If you squint, the AI boom in 2025 looks like a mash-up of every infrastructure mania America has ever seen. The railroad boom of the 1800s. The electrification push in the roaring 1920s. The fiber internet mania of the late 1990s.
Each era had the same three ingredients: breathless enthusiasm, an overbuilt network, and—after the crash—a foundation that quietly powered the next several decades of economic growth.
That is the irony of bubbles. They punish the wrong investors, but reward the right infrastructure. KKR's latest research makes a clean point: capital cycles come and go, but physical assets stick around and compound.
Right now, nothing is under construction faster or with more anxiety than AI data centers. McKinsey estimates that nearly $7 trillion will go into global data-center infrastructure by 2030, with more than 40% of that happening in the U.S. It is excessive, it is chaotic – yet it is also almost certainly necessary.
A Louder Déjà Vu
The late 1990s fiber buildout is the closest recent parallel. Telecom companies doubled capital spending in four years, stuffed the ground with more fiber than anyone could use, then watched the NASDAQ collapse by 78%. And yet, all that "excess" fiber became the backbone of the modern internet.
AI infrastructure is following the same script, but KKR highlights one key difference. This time, demand is lagging much less. U.S. data center vacancy rates are hovering near record lows.
Hyperscalers—Amazon (NASDAQ:AMZN), Google (NASDAQ:GOOG), Microsoft (NASDAQ:MSFT), Meta (NASDAQ:META)—are on track to spend north of $300 billion in capex this year. That number doesn't include everyone else trying to train models, deploy inference, or build GPU farms that look like small cities – and probably use more energy.
Power availability has become the new zoning fight. Substations and transformers now sit on the critical-path timeline. And in data-center-dense regions like Northern Virginia, land, permits, and grid access have become competitive moats.
So even if the broader AI ecosystem goes through a classic bubble pop—and it eventually will—these hardened assets won't evaporate. They'll simply get repriced, reallocated, and repurposed for the next wave of compute demand.
Which brings us to what historically separates winners from the casualties.
The Bubble Survivors
Across cycles, the same three factors tend to determine who makes it to the other side:
1. Hard-nosed underwriting
Not "AI TAM spreadsheets," but real project economics—returns after power costs, capital costs, and utilization risk. KKR stresses that many of the weakest models today rely on renting scarce GPUs or power at thin spreads. They don't survive long once credit tightens.
2. Moats that can't be faked
Power rights, land, grid connections, permits, and the operational chops to serve hyperscalers. These aren't optional. They're the equivalent of rail rights-of-way in the 19th century or long-haul fiber conduits in the 1990s.
3. Discipline and de-risking
Winners have long-term offtake agreements, balanced counterparties, build-to-suit construction, and flexible designs that can absorb the next generation of accelerators. In a sector where hardware refresh cycles are measured in quarters, not years, this is survival planning.
Finding The Equilibrium
McKinsey's study shows that the current expansion is straining state-level infrastructure in ways past bubbles didn't. Data center power demand could triple by 2030. Water use is turning into a political flashpoint. Local labor markets can't keep pace.
Communities are now pushing back—not because they dislike data centers, but because the resource trade-offs are getting harder to ignore.
Yet every major technology wave eventually found equilibrium. Railroads consolidated. Electrification standardized. Fiber got bought for pennies on the dollar by companies that later became telecom giants. And the infrastructure that remained created far more economic value than it ever destroyed.
Read Next:
Image: Shutterstock
