Flexible Data Centers May Need Near-Total Grid Cutoffs, Even When Annual Curtailment Is Tiny

An agreement to use slightly less electricity over a year can conceal a much tougher commitment: nearly disconnecting from the grid during its most stressed hours. A preprint by Nils Johnson, Geoff Blanford and John E. T. Bistline examines that distinction, arguing that the location, depth and duration of demand cuts matter as much as their annual total.

A small annual cut can mean a deep interruption

Connecting data centers sooner by making their electricity demand flexible has an appealing logic. If facilities reduce grid consumption when the power system is under greatest stress, existing infrastructure might accommodate them without waiting for new generation.

The resulting capacity to serve additional demand is called flexibility-enabled headroom. But an annual curtailment percentage does not describe what a facility must actually do. As the authors explain, losing 0.25% of annual electricity could mean one 22-hour interruption or 175 hours of shallower reductions. Those obligations require different operational capabilities.

Their preprint investigates this distinction using a national power-system model. Its central finding is conditional: substantial additional demand can fit within existing generation capacity, but only with favorable siting and potentially near-total cuts to grid-supplied power.

What the model compares

The researchers use US-REGEN to jointly optimize hourly generation, electricity transfers between 16 U.S. regions, storage operation and curtailment of new large loads. They add between 10 and 150 gigawatts of demand, assumed constant throughout the year.

In the fixed-capacity scenarios, generation and storage remain at 2024 levels, and new demand is allocated geographically to minimize curtailment. A comparison allows new capacity investment while retaining the same load locations.

Curtailment receives an intentionally high modeling price of $100,000 per megawatt-hour. This is not an estimate of operators’ actual costs: it makes curtailment a last resort when other resources cannot meet the model’s requirements.

A consequential assumption is that regional planning reserve margins—the capacity cushion above expected peak demand—are already binding in 2024. The model therefore assigns no headroom to any existing reserve surplus. Additional demand must instead rely on curtailment or changes in storage discharge to satisfy that constraint.

Fifty gigawatts, but demanding conditions

With optimized siting, the model accommodates 10 GW without curtailment. Ninety percent of that demand goes to California, Mountain-South and Texas, where storage discharge can be shifted into the hours that constrain reserve capacity.

At 50 GW, curtailed electricity remains below 0.1% of the new loads’ annual consumption nationally. No region requires curtailment in more than 20 hours, and no event exceeds eight hours.

Yet 14 of the 16 regions require reductions exceeding 90% of their allocated new load in at least one hour. Several require complete curtailment. That describes a regional requirement, not a prediction that every individual facility must disconnect simultaneously.

Facilities would need to stop consuming that grid electricity, whether by reducing activity or substituting on-site generation or storage. The authors caution that backup equipment used only occasionally can be expensive, and moving capacity behind the meter may still depend on constrained equipment supply chains.

At higher demand levels, frequency and duration become more challenging. With 150 GW added, annual curtailment reaches about 1.5%, and the longest event in one region lasts 137 hours.

Location changes the result

The study also compares two allocations of 49 GW of projected new data center peak demand between 2024 and 2030: announced locations and locations optimized for headroom.

The South Atlantic region, including Virginia, receives about 13 GW under announced siting but less than 4 GW under optimized siting. With existing capacity alone, announced demand there requires roughly 2% annual curtailment, spanning more than 550 hours across 88 events.

Six regions exceed a ceiling of 0.1% annual curtailed energy under announced siting. None exceeds it when the same national demand is allocated according to modeled headroom. National totals therefore do not establish whether particular proposed locations can be served.

Emissions, prices and boundaries

Fixed-capacity scenarios rely more heavily on existing coal and gas plants. Allowing investment produces incremental electricity with a carbon intensity up to 24% lower. At 150 GW, the national energy-only wholesale price is 28% higher with fixed capacity than with new investment. These are not retail-bill estimates.

Nor does the study establish that building capacity is cheaper overall. Its reported system costs are lower with fixed capacity across the examined range, but exclude facilities’ costs of delivering curtailment, including lost computing revenue and on-site equipment.

The authors describe their headroom estimates as upper bounds, not connection guarantees. The analysis uses one weather year and omits within-region network constraints, detailed plant operations and probabilistic reliability assessment. It also excludes backup-generation emissions and assumes new demand can curtail fully.

The practical message is to specify flexibility obligations through depth, frequency, duration and maximum event length—not annual energy alone—and evaluate them where facilities actually plan to connect.