It’s easy to think of AI as a kind of software magic floating somewhere in the cloud. In reality, the AI boom is rooted in the material world. Behind every chatbot response and image generator sits a sprawling physical machine — warehouses full of GPUs, giant cooling systems, substations, transformers, and power-hungry data centers that increasingly resemble factories more than server farms.
According to a new paper by Stijn Van Nieuwerburgh, the Earle W. Kazis and Benjamin Schore Professor of Real Estate at Columbia Business School, the scale of that buildout is booming. By late 2025, AI-related infrastructure investment was already driving nearly all observed U.S. GDP growth.
“We’d basically be in a recession right now without data centers,” Van Nieuwerburgh says. “That’s how big AI is.”
The boom shows no signs of slowing down. With the price tag of a single state-of-the-art AI campus running upward of $8 billion, expansion could require trillions in investment over the next several years — large enough to rival or exceed some of the biggest infrastructure booms in American history.
That buildout, Van Nieuwerburgh argues, forms the basis for a second, less-discussed AI revolution: a financial one. The biggest tech companies can’t — or don’t want to — pay for all this infrastructure themselves. Instead, they’re increasingly relying on outside investors, private credit firms, real estate players, and complex financing structures.
That opens the door to huge amounts of capital. It also spreads AI-related risk throughout the financial system in ways that are only beginning to emerge.
From railroads to GPUs
Van Nieuwerburgh compares today’s AI expansion to earlier eras of physical transformation like the building of the railroads, electrification, and the creation of the interstate highway system. He estimates that AI infrastructure investment could average roughly 2.8% of GDP during the current buildout — larger than the railroad boom at its peak.
The reason is simple: modern AI is incredibly resource-intensive. A traditional data-center server rack might consume 5 to 10 kilowatts of power, whereas a rack filled with cutting-edge AI chips can require more than 100 kilowatts. At that scale, conventional cooling no longer works. Facilities increasingly need liquid cooling systems, dedicated substations, and enormous amounts of electricity delivered continuously around the clock.
That’s changing not only the economics of computing, but also who finances it. Historically, cloud giants like Amazon, Microsoft, and Google largely owned the infrastructure they used. But the AI race has become so capital-intensive that hyperscalers are increasingly leasing facilities, partnering with outside developers, and turning to private-credit markets.
In practice, that means the company using the compute often doesn’t own the underlying infrastructure. Instead, pension funds, infrastructure investors, REITs, and structured-finance vehicles increasingly hold claims on the assets. Some firms are even exploring financing structures for GPUs themselves, essentially treating AI chips like aircraft or industrial equipment that can be leased and securitized.
That shift matters because it allows the AI buildout to scale far faster than tech company balance sheets alone would permit. But it also means AI risk is becoming more dispersed and more opaque.
The hidden risks beneath the AI buildout
One of Van Nieuwerburgh’s central arguments is that modern AI finance redistributes risk, rather than reducing it. He points to increasingly complex arrangements in which data centers are financed through off-balance-sheet entities, lease agreements, and layers of debt. One example involves a massive Meta-linked data center project financed with roughly $27 billion in debt through a maze of special-purpose vehicles.
The advantage for tech firms is obvious: they can preserve the appearance of being asset-light software companies rather than capital-heavy infrastructure businesses. That helps support lofty valuations, keeping investors happy.
But the structure creates vulnerabilities. To start with, AI hardware evolves rapidly, raising the possibility that today’s billion-dollar facilities could become obsolete faster than expected. There’s also the risk that power shortages, grid bottlenecks, and semiconductor supply constraints could slow projects or undermine returns.
Then there’s the financial risk. Many projects effectively depend on a handful of giant tenants. If enthusiasm for generative AI cools, the effects could ripple outward through lenders, private-credit funds, and infrastructure investors. For policymakers and investors, Van Nieuwerburgh argues, it’s important to keep an eye on how money is flowing through the system, and to watch out for signs of crisis.
“It’s very hard to get the timing right with these big buildouts, and often what ends up happening is we get overexcited and accrue too much debt and then a bunch of these investments go bust,” he says. “That doesn’t mean these are not important, productive investments. What I’m arguing is that we need to be careful and think through the risks.”
