U.S. coal retirements hit a 15-year low. The reason isn't policy. It's data center load.

The visible AI-power story is a procurement story. Hyperscalers are signing power purchase agreements with renewables developers and small modular reactor companies. Microsoft has restarted Three Mile Island Unit 1 through a 20-year offtake. Amazon, Meta, and Alphabet have each announced gigawatt-scale renewable and SMR commitments. The framing is that AI is paying for clean firm power, and on a press-release basis, that framing is correct.

The headline 2025 numbers from EIA's spring 2026 reporting reinforce that framing. Wind and solar generated roughly 760,000 GWh in 2025, a record 17% of the U.S. mix. Total U.S. generation hit 4.43 TWh, also a record. The growth side of the ledger is renewable.

The retirement side of the ledger tells a different story. EIA's April 2026 update reports operators retired 2.6 GW of coal capacity at four plants in 2025, the lowest annual coal retirement since 2010. Against an 8.5 GW planned-retirement schedule for the year, 4.8 GW were pushed out and 1.1 GW were cancelled. EIA's separate March 2026 piece on demand growth tells the other half: U.S. electricity demand grew about 1.7% annually between 2020 and 2025 versus 0.1% per year between 2005 and 2019, with data center electricity use named as the primary driver. EIA's April 2026 outlook is more direct: it concludes that fossil generation could rise with faster-than-expected growth in data center power demand.

Read together, those EIA notes describe a coupled system. When data center load surprises to the upside, planned coal retirements move out, not just because owners want margin but because regional reliability authorities and utility commissions need the megawatts on the bar to clear capacity auctions and reliability filings. This is why two-thirds of the planned 2025 coal retirements not happening on schedule is the load-bearing fact. The plants that were supposed to be off the grid are still on it. The marginal electron serving incremental data center load is, in many regions, a coal electron, even when the headline contract behind a hyperscaler campus is for solar plus storage or for an SMR that has not yet broken ground.

The verified facts, restated. A top-line decarbonization year by generation share, 17% wind and solar, sits next to a 15-year low in coal retirements. EIA has explicitly named data center load as the demand-side cause and explicitly said fossil generation could rise as that load grows. The deferral pattern is national, not regional.

The read: the visible PPA layer and the actual dispatch layer have decoupled. PPAs settle financially and shape long-term capex. Dispatch settles physically and shapes which plant fires up tonight. As long as data center load growth stays ahead of new firm low-carbon capacity coming online, retirement deferrals are the pressure-relief valve, and coal is the most over-indexed source of that pressure.

What this would mean if it holds:

First, hyperscaler 24/7 hourly carbon-matched commitments get harder to honor as coal stays in the dispatch stack. The matching math relies on either local clean firm power or grid imports that are themselves clean-firm at the hour of consumption. The deferral of 4.8 GW of coal retirements increases the probability that hyperscaler campuses in PJM, MISO, and SPP service territories are buying coal-attributable kilowatt-hours during peak training and inference windows, regardless of what the long-dated PPA on the campus says.

Second, the SMR and gas-peaker pipeline becomes more load-bearing for the AI capex thesis than the renewable pipeline. SMRs are still pre-construction in most U.S. permitting queues. Gas peakers are buildable in 18 to 30 months. If the deferral pattern continues into 2026 and 2027, gas is the realistic 2027 to 2029 incremental firm-power supply for AI campuses, and the SMR commitments become the 2030 and later story, not the 2027 story.

Third, regulator behavior is now an AI infrastructure variable. Public utility commissions and ISOs hold the actual lever on whether a coal plant retires on its filed date. Hyperscalers writing 5 to 15 year campus PPAs are buying assets whose carbon attribution depends on a regulator decision that has not yet been made and that may move based on data-center-driven reliability filings the hyperscaler did not author.

Reading: the 2025 coal-retirement print is not noise. It is the first numeric confirmation that the AI-power story has split into two layers. The procurement-and-press-release layer is renewable. The dispatch-and-deferral layer is not yet.