Loads That Leave the Grid: NERC's Reliability Insights on Large Computational Loads and Voltage Sensitivity

When a fault on the grid makes the voltage dip, some large computational loads, data centers among them, switch to their own backup power and remove themselves from the grid. The North American Electric Reliability Corporation, NERC, describes that pattern in a note headed "Reliability Insights," titled "Large Computational Load Risks Due to Voltage Sensitivity" and dated "September 2026."

Why NERC is watching these loads

The note starts from demand. Over the 10-year period of NERC's 2025 Long-Term Reliability Assessment, summer peak demand is projected to rise by more than 224 GW. Computational loads, which the note says include data centers and crypto-mining facilities, seek "rapid interconnection to the bulk power system (BPS)."

They seek to connect "very quickly (sometimes within one to two years)," the note says, "creating a risk that the demand will grow faster than available energy." And "In addition to resource- and energy-adequacy risks, computational loads also introduce challenges to grid operations during normal and event conditions."

A familiar pattern

The note reports "several incidents" in which such loads affected the grid by removing themselves "during system disturbances," in the Eastern Interconnection and in ERCOT, "over the past two years." A map beside that text, captioned "Figure 1: Customer-Initiated Load Reduction Events," carries a box headed "Eastern Interconnection Data Center Customer Initiated Load Reductions." Its entries, as printed:

• Jul 2024: 1,500 MW
• Feb 2025: 1,800 MW
• Feb 2025: 428 MW
• Mar 2025: 227 MW
• May 2025: 540 MW
• Jun 2025: 1,300 MW

Another box, headed "Crypto Mining Customer Initiated Load Reductions," is not reported here. The incidents, the note says, "largely unfold in a familiar pattern:"

• "The grid experiences a normally cleared fault, and voltage dips occur as part of normal reclosing to isolate the fault."
• "After several reclosing attempts, computational load entities react to the voltage dips by transferring their load to their backup power systems, removing themselves from the grid, with the goal of protecting their equipment from outages and any potential damage resulting from prolonged voltage dips."
• "As a result, large amounts of demand suddenly disappear from the BPS, creating oscillations on the systems as generation rebalances frequency and voltage to meet the remaining demand."

Historically, the note says, the grid "has not experienced, nor planned or operated for, these significant amounts of simultaneous load losses in response to a fault on the system." Planners and operators have traditionally focused instead on the risk of losing large generators. The recent incidents "did not create any near- or long-term issues for the system," but future ones "could result in severe impacts" as such sites grow in number and size, especially where they are concentrated, as in what the note calls "Data Center Alley" in Virginia. And since "all computational loads on a circuit could see the same disturbance and take the same action," the note says, "this creates significant common-mode failure risk."

A load that steps off the grid in a dip does so with the goal of protecting its own equipment. Taken at once by every load on a circuit, that same step could become the common-mode risk the note names.

What changed, and what has not

Some data centers have changed how they react. Working with grid operators, certain data centers "adjusted the protection and control schemes that they use to determine when their facilities transfer their load to a backup power supply." They "lengthened the time duration required for measuring voltage dips to be" counted, a word the note sets in quotation marks, "and altered the number of depressions during a specified period that would trigger disconnection from the grid." The transmission entity also included these requirements for "all newly connected facilities."

The note then sets two events side by side. On April 18, 2026, a fault "in the same vicinity and with similar characteristics" to an event in July 2024 that had led to 1,500 MW of load loss saw "only 500 MW of sustained data center load loss." The improved performance, the note says, "may be attributable to the changes made by the data centers."

The note says the risk is not abating. "Most recently, in July 2026, a nearly 4 GW customer-initiated load reduction occurred in Virgina, demonstrating that the risk is increasing, not abating." Its specifics "are still being analyzed as of September 2026." The spelling "Virgina" is the note's.

A change that may have helped in one vicinity has not settled the question. The largest event the note reports is also its most recent, and its specifics were still being analyzed when the note was written.

What NERC is doing

NERC is bringing these loads under its rules. It is developing registry criteria that "will define which entities will be required to register with NERC," and Reliability Standards that start with "foundational reliability activities, such as commissioning, communications, and modeling." The Federal Energy Regulatory Commission has directed NERC to complete the registry criteria and the foundational standards "by the end of 2026." Ride-through appears among the additional standards NERC "will consider" moving forward, beside modeling requirements and "critical cyber and physical security protections, including supply-chain risk management."

Until ride-through is a rule, how a campus behaves in a voltage dip is worked out between the campus and its grid operator.