A Fallen Power Line Exposed a New Grid Risk From AI Data Centers

Lead

A fallen power line would normally be a short-lived disturbance for a large power grid. This time, the recovery took more than 10 minutes. The reason was not simply the failed line itself, but the reaction that followed: data centers in Northern Virginia switched to backup power nearly simultaneously, removing more than 3 gigawatts of load from PJM’s system in a very short window.

The event did not cause a blackout, but it pushed voltages higher across a wide area, with flickering lights reported from Northern Virginia toward Chicago. For an AI industry racing to build ever larger computing campuses, the message is clear: data centers are no longer just large customers on the grid. When concentrated in one region, they can become dynamic actors that shape how grid disturbances unfold.

Key points

• A small fault became a large demand shock. PJM data cited in the report showed about 3.1 gigawatts of load disappearing in roughly 30 seconds. At the peak, the grid carried an extra 3.49 gigawatts of electricity. In a power system that depends on near-instant balance between supply and demand, even a few percent shift can matter.
• The core issue was simultaneous protection behavior. Individual data centers are designed to protect equipment by moving quickly to backup power when voltage dips. That is rational for one site. But when many nearby facilities make the same decision within seconds, the grid sees a sudden collapse in demand, which can turn an initial supply disruption into a voltage surge.
• Northern Virginia is the stress test. The region hosts the world’s densest concentration of data centers and sits inside PJM territory. A similar PJM event occurred in 2024, when 60 data centers disconnected at once and removed 1.5 gigawatts of load. This week’s event was roughly twice as large.
• Better coordination is needed. Experts argue that large loads should disconnect and reconnect sequentially, or be engineered to ride through disturbances. The goal is to prevent clusters of facilities from behaving like one enormous, abrupt switch.

Why it matters

AI training and inference are increasing the size and density of data center loads. TechCrunch notes that data centers accounted for about 6% of PJM load in an earlier estimate and are expected to reach 24% by 2040. If their response to grid trouble remains abrupt and synchronized, the issue will grow alongside the compute buildout.

One proposed fix is to place a full data center campus behind batteries and advanced power conversion equipment. ON.Energy says its approach makes the facility appear to the grid as a steady, well-behaved load. When there is excess power, batteries can absorb it; when supply dips, they can support servers, chillers, and other equipment without forcing an immediate disconnect.

Grid operators are also moving toward stricter expectations. ERCOT, for example, is expected to require large loads such as data centers to ride through disturbances. That points to a broader shift: the next phase of AI infrastructure will not be judged only by GPUs, power purchase agreements, and land availability. It will also be judged by whether massive compute campuses can coexist with the grid without turning minor faults into regional stress events.

Source: TechCrunch AI