The U.S. Department of Energy gave PJM Interconnection, which serves 67 million people across 13 states, emergency authority to curtail large data centers mid-operation to prevent blackouts. On July 2, 2026, as temperatures soared and demand approached record levels, PJM issued an emergency warning to prepare for possible curtailment of data centers and other large loads in Northern Virginia and nearby regions. While no actual cutoffs occurred, the episode signals a fundamental shift in how AI infrastructure constraints are now understood: power availability, not computing chips, has become the binding limit on AI scaling.

Why Did the Grid Operator Need This Power in the First Place?

PJM's preliminary peak demand on July 2 reached approximately 162,700 megawatts between 5 p.m. and 6 p.m., suppressed by roughly 6,000 megawatts of demand response measures. The grid operator later set a new all-time peak-load record of 168,158 megawatts during the heat wave, surpassing the previous summer peak from 2006 of roughly 165,600 megawatts. That margin represents the grid using nearly every available tool to avoid blackouts.

The underlying problem is growth velocity. PJM's 2026 long-term load forecast projects summer peak demand will grow at an average of 3.6 percent per year over the next decade, compared with just 0.3 percent in its 2021 forecast. Even after trimming some near-term AI-driven demand estimates, the direction remains clear: data center demand is expanding far faster than new power generation capacity can be built.

Northern Virginia's Data Center Alley, concentrated in Loudoun County with growth spilling into Prince William and Fairfax counties, sits directly inside Dominion Energy's service territory. Dominion serves more than 450 data centers from over 50 customers, making regional grid stress an immediate technology infrastructure crisis. The financial impact is already visible: PJM's independent market monitor found that forecast data center load caused $23.1 billion in capacity market costs in the last two base capacity auctions. Wholesale power costs rose 75.5 percent year over year, from $77.78 per megawatt-hour to $136.53.

How Are Tech Giants Responding to the Power Crunch?

Major cloud providers are taking matters into their own hands by securing dedicated power supplies. Amazon has tied data center growth to the Susquehanna nuclear plant in Pennsylvania through its Talen Energy relationship. Microsoft signed a 20-year power purchase agreement with Constellation Energy to support the restart of Three Mile Island Unit 1, now called the Crane Clean Energy Center. These deals make individual company spreadsheets look better, but they don't solve the underlying regional grid problem.

• Private Power Deals: Amazon and Microsoft are locking up their own generation capacity through long-term nuclear power agreements to ensure reliable electricity for AI data centers.
• Grid Timing Mismatch: New generation plants, nuclear restarts, interconnection approvals, and transmission upgrades take years to complete, while AI demand arrives in quarters.
• Transmission Constraints: Even with new power sources, the physical wires connecting generation to data centers remain under stress and require separate infrastructure upgrades.

What Does This Mean for Future AI Infrastructure Plans?

The July episode should fundamentally change how investors and executives evaluate AI infrastructure announcements. Graphics processing units (GPUs) still matter. Cloud capital expenditure still matters. But if the electricity behind a cluster can be interrupted when the grid is tight, the constraint is no longer only chips or capital. It becomes power delivered at the right place, at the right hour, through wires already under stress.

Cloud contracts typically discuss uptime, regions, availability zones, and service credits. The power system underneath those contracts operates on a different language: heat, reserve margins, transmission constraints, and generators tripping offline during peak evening hours. When those two systems collide, the cloud contract's promises become fragile.

Steps to Evaluate AI Data Center Power Reliability

• Actual Megawatts: Demand specific power capacity commitments from hyperscalers, not vague sustainability pledges or future procurement plans.
• Interconnection Timelines: Ask for documented timelines showing when new transmission capacity will be available and what regulatory approvals remain pending.
• Peak-Hour Performance: Understand what happens to your workloads on the hottest weekday of the year when the grid is operating at maximum stress.
• Backup Generation: Clarify whether your data center has on-site backup power and whether grid operators can curtail your service during emergencies.

The next serious AI infrastructure plan will need to show its power math explicitly. Not a slogan about clean energy. Not a slide about future procurement. Actual megawatts, actual interconnection timelines, and a clear answer for what happens when the grid is tight. The Department of Energy's June 30 order gave PJM the authority to direct backup generation resources to operate as a last resort before an Energy Emergency Alert 3 or during one, tied to forecast hot weather and blackout risk. That order made the constraint visible. The question now is whether the industry will acknowledge it.