the grid will need better ride-through requirements, smarter control systems and more robust campus-scale battery solutions. Otherwise, future faults may be harder to absorb, more visible to the public and far more expensive to fix.

The latest incident suggests the AI infrastructure buildout is entering a new phase: one in which power quality, fault tolerance and grid behavior matter just as much as computing capacity.