Here's a pattern you've probably seen before: wild enthusiasm about a transformative technology, obscene amounts of capital racing to build infrastructure, then a spectacular crash that leaves everyone wondering what they were thinking. Except the infrastructure doesn't actually disappear. It just gets cheaper and powers the next few decades of growth.

Welcome to the AI data center boom of 2025, which is basically a remix of the railroad mania of the 1800s, the electrification push of the 1920s, and the fiber optic frenzy of the late 1990s. Same ingredients, same trajectory, probably the same messy ending.

The twist, though, is that bubbles might punish investors who show up at the wrong time, but they're remarkably good at building stuff that sticks around. KKR's latest research hammers this point home: capital cycles are temperamental and fickle, but physical assets compound quietly for years after everyone stops paying attention.

Right now, AI data centers are being constructed at a pace that feels equal parts exciting and unhinged. McKinsey estimates nearly $7 trillion will flow into global data center infrastructure by 2030, with more than 40% landing in the United States. It's excessive. It's chaotic. It might also be completely necessary.

History Doesn't Repeat, But It Sure Rhymes Loudly

The closest historical parallel is the late 1990s fiber optic buildout. Telecom companies doubled their capital spending over four years, buried more fiber than the world could possibly use, and then watched the NASDAQ crater by 78%. Brutal for shareholders, obviously. But all that "wasted" fiber became the literal backbone of the modern internet.

The AI infrastructure wave is following a remarkably similar playbook, with one crucial difference that KKR points out: demand isn't lagging nearly as much this time. U.S. data center vacancy rates are sitting near record lows, which suggests the supply frenzy might actually have some fundamental support.

The hyperscalers—Amazon (AMZN), Google (GOOG), Microsoft (MSFT), Meta (META)—are on pace to spend over $300 billion in capital expenditures this year. That figure doesn't even include the parade of startups and enterprises building GPU farms that resemble small cities and consume alarming amounts of electricity.

Power availability has morphed into the new bottleneck. Substations and transformers are now critical-path items. In data-center-dense regions like Northern Virginia, land, permits, and grid access have become legitimate competitive moats—the kind you can't replicate by hiring better engineers or raising another funding round.

So even when the broader AI ecosystem goes through its inevitable bubble pop, these physical assets won't vanish into thin air. They'll get repriced, reallocated, and repurposed for whatever comes next in the compute demand cycle.

Which raises the obvious question: who actually survives when the music stops?

Three Things That Separate Winners From Wreckage

History offers a pretty clear playbook. Across cycles, three factors consistently determine who makes it through to the other side:

Real Economics, Not Fantasy Spreadsheets

KKR emphasizes that rigorous project underwriting matters more than ambitious "total addressable market" presentations. That means calculating actual returns after accounting for power costs, capital costs, and utilization risk. Many current players are renting scarce GPUs or power capacity at razor-thin margins. Those business models don't survive long once credit markets tighten.

Moats You Can't Manufacture Overnight

Power rights, land ownership, grid connections, permits, and the operational expertise to serve hyperscale customers—these aren't nice-to-haves. They're the modern equivalent of railroad rights-of-way in the 1800s or long-haul fiber conduits in the 1990s. You either have them or you don't, and acquiring them takes years, not quarters.

Discipline When Everyone Else Is Losing Their Minds

Winners structure long-term offtake agreements, diversify their counterparty risk, build to suit rather than speculate, and design facilities flexible enough to absorb the next generation of accelerators. In a sector where hardware refresh cycles happen in quarters instead of years, this kind of planning isn't optional—it's existential.

The Reckoning That's Already Starting

McKinsey's research shows this expansion is straining state-level infrastructure in ways previous bubbles didn't. Data center power demand could triple by 2030. Water consumption is becoming a genuine political issue. Local labor markets can't scale fast enough to meet demand.

Communities are starting to push back, not because they're anti-technology, but because the resource trade-offs are getting harder to ignore. Power grids built for residential and light commercial use are suddenly supporting facilities that draw electricity like small industrial cities.

But here's the thing about major technology waves: they always find equilibrium eventually. Railroads consolidated into a handful of survivors. Electrification standardized around winning formats. Fiber optic networks got acquired for pennies on the dollar by companies that became telecom giants. And the infrastructure that remained created vastly more economic value than it destroyed during the buildup.

The AI boom will probably follow the same arc. There will be casualties, probably spectacular ones. But the data centers, the power infrastructure, the grid connections—those aren't going anywhere. They'll just get cheaper and more accessible, which is exactly how infrastructure manias are supposed to work.