The phrase "energy collapse" is no longer a distant possibility but a real threat facing national power grids, and its primary accelerant is the explosive growth of artificial intelligence. In regions like the PJM, which serves 13 states and D.C., the strain is already evident. A recent capacity auction hit a record $329.17 per MW-day, a price that will translate into billions of dollars in payments to generators for 2026-27. This financial marker, coupled with frequent "maximum generation" alerts, signals a grid under immense stress, a reality that the International Energy Agency projects will see global data center electricity demand roughly double by 2030.
The problem is multifaceted, originating from AI's two primary power demands: massive one-off training runs and the continuous "relentless drip of inference" from millions of daily queries. This demand is not just outstripping power generation but also exposing critical weaknesses in infrastructure. Transmission lines are bogged down in long interconnection queues, transformer shortages are common, and permits lag years behind construction plans, creating a dangerous bottleneck that prevents power from reaching the data racks when it is needed most.
To avert a crisis, utilities are beginning to push for new, equitable solutions. Companies like Dominion Energy in Virginia have proposed new rate classes with long-term commitments for data centers, ensuring regular customers are not forced to subsidize hyperscale growth. Simultaneously, the tech industry is forging its own path forward. Giants like Microsoft, Google, and Amazon are increasingly turning to nuclear power, inking deals for both traditional and small modular reactors (SMRs). This strategy provides a source of clean, reliable, 24/7 power that can be sited near data centers, relieving grid pressure and bypassing the slow, costly process of grid upgrades.
