A transmission-line fault outside Washington, DC this week did not cause a blackout. But it did expose a bigger infrastructure problem now tied to the AI buildout: large clusters of data centers can respond to ordinary grid disturbances in ways that make the disturbance harder to manage.

According to reporting from TechCrunch, the event triggered a wave of near-simultaneous switching to backup power across data centers in Northern Virginia, causing about 3.1 gigawatts of load to disappear from the PJM grid in roughly 30 seconds. PJM data cited by TechCrunch showed the imbalance later reached 3.49 gigawatts, and the grid took about 11 minutes to stabilize. Reuters, as cited in the same report, said the disconnected facilities represented around 3% of PJM demand at the time.

That matters well beyond one region. Northern Virginia is the world’s largest concentration of data centers, and much of the current buildout is being justified by demand for AI training and inference. The lesson from this week’s incident is that the sector’s power problem is no longer just about securing enough electricity. It is also about how AI infrastructure behaves during faults, voltage dips, and recovery periods on a stressed grid.

What happened on the PJM grid

The basic sequence, as described by TechCrunch using PJM data and outside expert commentary, was straightforward. A fallen power line created a grid disturbance. Under normal conditions, the broader system would likely recover quickly. Instead, many nearby data centers appear to have interpreted the voltage dip as a trigger to transfer to onsite backup systems.

That response protected the individual facilities, but it abruptly removed a very large amount of demand from the grid. Because power systems must keep supply and demand in tight balance, a sudden loss of load can be just as destabilizing as a sudden loss of generation. In this case, excess supply pushed voltage higher across a wide area.

TechCrunch reported that Ting Labs, which operates an IoT sensor network through residential electrical outlets, observed voltage spikes from Northern Virginia to Chicago. The event reportedly caused visible flickering lights across parts of the region, even though it stopped short of a blackout.

The technical point is simple but important for AI builders: if many large facilities in the same area are designed to make the same split-second decision, they can behave like one enormous machine. That turns a localized disturbance into a broader grid-management problem.

Why AI data centers make the risk more urgent

This problem is not unique to AI workloads, but AI increases the stakes. Large-model training clusters and dense inference deployments are pushing operators to build bigger campuses with heavier and more concentrated power demand. That concentration can amplify the consequences when multiple sites react in lockstep.

TechCrunch cited Ricardo de Azevedo, CTO of ON.Energy, calling the incident a "canary in the coal mine" and saying similar events involving large loads are happening more often. The report also connected this week’s disturbance to a comparable PJM event in 2024, when about 60 data centers disconnected at once and roughly 1.5 gigawatts of load dropped off the system.

The growth trajectory matters here. TechCrunch cited Synapse Energy Economics saying data centers accounted for about 6% of PJM load in 2024 and are expected to reach 24% by 2040. Those figures are not specific to AI alone, but they show why AI infrastructure planning now intersects directly with grid operations. A response pattern that looked manageable at lower penetration levels becomes much more consequential as campus scale rises.

For enterprises buying AI capacity and for cloud providers selling it, this creates a new reliability variable. Compute availability is not only about GPUs, networking, or cooling anymore. It is also about whether the surrounding electrical system can absorb faults without mass load-shedding from colocated facilities.

The proposed fixes: stagger responses or make campuses ride through

The emerging fixes fall into two broad categories.

The first is operational coordination. TechCrunch quoted Ali Zain Banatwala, a senior market models specialist at the Independent Electricity System Operator, saying grid operators and facility owners need a way for large colocated loads to disconnect and reconnect sequentially rather than all at once. In practice, that would mean staged response logic, coordinated recovery windows, and grid-aware controls instead of purely site-level trigger thresholds.

The second is more architectural: design campuses to absorb short disruptions without dropping off the grid. TechCrunch highlighted ON.Energy’s approach as one example. According to the company’s description in the report, it is building a campus-scale uninterruptible power supply that covers not just servers but also chillers and other supporting equipment. The idea is to place batteries and power-conversion systems between the grid and the entire data center load so the external grid sees a steadier demand profile.

If that works as described, a disturbance would not force the facility to suddenly vanish as load. Extra incoming power could be diverted into batteries, while shortfalls could be covered by stored energy. TechCrunch reported that ON.Energy says its system can follow grid conditions within milliseconds and smooth fluctuations from AI training ramps as well as grid faults.

That design logic is increasingly relevant because AI workloads are not only large; they can also be bursty. Training jobs, inference surges, and cooling changes can create sharp internal load movements. A buffer layer between the campus and the grid could help with both self-inflicted variability and external disturbances.

Evidence, benchmarks, and what remains unverified

The strongest factual elements in this story are the event itself and the scale of the load drop cited from PJM data through TechCrunch reporting. The publication also attributed regional voltage observations to Ting Labs and contextual demand-share figures to Reuters and Synapse Energy Economics.

Several forward-looking claims, however, should be treated as company or expert assertions rather than settled industry fact. ON.Energy’s description of its system’s performance, including millisecond-level response and its ability to shield entire campuses from grid volatility, comes from the company as reported by TechCrunch. Likewise, the statement that ON.Energy is currently installing 3 gigawatts of systems across four data center campuses was attributed to de Azevedo. No independent project documentation was included in the source material provided here.

The report also said ERCOT is moving toward requiring large loads such as data centers to "ride through" disruptions, again attributed to de Azevedo. That may point to a real policy direction, but based on the available evidence, readers should treat it as a cited executive comment rather than a fully documented regulatory summary.

What is well supported is the broader pattern: concentrated data center load can worsen grid instability if facilities disconnect simultaneously, and the issue is becoming more significant as power demand from AI infrastructure rises.

What this means for builders and enterprise buyers

For AI builders, this story changes the checklist for site design. Backup generation alone is not enough if the transfer logic itself destabilizes the grid or increases the risk of regional disturbances. Developers may need to invest more in power electronics, onsite storage, dynamic controls, and utility coordination at the campus level.

For enterprise AI buyers, especially those relying on hyperscale capacity in data-center-heavy markets, power resilience is becoming part of service resilience. A region with abundant capacity but brittle fault behavior may be less attractive than one with stronger ride-through standards, even if power is more expensive.

For utilities and grid operators such as PJM and possibly ERCOT, the operational model for "large loads" is changing. Data centers used to be treated mostly as passive demand with backup systems behind the meter. AI-scale campuses increasingly act more like grid-relevant actors whose protection settings and recovery timing can influence regional stability.

This also adds pressure to the debate over where to build. Northern Virginia remains strategically important because of fiber density, cloud presence, and existing ecosystems. But if the concentration of AI data center capacity creates recurring grid-management risks, developers may face stronger incentives to spread load geographically or adopt stricter interconnection requirements.

What to watch next

First, watch for a formal incident analysis from PJM or related utility and reliability bodies. The key questions are how many facilities disconnected, what protection settings were involved, and whether reconnection behavior worsened the recovery.

Second, look for interconnection rule changes covering AI data center campuses and other large loads. If ride-through capability becomes a requirement rather than an optional design feature, that would materially affect project costs, timelines, and vendor selection.

Third, track whether Northern Virginia becomes a proving ground for campus-scale storage and power-conditioning systems from companies like ON.Energy. If these systems move from niche deployments to standard design, that would signal a new phase in enterprise AI infrastructure.

Finally, pay attention to whether this issue spreads beyond PJM. If similar events appear in other fast-growing data center regions, the market will likely treat grid-behavior controls as core infrastructure rather than an engineering afterthought.

Creati.ai perspective

The AI industry has spent the last two years talking about a coming power shortage. This week’s PJM incident suggests the more immediate problem may be power behavior. When multi-gigawatt clusters react identically to a routine fault, the result is not just a site-level resilience issue; it becomes a systems issue for the grid around them.

That has a practical consequence for anyone building or buying enterprise AI capacity. The next competitive layer in AI infrastructure may not be only cheaper GPUs or faster networking, but better electrical integration. The winners could be operators that can prove stable ride-through, controllable load ramps, and tighter coordination with utilities in places like Northern Virginia. In other words, reliable AI may increasingly depend on power engineering as much as compute engineering.