Tom’s Hardware reports that coal power generation in the United States has jumped nearly 20%, driven largely by the rapid rise of AI data centers. Although major AI sites don’t rely on coal exclusively, they stress the power grid and coal is being pressed into service to maintain stability.
The distinction lies in the type of AI workload. For model training, i.e., long, continuous computations, nuclear power is a natural match, offering stable base loads. But during training, brief interruptions known as checkpointing (when model states are saved) momentarily reduce GPU demand. Since nuclear plants can’t quickly adjust down output, grid operators must compensate via flexible power sources such as coal or renewables.
Inference workloads are different: they consume power in unpredictable bursts, as user requests flood in. These surges stress local grid infrastructure, requiring instantaneous response from dispatchable energy sources. Coal and renewable plants are poised to fill that role when the load spikes.
So although the narrative “data centers are powered by coal” is misleading, coal increasingly acts as the margin fuel that balances grid fluctuations caused by AI. As natural gas prices rise and variability in demand amplifies, coal’s role may not just persist but grow.
This dynamic reveals an uncomfortable irony: efforts to decarbonize the power system collide with new electrical demands from AI, especially demand peaks. Until energy storage or flexible renewables mature, coal has reentered the grid, quietly, as a stabilizer.