Ricardo Semler on the prisoner’s dilemma on data centers: trillions of dollars obsolete in just a few years

The conversation around data centers has turned into a nose-to-nose shouting match between politicians, NIMBYs, and tech billionaires. Ask about the concerns and you’ll get an earful about water consumption, farmland, rural electricity bills and tax abatements.

Every objection is reasonable, but each assumes the buildings these data centers will occupy will still be worth something in 2035. Once again, everyone is asking the wrong questions. Here’s the right one: “What happens to data centers and the towns they occupy when the structures outlive the infrastructure?”

Demand for data center space is ravenous. According to Jones Lang LaSalle (JLL), one of the world’s largest commercial real estate companies, only one percent of North America’s data-center space sits empty. JLL says data center customers also contracted for a record amount of electric generating capacity—twenty-five gigawatts—in the first six months of 2026. That’s five percent of all the electricity the U.S. generates in an entire year (or 20.66 flux capacitors).

Because traditional data-center markets are running out of generation capacity, 77% of new data centers in the pipeline are shifting toward “frontier” markets like West Texas, northeast Louisiana, and east of Columbus, Ohio. Around $7 trillion in financing and investment is being lined up to keep the concrete and electrons flowing.

If that scenario sounds familiar, think Rust Belt. The beating heart of industrial America had capillaries running into hundreds of towns in the Midwest, Plains, and Appalachia—row upon row of factories turning out everything from lawnmowers to living room sets, creating jobs and mini-booms of economic prosperity.

That is, until globalization and automation turned out the lights, leaving the crumbling brick hulks and rusting machinery that have become avatars of exurban poverty and rage from Youngstown, Ohio to Gary, Indiana.

But the Rust Belt was built from things everybody wanted…until the economic equations stopped balancing. As the AI land rush reaches a fever pitch, we need to be talking about how long each layer of a hyperscale facility can expect to be functional.

Servers and networking are good for three to six years. Cooling and electrical architecture, seven to fifteen. Functional design, ten to fifteen. The building itself sees a thirty- to sixty-year slide to obsolescence. But what happens to those buildings—and the tax revenues they generate—when infrastructure inevitably gives way to greater speed and efficiency? Do AI titans upgrade or split town for more land, cheaper power, and bigger tax breaks elsewhere?

The scale of tech is a moving target, making all infrastructure bets longshots. That’s why one former AT&T underground communications center in Nebraska—54,000 square feet with its own cell tower—recently listed for just $7.95 million. When obsolescence happens at terabytes-per-second speed, we risk a landscape littered with vacant, crumbling, toxic leviathans.

We ran this experiment in the late 1990s. Everyone said the Internet would change the world, and they were right. What they got wrong was how much infrastructure it would take. Technology improved faster than crews could dig. Billions of dollars of cable ended up buried and dark. Promises might be made with ones and zeroes, but they’re kept with atoms and molecules.

That’s the cautionary tale for AI. You don’t have to believe it’s a bubble, just that we’ll continue to get better at delivering computing than the people pouring concrete in 2026 realize. History backs this up. Five years ago, a conventional data center rack drew five to ten kilowatts. Today, AI racks can draw 100 to 250 kilowatts, and JLL has seen proposals for racks requiring as much as 600 kilowatts.

If each rack does 10x the work, demand doesn’t have to fall for the building boom to overshoot. Everyone can want more computing every year and need less space to get it. Scarcity just moves from acres to megawatts.

Software is doing the same thing. OpenAI found that between 2012 and 2019 the computing needed to hit a fixed ImageNet benchmark fell by a factor of forty-four. Extend that ten years from today, with 2026 benchmarks, and the numbers become ridiculous.

The big problem is the Jevons Paradox: if demand for a thing is highly responsive to price, making it cheaper to produce expands demand more than the savings shrink it. LED bulbs cut household lighting energy usage because nobody needs one hundred times more light. Make computing ten times cheaper and we may use a hundred times more of it.

That doesn’t mean data centers will be unnecessary. The current generation of data centers will become unnecessary. Demand will press companies to continuously upgrade computing speed and power, requiring greater supplies of electricity and coolant as well as engineering built to accommodate bleeding-edge hardware. But real estate is inelastic.

Warehouses need forty-foot ceilings, dozens of loading docks, and acres of truck yard. Picture a building in rural Ohio designed around one era of silicon—switchgear, busways, chillers, reinforced walls, a layout tuned to chips that somebody wanted—until nobody wants them because it’s now a new era. That building is too specialized for a warehouse, too remote for housing, and too expensive to convert into anything. The AI company, hungry for ever-escalating computational might, breaks contract, takes its billion-dollar toys and departs with a hearty cry of, “Sue us, suckers!” In their wake, a useless shell in another fiscally crippled town.

But why are so many brilliant, wealthy people going into massive debt to build data centers that will be obsolete in five years? It’s the prisoner’s dilemma: No one dares exit the race. Nvidia is supreme for now, but the tech titans can’t afford to let the others pass them, so they keep pushing in their chips. The pensioner and mutual fund investor have no seat at the table.

The major players know the hardware won’t remain competitive. They’re betting—with taxpayer dollars—that electricity will continue to be scarce enough to act as a brake on growth. If they’re wrong and rising demand leads to town-sized compounds flying past their sell-by dates, they’re covered. The banks will have made their commissions, and the tech companies have covered themselves with smart contracts and smarter bookkeeping.

For instance, Meta financed its $27 billion Louisiana Hyperion data center through a joint venture with Blue Owl, which owns eighty percent. The venture issued the bonds, so the debt is off Meta’s balance sheet. Meta occupies the site under a four-year lease with renewal options out to sixteen years and a residual value guarantee. If the venture goes bust, it’s mostly taxpayers on the hook for incentives, utilities, and pension-fund exposure.

This is the AI cautionary tale no one is telling. According to Epoch AI, the performance of leading AI supercomputers has doubled every nine months, laughing at Moore’s Law as it sprints by. Silicon runs into practical obsolescence far sooner than the structures that house it.

That imbalance threatens the U.S. with a future of dark, silent unusable buildings standing watch over revenue-starved municipalities—a Silicon Belt. The question isn’t whether data centers will become obsolete, but whether we can distribute the risk in a way that doesn’t turn predictable failure into economic catastrophe. If we can, as long as DeepSeek doesn’t come up with something better, we’ll be okay.

Then again, AI may kill us all in 10 years, in which case…never mind.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on