Why AI Data Centers Could Cause the Next American Blackout

In May 2026, approximately 1,000 megawatts of computing infrastructure dropped offline in seconds. That single event triggered the North American Electric Reliability Corporation (NERC) to issue its most urgent class of reliability warning, a Level 3 Essential Actions alert targeting every major operator and planner across the bulk power system [1]. The grid had experienced a rapid load loss before, but never one that moved so fast or involved loads so large that operators had no realistic chance to respond in real time.

The May 2026 alert was not an isolated alarm. It was the regulatory reckoning for a problem that had been building for years inside data centers from Virginia to California, where the explosive growth of AI training and inference workloads has fundamentally changed what these facilities demand from the power grid.

The Hidden Instability in America's Data Centers

Grid operators need balance. Every megawatt of electricity consumed must be matched by generation elsewhere on the system. When that balance breaks, grid frequency drifts, protective systems engage, and cascading outages become a real possibility.

What makes AI data centers uniquely dangerous to this balance is speed. Traditional commercial loads change gradually. AI facilities do not. A large language model training run can shift power demand by hundreds of megawatts within seconds as servers ramp up or complete processing batches. NERC observed customer-initiated large load reductions and significant oscillations occurring in seconds, leaving little room for real-time responses from system operators [1].

The incidents have already happened. In July 2024, 60 data centers in Northern Virginia dropped 1,500 megawatts simultaneously over the course of 82 seconds. Grid operators narrowly avoided invoking emergency procedures, according to subsequent reporting [4][8]. The event demonstrated exactly how quickly large computational loads can destabilize the system and exposed protection circuit behaviors that trip within electrical cycles, faster than any human operator can respond [3].

Grid operators have limited visibility into fast-moving load changes from data centers [3]. They know what these facilities draw at any given moment, but they cannot see the rapid-fire shifts happening inside them. This creates a dangerous gap between what the grid needs to stay stable and what operators can actually do about sudden imbalances.

A 16-Fold Jump in Power Demand

The underlying cause of this crisis is a staggering increase in the power density of individual servers.

GPU racks that drew 8 kilowatts five years ago now require 132 kilowatts [9]. That is a sixteen-fold increase in power density per rack, a change that has made modern AI data centers fundamentally different infrastructure from their predecessors. A single three-story facility under construction in Vernon, California, in April 2026 was designed to consume 49.5 megawatts [2]. Facilities of this scale were rare a decade ago. They are becoming commonplace.

Lawrence Berkeley National Laboratory projects that U.S. data center power consumption will grow from 176 to 192 terawatt-hours in 2024 to a range of 325 to 580 terawatt-hours by 2028 [6]. The Electric Power Research Institute forecasts that data centers could account for 9 to 17 percent of all U.S. electricity consumption by 2030 [6][7]. Summer peak demand across the bulk power system is expected to rise 24 percent over the next decade, with data centers driving most of that increase [1].

These are not incremental changes. They represent a fundamental reshaping of when and how much electricity the grid must supply.

Why Frequency Stability Is Now at Risk

Grid frequency, measured in hertz, must stay close to 60 cycles per second across the Eastern or Western Interconnection to maintain stability. Frequency deviations trigger protective relays and, if severe enough, cause generators to trip offline.

The traditional understanding of large loads treated them as steady, predictable consumers. A steel mill or an aluminum smelter draws enormous power, but it draws it consistently. Grid operators can plan around that. AI data centers under heavy training or inference loads can change their consumption profile entirely within seconds, creating frequency and voltage instability that existing grid models did not anticipate [5].

The July 2024 Virginia incident made this concrete. Sixty facilities disconnected from the grid almost simultaneously, removing 1,500 megawatts of load that was helping to stabilize system frequency. Had emergency procedures been triggered, the consequences could have extended well beyond those facilities [8]. Data centers can act as both grid assets and grid liabilities depending on their control systems [5]. Right now, many of them are operating as liabilities.

Capacity Shortfalls and Regulatory Responses

The problem extends beyond stability. PJM Interconnection, which operates the grid covering 65 million people from New Jersey to Illinois, projects a generation shortfall of up to 49 gigawatts by 2028 as AI data center load growth outpaces new capacity additions [7]. Capacity auction prices in PJM rose 833 percent between the 2024-2025 and 2025-2026 delivery years [7]. The signal from the market is unambiguous: new generation cannot be built fast enough to meet the coming demand.

States are beginning to act. Texas passed Senate Bill 6 in response to grid concerns driven by data center concentration [6]. Germany implemented a new maturity assessment procedure for data centers effective April 1, 2026 [4]. NERC's May 2026 alert requires seven mandatory actions from registered entities including balancing authorities, transmission operators, transmission planners, and generator owners [1][2]. Registered entities must acknowledge receipt by May 11 and respond by August 3, 2026 [1]. The alert notes that noncompliance can lead to referral to FERC for compliance action and potential monetary penalties [1].

NERC last issued a Level 3 alert approximately a year earlier, also regarding rapid changes in power system loads, this time from inverter-based resources [1]. Level 3 alerts are rare. The fact that the grid now faces this category of risk from two different directions, simultaneously, underscores the scale of the challenge.

The Path Forward Takes Years

Fixing this will not be quick. NERC's own alert acknowledges that addressing the underlying issues will take years of coordinated effort, including drafting new regulations, understanding equipment limitations, and re-prioritizing loads [2].

What is needed, at minimum, is faster communication between data centers and grid operators, coordinated load shedding procedures that prevent simultaneous disconnection, and new planning standards that account for rapid, large-magnitude swings in computational power demand. The industry also needs to examine whether data center protection schemes can be adjusted to coordinate with grid stability requirements rather than simply protecting individual facilities [1][3].

The lights stayed on in July 2024. The lights stayed on in May 2026. That is not a reason for confidence. It is a warning.