The U.S. AI boom may require about $10.3 trillion in data center investment between 2025 and 2032, roughly 3.6% of GDP per year, according to a new Brookings study. That is about 50% above the peak of the 19th-century railroad boom and more than three times the internet-era telecom buildout.

Speaking on the Brookings Podcast on Economic Activity, the paper’s author, Stijn Van Nieuwerburgh of Columbia Business School, said the way the spending is financed is changing. Tech giants such as Meta and Microsoft now spend more on AI than they earn in operating cash flow, and are turning to leases, joint ventures, and private credit.

The core pattern is a shift from “build” to “rent.” Instead of funding data centers from their own balance sheets, hyperscalers increasingly pass the capital burden to third parties through leases, joint ventures, special purpose vehicles and private credit. A typical structure has a thin equity slice and a large block of long-term project debt.

The lender’s comfort comes not from the asset itself but from the tenant’s lease commitment. The tenant keeps that commitment flexible through short lease terms with renewal options, and a residual value guarantee backstops it by paying off the debt if the tenant walks away. Because accounting rules require recognition only when obligations are at least 70% likely, the future leases and guarantees mostly stay off the tenant’s balance sheet.

As a result, leverage migrates to outside investors while the sponsor looks lightly indebted. Chips are following the same path, financed through shorter-dated debt tied to fast-depreciating assets. Van Nieuwerburgh said the structure echoes the off-balance-sheet vehicles of the 2008 crisis, but noted differences: long-term debt funds long-term assets, and the tenant is investment-grade. About $1.4 trillion of related debt has been issued, roughly half with banks.

Possible outcomes range from a productivity-driven soft landing to an internet-style bust confined to equity investors, or a 2008-style crisis if banks are hit. While admitting we don’t have the measurement framework to discuss possible outcomes, Van Nieuwerburgh called for comparable disclosure of who owns, owes, leases, and guarantees each major asset, and for regulators to map exposures across banks, insurers, and private credit. The aim, he said, is transparency, not stopping the investment.

End Notes

Source: https://www.brookings.edu/wp-content/uploads/2026/10/BPEA-S9E1-financing-the-ai-buildout.pdf

Source: https://www.brookings.edu/articles/the-10-trillion-question-financing-the-ai-buildout/