AI Buildout's $10 Trillion Price Tag Is Starting to Look Like a Financial Stability Problem

The artificial intelligence buildout has quietly become the most capital-intensive technology rollout in American history, and the bill is coming due in a form that should worry anyone tracking financial stability.

A new study prepared for a Brookings Institution conference this week puts hard numbers on what has largely been a qualitative debate. Stijn Van Nieuwerburgh, a finance and real estate professor at Columbia Business School, estimates the AI buildout will consume roughly 3.6% of US GDP annually through 2032, translating to more than $10 trillion in required investment. For context, that exceeds the 2.2% of annual GDP absorbed by the railroad expansion of the late 1800s, previously the most expensive technology rollout on record. It dwarfs the roughly 1% annual GDP commitment tied to the interstate highway system in the 1950s or the telecom buildout of the mid-1990s.

Scale alone would be notable. What makes this different, and more dangerous, is how it is being financed.

The Financing Shift

Amazon, Meta, and Alphabet's Google initially funded data center expansion largely out of retained cash. That era is over. The investment required has outstripped what even the largest hyperscalers can self-finance, pushing the industry toward external capital: banks, private credit lenders, real estate firms, and an expanding cast of intermediaries structuring the deals.

Van Nieuwerburgh conservatively estimates the buildout will add 183 gigawatts of new data center capacity over the next seven years, more than three times the roughly 57 gigawatts currently installed. Financing that expansion has meant higher leverage, risk redistributed across parts of the economy that have little direct exposure to AI's fortunes, and growing reliance on revenue streams that remain unproven.

"This is freaking complicated," Van Nieuwerburgh told reporters, describing the web of arrangements between AI firms, hyperscalers, lenders, and real estate players now underpinning the sector's growth. That complexity is not incidental. It is the risk.

He draws a direct parallel to 2008. "This opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis," he said, referring to the layered, hard-to-value financing structures that unraveled and triggered the 2007-2009 recession. The comparison is deliberate, not rhetorical. Special purpose vehicles obscure who actually bears the risk when cash flows disappoint. That opacity is precisely what turned localized mortgage losses into a systemic event.

None of this means a crisis is imminent. Van Nieuwerburgh is careful to note that strong AI adoption, high data center utilization, and continued model improvements could generate the cash flows needed to support the investment as planned. The downside case depends on assumptions proving wrong, not on assumptions already having failed.

But the math on required returns is unforgiving. Van Nieuwerburgh calculates the industry needs to generate about $3.7 trillion in annual revenue by 2032 to justify the expected return on this investment. Current combined annual revenue for OpenAI and Anthropic sits around $100 billion. Closing that gap requires roughly 80% annual revenue growth, sustained for seven straight years, with no interruption.

That is not a modest ask. Software companies have posted extended stretches of 80% growth before, but rarely at this scale and rarely while facing execution bottlenecks on the infrastructure side simultaneously. Power availability, chip supply, and permitting delays all threaten to slow deployment even if demand materializes as hoped.

The stakes are already visible in the market. Oracle recently invoked force majeure on a data center project over power delivery delays, a move that reportedly sent ripples through AI financing arrangements tied to partners including Blue Owl. That is a live example of exactly the kind of execution risk Van Nieuwerburgh flags: complex, interlinked financing structures where a single project's delay can cascade into counterparties several steps removed from the original deal.

Political and regulatory pressure is building alongside the financial risk. Some localities are growing reluctant to host new data centers, citing strain on power grids and water resources. Federal Reserve officials are reportedly examining whether the construction boom itself is contributing to inflationary pressure. Even inside the industry, caution is surfacing. Anthropic's CEO has publicly urged AI companies to slow the pace of model development, a striking admission from a firm competing directly in the race Van Nieuwerburgh is studying.

Investors have generally treated AI infrastructure spending as a straightforward growth story: more compute, more capability, more revenue. This research reframes it as a leverage and financing-structure story, closer in character to a credit market puzzle than a technology adoption curve. The distinction matters because the two carry very different risk profiles. A slower-than-expected adoption curve is a growth disappointment. A leverage-driven unwind, propagated through opaque special purpose vehicles, is a systemic event.

What Investors Should Watch

Three signals will matter most in the months ahead. First, watch combined revenue growth at OpenAI and Anthropic against that 80% annual benchmark; any visible deceleration should prompt a reassessment of return assumptions embedded across the financing stack. Second, track how private credit and real estate financing vehicles tied to data center construction perform when individual projects hit delays, as Oracle's force majeure episode illustrates the transmission mechanism in real time. Third, monitor Federal Reserve commentary on whether AI-related construction spending is feeding into broader inflation readings, since that would signal the buildout has grown large enough to move macroeconomic policy, not just sector-specific valuations.

None of this argues for abandoning AI infrastructure exposure. It argues for pricing the leverage and opacity properly, something the market has not yet been forced to do. The railroads and telecom sectors both delivered enormous long-term value after their respective bubbles burst, but investors and lenders who were overexposed at the peak absorbed painful losses along the way. The AI buildout looks set to follow a similar arc. The only open question is how large the eventual correction proves to be, and who ends up holding the risk when it arrives.