Data Centers May Face Temporary Power Cuts

Starting June 2027, PJM Interconnection — the electrical grid operator serving roughly 67 million customers across 13 states from Virginia to Illinois — will begin deliberately cutting power to data centers of 50 megawatts or larger during periods of grid stress. The decision, reported by Tim De Chant at TechCrunch, arrives after a PJM capacity auction fell short of what the grid needs, and follows a year in which wholesale electricity prices on PJM's network have nearly doubled — a jump the grid's own market monitor has attributed directly to surging data center demand.

What Changed

PJM's curtailment policy targets the largest facilities: those at or above 50 megawatts, a threshold that catches hyperscale cloud campuses and large colocation facilities but leaves smaller edge deployments untouched. When the grid operator declares a shortage event, affected data centers will receive somewhere between 30 minutes and several days of advance notice before power is reduced. Operators won't simply absorb the loss — they will be compensated, structured similarly to existing demand-response programs that utilities have long used with industrial customers like aluminum smelters and auto plants.

The timing is significant. PJM's territory is not a generic stretch of the American grid. The Northern Virginia corridor — home to what infrastructure analysts routinely describe as the world's largest concentration of data center capacity — sits squarely inside PJM's footprint. AWS us-east-1, Azure East US, and GCP us-east4 are all PJM customers. So is the dense cluster of colocation facilities in Ashburn, Virginia that routes a substantial share of global internet traffic. A grid-level curtailment policy here is not an abstraction; it touches infrastructure that developers interact with every day through API calls, database queries, and compute jobs.

The price signal driving this decision is stark. PJM's market monitor has explicitly named data center growth as the primary driver behind wholesale electricity costs that have roughly doubled over the past year. Multiple outlets covering the story — including aggregated reporting via daily.dev — confirm this framing without disputing it. What the sources collectively leave unclear is whether the June 2027 date represents PJM's first curtailment action or the formal start of the policy framework; the distinction matters for operators trying to plan around it.

Looking further out, the pressure compounds rather than eases. Data centers are projected to consume four times their current electricity load on the PJM grid by 2035. That means today's curtailment policy, if anything, represents the conservative opening move of a longer regulatory negotiation between grid operators and the AI industry.

How It Works

When PJM signals a shortage, affected data centers will need somewhere to put their compute load — or simply shut parts of it down. The most likely near-term fallback is on-site diesel generation, and this is where the policy gets complicated. Federal environmental rules cap backup diesel generator use at 50 hours per year when operating under a demand-response designation, and at 100 hours per year for declared emergencies. Those caps exist because diesel generators are not clean infrastructure: they emit particulate matter, nitrogen oxides, and other pollutants at meaningful scale.

The health dimension is already a live controversy in Northern Virginia. Vantage Data Centers has faced scrutiny over diesel generator use in the region, a preview of the kind of community conflict that could intensify if curtailment events push dozens of large facilities to fire up backup power simultaneously. One analysis in the Reddit thread attached to this story — though the post itself required verification to access — points toward this pattern of local resistance as something data center operators have consistently underestimated. Our read is that the diesel question is where PJM's curtailment plan is most likely to hit friction: the grid workaround that operators reach for first happens to be the one that generates the sharpest local opposition.

The compensation structure PJM has proposed is meant to give operators a financial reason to participate voluntarily in demand response rather than simply treating curtailment as an adversarial act. This mirrors how grid operators have historically managed large industrial loads — paying customers to reduce consumption at peak moments rather than building additional peaker plants that sit idle most of the year. Whether that framing holds when the "load" in question is a machine learning training run or a customer-facing API endpoint is a question the policy does not yet fully answer.

What It Means for Developers

The most direct consequence is reliability risk in the US-East region specifically. Developers and engineering teams that have treated Northern Virginia cloud availability as effectively guaranteed now have a named, scheduled policy mechanism that can reduce that availability — not through accident or hardware failure, but through deliberate grid management with a known start date. The 30-minute minimum notice window is short enough that automated failover, not human response, is the only realistic mitigation.

Multi-region architecture has long been the orthodox recommendation for resilience, but the economics have pushed many teams toward single-region deployments, particularly when the target region offers the lowest latency to end users. June 2027 gives those teams a concrete deadline to reconsider. It also gives infrastructure procurement teams a reason to ask cloud vendors directly what their demand-response obligations are and how they intend to sequence load shedding across customer workloads when curtailment hits.

The longer arc matters too. If data center electricity demand reaches four times current levels by 2035, and if PJM's curtailment policy proves effective enough to be emulated by other grid operators, then what looks today like a single jurisdiction's regulatory experiment becomes a design constraint that shapes where compute infrastructure gets built for the next decade. Regions with surplus renewable generation, less constrained transmission capacity, or looser demand-response frameworks will become more attractive — not just on paper, but in actual capital allocation decisions. Developers building systems intended to run for years rather than months should treat grid policy as part of the infrastructure risk profile, not a footnote to it.

Sources