US Grid Operator Plans Emergency Power Cuts to Data Centers
PJM Interconnection, the largest grid operator in the US, announced it will cut power to large data centers during peak demand periods starting next year to prevent widespread blackouts. The policy signals a direct collision between the AI infrastructure buildout and physical energy limits.
Original sourcePJM Interconnection, which manages the power grid across 13 states and the District of Columbia, has disclosed plans to implement temporary, targeted power interruptions to large data center customers as a demand-response measure to prevent grid-wide blackouts. The move affects one of the most data-center-dense regions in the world, including Northern Virginia's so-called "Data Center Alley," which hosts a significant share of global cloud and AI compute infrastructure.
The announcement reflects a growing tension between explosive demand for AI compute and the physical constraints of the electrical grid. Data centers have become among the largest and fastest-growing consumers of electricity in the US, with hyperscalers and AI companies racing to bring new capacity online faster than grid upgrades can keep pace. PJM has been warning for several years that reserve margins are tightening as legacy power plants retire and load growth accelerates.
Under the proposed framework, data centers classified as interruptible load customers would be subject to curtailment during emergency conditions, likely tied to extreme weather events or unexpected generation shortfalls. Data center operators have invested heavily in backup generation and power redundancy, but those systems are designed for brief outages — not coordinated, grid-operator-directed load shedding that could last hours. The practical and contractual implications for uptime SLAs, particularly for AI training runs and latency-sensitive inference workloads, are significant.
The policy is a bellwether for a broader national infrastructure problem: the AI industry's power appetite is outrunning the grid's ability to supply it. With transmission buildout measured in decades and new generation capacity still years away, demand-response curtailment may become a standard operating condition rather than an emergency measure. Other grid operators are watching closely, and similar policies could follow in regions where data center growth is concentrated.
Panel Takes
The Futurist
Big Picture
“The thesis here is stark and falsifiable: AI compute demand will outpace grid capacity additions through at least 2030, making curtailment a structural feature rather than an edge case. For this to resolve cleanly, either nuclear and storage come online faster than current timelines suggest, or AI training workloads become significantly more geographically mobile — shifting to wherever electrons are cheap and available. The second-order effect nobody is pricing in yet is that interruptible power contracts become a new axis of competitive advantage: the hyperscaler that can tolerate a 4-hour training interruption cheapest wins a margin war that has nothing to do with model architecture.”
The Founder
Business & Market
“The buyer here is every cloud provider and AI infrastructure company with colocation in PJM territory, and this is a direct tax on their uptime SLAs — which means it flows straight to their enterprise customers' contracts. The companies that built their moat on five-nines availability now have a grid operator as an uncontrollable upstream dependency, and no amount of backup diesel changes the calculus when the curtailment is coordinated and prolonged. Watch for a land grab in ERCOT and other grids with looser demand-response rules, and watch for the first hyperscaler to announce a dedicated power purchase agreement with a behind-the-meter generation asset as a competitive differentiator.”
The Skeptic
Reality Check
“The part of this story that deserves more scrutiny is the word "temporary" — because the conditions that make curtailment necessary are not temporary, they are structural, and the gap between new generation coming online and current load growth means PJM is managing a slow-motion crisis with emergency tools. Data centers will negotiate interruptible rate tariffs that look cheap until the first major training run gets killed mid-epoch, at which point the true cost of "temporary" becomes very clear very fast. This kills exactly the workloads that can't be checkpointed cheaply: long inference batches, real-time serving under SLA, and any operator naive enough not to have modeled grid risk into their infrastructure stack.”
The PM
Product Strategy
“The job-to-be-done for AI infrastructure teams just got a new mandatory requirement: model grid curtailment risk into capacity planning, the same way you model AZ failure. That's not a nice-to-have — it's a gap in every current capacity planning tool I've seen, none of which have a field for "interruptible load classification" or "PJM emergency event probability by month." The product opportunity is real: whoever builds the ops tooling that gives ML platform teams visibility into curtailment risk windows, auto-schedules deferrable training jobs around them, and surfaces the cost delta of interruptible vs. firm power contracts will have a very clear buyer with a very clear budget.”