PJM Prepares to Curtail Power to Data Centers as Grid Capacity Crunch Worsens

A Grid Under Pressure

PJM Interconnection, which coordinates electricity delivery across thirteen states and the District of Columbia, announced it will begin curtailing power to large data centers during periods of supply shortfall. The policy takes effect in June 2027 and applies to facilities drawing 50 megawatts or more, a threshold that captures most hyperscale operations but exempts smaller colocation sites and enterprise builds.

The decision follows a capacity auction in which PJM failed to secure enough new generation to meet projected demand. At DailyTechWire, we've tracked similar capacity warnings from ERCOT in Texas and CAISO in California, but PJM's territory is larger by customer count and spans industrial heartland states where manufacturing and data center loads increasingly compete for the same electrons.

Wholesale electricity prices across PJM's footprint have roughly doubled over the past twelve months. The grid's independent market monitor pointed to data center expansion as a primary driver, a conclusion that has intensified scrutiny of how interconnection queues are managed and how new large loads are approved.

Demand Response with a Tech Twist

The curtailment framework mirrors demand response programs that have existed for decades in manufacturing and heavy industry. Participants receive advance notice ranging from thirty minutes to several days, depending on forecasted conditions, and are compensated for agreeing to reduce or disconnect load. PJM has not yet published the compensation structure for data center participants, but precedent suggests payments will track avoided capacity costs and real-time locational marginal pricing.

What sets this iteration apart is the asset class. Data centers run latency-sensitive workloads, inference clusters that cannot tolerate interruption, and training jobs measured in GPU-weeks. Unlike a steel mill that can idle a furnace with predictable restart costs, a hyperscale facility serving real-time API traffic faces reputational and contractual risk if availability drops below service-level agreements.

The On-Site Generation Calculus

Many operators are expected to respond by deploying on-site generation, sidestepping the grid entirely during curtailment windows. Natural gas turbines, fuel cells, and modular nuclear units are all under consideration, but diesel gensets remain the most common fallback because fuel can be stored on-site and the technology is proven at scale.

Federal rules permit backup diesel generators to run up to fifty hours per year for demand response and up to one hundred hours annually when emergencies, maintenance, or testing are included. That window is narrow but sufficient for most grid events, which historically last hours rather than days.

The environmental cost is non-trivial. A 96-megawatt data center in Northern Virginia recently drew scrutiny after a health impact assessment estimated that diesel backup operation could impose tens of millions of dollars in annual damages to nearby communities through particulate and nitrogen oxide emissions. Vantage Data Centers, the facility's operator, was accused of coordinating with state environmental regulators to question the methodology behind the report, an allegation that underscores the tension between infrastructure urgency and local air quality.

The 4x Demand Horizon

Projections show data center electricity consumption in the United States will quadruple by 2035, driven overwhelmingly by AI training and inference. Hyperscalers are already pre-leasing gigawatts of capacity in markets with available interconnection, and utilities in Virginia, Ohio, and Georgia have filed integrated resource plans that assume multi-gigawatt data center load growth over the next decade.

PJM is running a second capacity auction to close the gap, but lead times for new gas, nuclear, or renewable generation stretch years, and interconnection queues are congested. In the interim, the grid operator faces a binary choice: either curtail interruptible loads or risk broader blackouts that would affect hospitals, transit systems, and residential customers.

Asia's Parallel Path

The infrastructure bind is not unique to the United States. In Singapore, the government imposed a moratorium on new data center construction in 2019, lifting it selectively in 2022 with strict energy efficiency requirements. Seoul and Tokyo have introduced tiered electricity pricing for large compute facilities, and Bangalore has seen rolling industrial load-shedding during peak summer months as renewable intermittency and coal plant retirements collide with data center growth.

China's approach has been more directive. Regulators steered new builds toward Inner Mongolia, Guizhou, and Gansu, provinces with surplus renewable capacity and cooler climates, while limiting expansion in coastal tier-one cities. The policy succeeded in distributing load but introduced latency penalties for workloads requiring low-millisecond access to end users.

What Comes Next

PJM's curtailment policy is likely a harbinger. As AI inference moves from cloud batch jobs to edge and real-time applications, the mismatch between where compute capacity is built and where generation capacity exists will sharpen. Operators will face a choice: pay for firm, uninterruptible power at a premium, build redundant on-site generation, or architect workloads to tolerate graceful degradation during grid events.

The third option remains underexplored. Distributed inference, workload migration across regions, and tiered service models that prioritize critical traffic during curtailment could reduce dependence on diesel gensets and fossil peaker plants. But those architectures require coordination between hyperscalers, utilities, and grid operators, a level of integration that has proven difficult to achieve at the pace the industry is moving.

For now, the largest grid in the country is preparing to flip the switch on its largest new customers. Whether that accelerates cleaner on-site generation or entrenches diesel dependency will depend on the economics of the next eighteen months and the willingness of operators to treat power as a design constraint, not an assumption.