A power line fault in the eastern United States this week, while not causing a blackout, brought to the forefront the power supply problems of AI data centers. Following the incident, numerous data centers in northern Virginia switched to backup power almost simultaneously, removing several gigawatts of load from the PJM grid in a short period, causing significant voltage fluctuations in the region.
Remove 3.1 gigawatts of load within 30 seconds
TechCrunch, citing data from PJM, reported that approximately 3.1 gigawatts of load disappeared within 30 seconds after the outage. More load subsequently dropped, resulting in an additional 3.49 gigawatts of power on the grid at one point. The system took approximately 11 minutes to stabilize.
Reuters reported that the disconnected data center loads accounted for approximately 3% of PJM's total demand at the time. While this percentage may seem small, the power grid relies on a real-time balance between supply and demand, and a sudden change of just a few percentage points can trigger voltage spikes or drops.
Northern Virginia becomes a high-risk area
The impact of this fluctuation extended from Northern Virginia to Chicago, one of the world's most densely populated data center areas and also within the service area of PJM, the largest power grid operator in the United States. PJM covers a region from New Jersey to Illinois, serving approximately 67 million customers.
The problem is that data centers typically take protective measures in a very short time. Once a voltage drop is detected, they quickly switch to backup power. This is standard safety design for a single campus, but when a large number of facilities act synchronously within seconds, it amplifies what would otherwise be a localized power supply disturbance into a much larger demand-side fluctuation.
The industry is trying to smooth out shocks with batteries.
To mitigate the impact of such mass shutdowns, some companies are pushing to connect entire data center campuses to larger-scale uninterruptible power supply (UPS) systems. The report mentions that ON.Energy is developing a campus-wide solution that covers not only servers but also critical facilities such as cooling equipment.
The core approach of this type of system is to add batteries and power conversion equipment between the data center and the power grid. This way, the power grid sees a more stable load, rather than the spikes caused by the simultaneous fluctuations of server training tasks, cooling equipment, and backup systems.
If the power grid experiences a brief period of excess power, the system can charge the battery; if the power supply experiences a temporary drop, the battery can immediately replenish it. The article states that this approach can also support AI training load scaling while reducing transient disturbances to the power grid.
Additional information:The report mentions a similar incident in 2024, when approximately 60 data centers simultaneously went offline, removing 1.5 gigawatts of load. According to Synapse Energy Economics, data center load as a percentage of PJM's total load could rise from 6% in 2024 to 24% in 2040.
