Columbia Professor Estimates US AI Buildout at Over $10 Trillion, Warns of Financing Risk

The US AI buildout may require more money than its biggest backers can generate themselves, pushing more risk onto outside financiers. In a new study, Columbia Business School professor Stijn Van Nieuwerburgh estimates the expansion will cost more than $10 trillion through 2032. He warns that rising leverage and uncertain revenue create meaningful downside risk, though those developments do not imply imminent financial distress.

A buildout larger than the rail era

In Reuters’ account, Van Nieuwerburgh’s study, prepared for a Brookings Institution conference, puts AI investment at about 3.6% of US gross domestic product annually through 2032. By his calculation, that exceeds the annual share absorbed by the rollout of railroads in the late 1800s, as well as the construction of interstate highways and the telecommunications expansion that began in the mid-1990s.

The physical forecast is just as ambitious. The paper estimates 183 gigawatts of new US data-center capacity over the next seven years, against about 57 gigawatts installed now. Those figures describe a projected addition and the current base, not capacity already under construction. The scale matters to the financing question: facilities must be funded before their future use can produce the cash to support them.

The financing changes hands

Earlier spending could be covered by cash accumulated at companies such as Amazon, Meta and Google. Van Nieuwerburgh says the investment now underway exceeds what major players can fund from their own cash flows. More of the expansion therefore depends on outside financing, increasing leverage—the use of borrowed money—and spreading exposure beyond the companies ordering the infrastructure.

In a briefing with reporters, Van Nieuwerburgh pointed to arrangements involving AI firms, major technology companies, banks, private-credit lenders and real estate firms. He likened the opacity of special-purpose vehicles—separate entities used in financing—to a feature of the subprime mortgage crisis. That is a warning about how difficult exposure can be to see, not a claim that an AI-driven financial crisis has begun.

The revenue the investment needs

The paper calculates that AI would need to bring in about $3.7 trillion a year by 2032 to deliver the expected return on the investment. Van Nieuwerburgh contrasts that target with current estimates of roughly $100 billion in combined annual revenue for OpenAI and Anthropic, and says revenue would need to grow about 80% a year. The comparison is a measure of the hurdle, not a like-for-like tally: the target is for the industry, while the current figure covers two companies.

That gap does not settle the outcome. Stronger growth in AI applications, heavy use of the new facilities and better models could produce stable cash flows, Van Nieuwerburgh writes. His downside case combines demand that falls short with fast-changing technology, obstacles to completing projects and high leverage. If expectations weaken, those financing commitments could make the correction more painful.