Power plants once destined for demolition are now being brought back to life, extended for years, or sold at premium prices because artificial intelligence data centers need unprecedented amounts of electricity. Between 2024 and August 2026, the United States has witnessed the first commercial nuclear plant ever brought back from decommissioning status, a second shuttered reactor rebuilt under a twenty-year contract with a major software company, at least five fossil fuel plants ordered by the federal government to keep operating past their approved retirement dates, and a state governor negotiating to allow two of his state's largest coal plants to run four extra years.

This reversal represents what researchers call "Retirement Repricing," a phenomenon where assets once considered worthless have suddenly acquired enormous market value. The retirement date of a power plant, which used to be an administrative milestone processed quietly by utility commissions, has become a contested public decision fought over by governors, federal secretaries, hyperscalers like Microsoft and Amazon, unions, and local communities.

Why Are Data Centers Consuming So Much Electricity?

The demand is staggering. The International Energy Agency projects that global data center electricity consumption will roughly double from the mid-2020s to approximately 945 to 950 terawatt-hours by 2030, more than the entire electricity consumption of Japan today. AI-optimized facilities are expected to more than quadruple their power draw, with data centers accounting for nearly half of all U.S. electricity-demand growth this decade.

The four largest hyperscalers, Amazon, Alphabet (Google's parent company), Microsoft, and Meta, committed a combined total between $650 billion and $725 billion in capital expenditure for a single calendar year in their fiscal 2026 guidance, up more than 60 percent from 2025's already record levels, with the overwhelming majority directed at AI data centers, chips, and the power to run them. Stanford's AI Index 2026 counts 5,427 AI data centers in the United States alone, more than ten times any other country, with U.S. AI data center power capacity reaching 29.6 gigawatts, roughly comparable to New York State at peak demand.

"There is no AI without energy, specifically electricity," stated Fatih Birol, Executive Director of the International Energy Agency.

Fatih Birol, Executive Director, International Energy Agency

How Is the Grid Struggling to Keep Up?

The American electricity grid spent twenty years planning for decline. Utilities and regional operators scheduled hundreds of retirements on the assumption of flat demand, cheap natural gas, and tightening environmental rules. That assumption has evaporated. PJM Interconnection, the thirteen-state market serving 67 million people, now forecasts its summer peak to climb roughly 85 gigawatts over fifteen years, to more than 241 gigawatts by 2040, against an all-time record of about 167 gigawatts set in 2006. Its ten-year annual growth rate, forecast at 0.3 percent as recently as 2021, is now 3.6 percent.

When demand accelerates that fast while new power supply additions crawl forward, every scheduled retirement becomes a negotiation. Harvard's Belfer Center, launching a joint Kennedy School and School of Engineering and Applied Sciences project on the question in February 2026, calls the collision of AI data center demand with the American grid "a watershed moment" for the electricity system.

What Makes Old Power Plants Valuable Again?

The economics of Retirement Repricing operate on several levels. First, existing power plants already have grid connections, transmission infrastructure, and decades of operational history. Building equivalent new grid connections would take seven to twelve years to recreate, according to transmission engineers cited in the research. Second, the federal government has begun invoking wartime-era statutes to command dying plants to keep operating. Third, capacity prices in regional electricity auctions have slammed into their administrative ceilings, making the option to extend or restart a plant financially attractive.

What used to be a worthless asset, scheduled for demolition and environmental remediation, has moved back into the money. The same steel and copper that was destined for scrap now commands offers measured in federal loans, purchase-power agreements, consent decrees, and auction premiums. The funeral has become a bidding war.

How Are Utilities and Governments Responding?

The response has been rapid and multifaceted. A regulated utility outbid a data center developer for a bankrupt coal plant, largely to keep its grid connection on the public network rather than lose it to private control. Federal emergency orders have commanded plants to remain operational. State governors have negotiated consent decrees to extend coal plant operations. The largest wholesale electricity market in the world has fallen short of its reliability requirement for the third consecutive year while clearing at its price cap.

These are not isolated incidents. Taken together, they reveal a fundamental restructuring of how the electricity system values assets. In the AI economy, the retirement of a power plant is no longer a terminal event. It is a strategic option, and options get repriced when the state of the world changes.

Steps to Understanding the Five-Layer AI Economy

Retirement Repricing occurs within a broader framework of how the AI economy functions as an interconnected system. Understanding this structure helps explain why power plants matter so much to AI companies:

• Layer One, Energy: The electrons, generation assets, transmission infrastructure, and grid positions without which nothing above exists. This is where Retirement Repricing occurs.
• Layer Two, Chips: The accelerators, memory, and fabrication capacity that convert electricity into computation. Demand at this layer drives demand at Layer One.
• Layer Three, Data Centers: The buildings, cooling systems, and siting decisions that convert abstract chip demand into geographically concentrated megawatt demand on the grid.
• Layer Four, Models: The training runs and inference fleets whose load profiles and flexibility determine what the lower layers must deliver.
• Layer Five, Applications and Agents: The consumer and enterprise adoption that makes inference demand persistent, structural, and politically undeniable.

Retirement Repricing's causes descend from Layer Five, where consumer and enterprise adoption of AI creates persistent demand, and its consequences propagate back upward through all five layers. A single data center developer with a nine-figure checkbook can now appear at a county courthouse asking about a substation, and suddenly the entire calculus of whether a power plant lives or dies becomes a matter of national infrastructure policy.

The implications extend far beyond the power sector. As AI infrastructure demands reshape electricity markets, the scarcest asset in the five-layer AI economy is no longer chips, cooling capacity, or even land. It is the right to connect to the grid, and that right now belongs to whoever can afford to keep the lights on.