PJM will reduce data center operations to prevent the largest US grid from failing

In plain terms - America's largest grid operator, PJM, is preparing to briefly cut power to data centers when the network gets too stressed. The AI construction boom is now visibly outrunning the physical grid, and the fix - for now - is rationing.

Context

PJM Interconnection covers 13 states in the Northeast and Mid-Atlantic - the largest electrical interconnection zone in the United States, and the first corridor of the DC boom (Northern Virginia in the lead). For 18 months, moratoriums (New York, publi #1097), connection delays, and tariff negotiations have multiplied. The announcement of July 28, 2026 marks a turning point: rather than delaying new DCs, PJM is preparing to cut - briefly - the power supply of those already in operation, if the network falters.

The Data

• PJM Interconnection is considering targeted temporary outages of data centers to avoid blackouts (TechCrunch, July 28, 2026).
• The pace of AI DC construction is described as "breakneck" by the same coverage, with grid operators "scrambling to generate power."
• On the Asia axis, a Rest of World coverage (July 28) makes the same observation: "Asia's AI race won't be won by capital or talent, but by whoever can keep the lights on" - the constraint shifts from silicon to electronics.

Analysis

Three things to see. Systemic: the same pinch is rising in Europe, in Asia, everywhere the pace of GPU installation exceeds the formation of new capacity (nuclear, gas, PPA in 4-8 years, DC in 18-30 months). Risk transfer: the scheduled outage shifts the risk from the utility (blackout) to the DC operator (client SLA). Pricing: interruptible rates will populate new contracts - down for flexible workloads (training), up for production inference.

Scenarios (12 months)

• 50% - A major AI DC experiences a publicized PJM outage, public debate on priority (residential vs cloud).
• 35% - Generalization of interruptible contracts, explicit split between flexible and non-flexible workloads.
• 15% - Political lifting of the constraint (fast-track nuclear, grid exemptions) - unlikely at 12 months.

Implications for the Practitioner

For a cloud CTO: map your workloads by tolerance to outages; training can become a monetizable "demand response," production inference cannot. For a leader: DC CAPEX is a bet with two unknowns - chip + electronics - not just the first.

To Watch

The first effective outage of a PJM DC, the FERC reaction (federal regulator), the translation into hyperscaler contracts.