Data Centers Are Creating Grid Instability No One Planned For

The Cascade That Shouldn't Have Happened

A minor fault on a transmission line outside Washington, DC, should have been routine. Grid operators deal with downed lines constantly, and modern systems typically recover in seconds. Instead, the incident this week took more than 10 minutes to stabilize and sent voltage spikes rippling through eight states.

The culprit was not the line itself but what happened next. Within 30 seconds of the initial fault, roughly 3.1 gigawatts of data center load vanished from the PJM Interconnection grid. More facilities dropped offline shortly after, pushing the surplus to 3.49 gigawatts at its peak. The grid, suddenly flooded with power it had nowhere to send, experienced voltage surges that caused lights to flicker from Northern Virginia to Chicago.

No blackout occurred, but the event revealed something more troubling: the grid is increasingly vulnerable to the synchronized behavior of large compute facilities, and the problem is accelerating.

A Perfectly Rational Decision, Multiplied by Hundreds

Data centers are designed to protect their workloads above all else. When sensors detect voltage irregularities, uninterruptible power supplies kick in and the facility seamlessly switches to backup generators or batteries. The transition happens in milliseconds, and from the perspective of any individual operator, it is the correct decision.

The issue arises when hundreds of facilities, clustered in the same region and connected to the same grid, all make that decision within seconds of each other. Northern Virginia hosts the highest concentration of data centers on the planet, and PJM Interconnection, which manages the grid from New Jersey to Illinois and serves 67 million customers, has become ground zero for this new class of grid event.

Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator, points out that the facilities responded almost simultaneously because they all experienced the same voltage dip at roughly the same time. "We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect," he explained. Without coordination, each facility acts independently, and the aggregate effect can destabilize the very grid they depend on.

The Math Is Getting Worse

This week's event was not an isolated incident. Two years ago, a similar disruption on PJM's grid caused 60 data centers to disconnect simultaneously, pulling 1.5 gigawatts offline. The latest event was more than twice as large.

At the time of this week's fault, the disconnected load represented roughly 3 percent of total demand on PJM. That may sound manageable, but electrical grids operate on razor-thin tolerances. Supply and demand must remain in near-perfect balance; even small mismatches cause voltage to sag or spike, and larger ones trigger automatic disconnections that can cascade into broader failures.

The trajectory is concerning. According to Synapse Energy Economics, data centers accounted for about 6 percent of PJM's load in 2024. By 2040, that figure is projected to reach 24 percent. If current behavior patterns hold, the grid could face synchronized disconnections measured in tens of gigawatts, dwarfing this week's disruption.

The Sensor Network That Caught It

The voltage spike was documented in granular detail by Ting Labs, a startup that operates an IoT sensor network embedded in residential electrical outlets across the United States. The company's distributed sensors captured voltage fluctuations in real time, providing a view of the event that traditional grid monitoring infrastructure often misses.

Ricardo de Azevedo, CTO at ON.Energy, called the incident "the canary in the coal mine." Events involving large, synchronized loads are "happening more and more," he noted, and Northern Virginia's concentration of compute capacity makes it an early warning system for what other regions will face as AI infrastructure expands.

Two Paths Forward

Grid operators and data center developers are beginning to grapple with the problem, though solutions remain nascent. The most straightforward approach involves coordination: staggering disconnection and reconnection sequences so that facilities do not all act at once. This would require data centers to communicate with grid operators in real time and accept some degree of external control over their failsafe systems, a significant shift for an industry built on autonomy and uptime guarantees.

The alternative is to build data centers that can absorb grid disturbances rather than flee from them. ON.Energy has developed an uninterruptible power supply system designed to cover entire data center campuses, including not just servers but also cooling infrastructure and auxiliary equipment. The system places a large battery bank and power conversion hardware between the grid and the facility, presenting the grid with a single, stable load regardless of what is happening inside the data center.

When the grid experiences a voltage surge, ON.Energy's system absorbs the excess power by charging its batteries. When voltage dips, the system dispatches stored energy to keep servers running. The response time is measured in milliseconds, fast enough to prevent the facility from disconnecting and contributing to a cascade.

The company is currently installing 3 gigawatts of capacity across four data center campuses, according to de Azevedo. The system also allows operators to ramp AI training workloads up and down without creating corresponding spikes in grid demand, smoothing out the peaks and valleys that make large compute facilities difficult to integrate into regional grids.

Regulators Are Starting to Notice

Some grid operators are moving toward mandatory "ride-through" requirements for large loads. ERCOT, which manages the Texas grid, is expected to implement rules requiring data centers to remain connected during minor disturbances rather than automatically switching to backup power. The policy would force facilities to install hardware capable of tolerating voltage fluctuations, effectively mandating the kind of technology that ON.Energy and others are developing.

PJM has not announced similar measures, but this week's event is likely to accelerate internal discussions. The grid operator has been dealing with the rapid buildout of data center capacity in Northern Virginia for years, and the region's dominance in the sector shows no sign of slowing. Hyperscalers continue to lease and build facilities at a pace that outstrips grid planning cycles, leaving operators to retrofit solutions onto infrastructure that was never designed for this kind of load behavior.

The Clock Is Running

The problem is not hypothetical, and it is not distant. Data center capacity is growing at double-digit rates annually, driven primarily by AI training and inference workloads that require far more power per rack than traditional cloud computing. Northern Virginia remains the epicenter, but other regions are following similar trajectories as developers seek land, power, and connectivity.

If the industry and grid operators do not implement coordination mechanisms or ride-through technology soon, the next event could be significantly larger. A 3.5-gigawatt disruption caused voltage spikes and flickering lights. A 10-gigawatt event, or a 20-gigawatt event, could trigger cascading failures that take portions of the grid offline entirely.

At DailyTechWire, we have tracked the collision between AI infrastructure growth and legacy grid architecture for the past two years. This week's incident in PJM's territory is the clearest signal yet that the collision is no longer theoretical. The grid is being reshaped by compute demand, and the institutions responsible for keeping the lights on are scrambling to catch up.