Two facts, both from the federal government, both published this year. First: after fifteen years of flat electricity demand, the Energy Information Administration now projects U.S. generation growing 25% to 50% through 2050, with installed capacity rising as much as 90%, a transformation the agency attributes overwhelmingly to data center construction. Second: in 2025, the U.S. power sector retired 2.6 gigawatts of coal capacity, the least since 2010 and less than a third of what was planned, as Department of Energy emergency orders kept scheduled retirements running.
These are not two stories. They are one story, and almost no one in either industry wants to tell it: the artificial intelligence buildout has become the most powerful force extending the operating life of America’s coal fleet.
The Arithmetic Nobody Ordered
No hyperscaler chose this. The mechanism is system lag, not preference. Load growth arrives on construction timelines, eighteen to thirty months from groundbreaking to energization. New supply arrives on infrastructure timelines: gas turbines carry multi-year manufacturer queues, and interconnection backlogs hold new generation of every kind in line for years. When demand shows up faster than new supply can physically be built, the grid does the only thing a grid can do. It runs what already exists, harder and longer. In much of the country, what already exists and can run on command is coal.
Hence the pattern of the past two years. Plants with published retirement dates receiving 90-day must-run orders, renewed again and again. Retirements postponed, then postponed again. Utilities in data center corridors revising capital plans upward by billions while quietly extending the very units those plans were supposed to replace. The EIA now hedges its own retirement schedule with the caveat that plans are more subject to change than usual, which is a statistical agency’s way of saying the dates are fiction.
Why the Silence
The silence is structural, and it is worth being fair about why. The companies driving the load growth carry public clean energy commitments and meet them honestly, through power purchase agreements that fund real renewable construction. The utilities serving them are simultaneously building that new capacity and running the old units that keep the lights on while it arrives. Neither has any incentive to advertise that the marginal megawatt-hour stabilizing a data center corridor is, in several regions and several hours of the day, coal-fired. The contracts are clean. The grid is a mixing bowl. Everyone reports the contracts.
This is not a hypocrisy story, and readers hunting for one will miss the point. Announced energy futures and operating reality always diverge during a demand shock; that is what a demand shock is. The interesting question is never whether the gap exists. It is what the gap is worth, and who is positioned to finance it.
The Shadow CapEx
Because here is what the silence conceals: a capital requirement. Every extended plant is a customer that keeps buying. Every plant that keeps buying is a mine that keeps producing. And every mine asked to produce years past its planned horizon is a fleet of draglines, haul trucks, longwall systems, and preparation plants that must be rebuilt, recommissioned, and in some cases expanded, after a decade in which the sector was starved of investment capital by lenders who exited and never returned. The AI buildout’s headline capex, the hundreds of billions flowing into chips, shells, and substations, casts a shadow capex that no one models: the equipment spending required to keep legacy generation supplied for however long the extension lasts.
That spending is happening now, in a sector where the traditional financing channel is closed by institutional mandate rather than credit judgment. The operators funding it are among the most consolidated and conservatively capitalized in the industrial economy, and they are doing it largely without the banking system.
Markets eventually finance what is actually happening rather than what has been announced. The distance between those two things, between the energy transition’s stated schedule and the grid’s operating reality, is not a scandal. It is a window. The capital that moves through it will be supplied by institutions willing to underwrite the world as it runs, not as it is described. For now, that list is short, and the window is wide open.