Surging AI Energy Demand Forces Revival of Aging Fossil Fuel Plants
New York, Wednesday, 24 December 2025.
Utilities are canceling fossil fuel plant retirements to meet AI’s energy surge, a desperate move that drove electricity prices in the largest US grid up 800 percent this summer.
The Return of the Peaker Plant
The strain on the electrical infrastructure is most acute within the PJM Interconnection, the largest power grid in the United States, which services 13 states and the District of Columbia [1]. As of December 23, 2025, a detailed analysis of company filings reveals that approximately 60 percent of the fossil fuel power plants in this region that were scheduled for retirement have either delayed or canceled those plans [1][2]. The majority of these retained facilities are “peaker” units—plants designed to operate only during periods of maximum electricity consumption [1]. While these units provide a critical safety valve for grid reliability, they are often older, less efficient, and significantly more expensive to operate than baseload facilities, a factor directly contributing to the soaring costs observed in wholesale markets [2].
Environmental Costs and Policy Shifts
The reactivation of these facilities presents a quantifiable setback for environmental targets. Peaker plants running on natural gas are documented to emit 1.6 times more sulfur dioxide per unit of electricity than non-peaker plants, and oil-fired units, such as those remaining at the Fisk site, can emit up to 22.7 metric tons of sulfur dioxide annually [2]. This resurgence aligns with broader shifts in U.S. energy policy under the current administration, which has voiced strong support for utilizing all available power sources to meet surging demand [1]. The U.S. Energy Information Administration (EIA) notes that coal production actually increased in 2025, driven by higher natural gas prices and the delayed retirement of coal plants [1].
Quantifying the AI Power Hunger
The urgency to secure power is driven by the computational intensity of modern artificial intelligence applications. The energy footprint of generative AI is vastly larger than traditional internet usage; a single query on ChatGPT consumes seven times the energy of a standard Google search [4]. The disparity widens with more complex tasks: generating just one minute of AI video requires as much electricity as a typical Western household consumes in a full hour, while training a large AI model can draw as much power as 100 homes use in a year [4].