Invest1 publisher3 min readPublished
A Columbia estimate puts $800 billion of bank credit behind the AI buildout
Stijn Van Nieuwerburgh's $800 billion adds up mortgages, syndicated loans and project debt across AI's physical infrastructure, and he says loans going bad there would cut what banks can lend everyone else.
The Investor · Invest desk
What happened
- Columbia professor Stijn Van Nieuwerburgh estimates banks have about $800 billion riding on AI, counting mortgages, syndicated loans and project debt, and he expects the figure to rise as the buildout continues.
- Data center projects, he said, are sometimes financed with debt at as much as 90 percent of overall project cost against equity of only 10 percent.
- KeyCorp's loans to utility companies reached $10.1 billion in the second quarter of 2026, up almost 40 percent in three years, and utilities made up 61 percent of the bank's loan growth over the past year.
- Truist Securities analyst Brian Foran said the AI boom turbocharged that utility lending, and that he does not think most people realized how big it had gotten.
- Van Nieuwerburgh said weak product demand, or cheaper Chinese alternatives that drive prices lower, could each be enough on its own to push AI-related loans underwater.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint Banks already at their data-center concentration limits cannot add the next project loan without displacing a borrower elsewhere on the book, so AI's marginal dollar of credit now competes with companies that have nothing to do with AI.
- exposure At ten percent equity, the first tenth of any shortfall in project value exhausts the sponsor and the remainder lands on bank capital, which makes a moderate revenue miss a lender's problem rather than a sponsor's.
- contradiction Van Nieuwerburgh treats a small demand shock as enough to impair these loans, while Foran argues the utility slice is regulated and probably fine, conceding only that you never really know until the tide goes out.
- decision Anyone pricing a regional bank's equity has to decide whether utility loan growth of this composition is safe regulated spread income or AI credit arriving under another name.
"A lot of banks are pretty much up against their concentration limits in their data-center lending, so they have tons of exposure," Van Nieuwerburgh said [3]. A bank at its limit has already allocated that line. The next grid or data-center loan it writes displaces a borrower somewhere else on the book. "Banks play a central role in the functioning of the broader economy," he told American Banker, and "if banks get exposed to data-center loans that are going bust, then that impacts their ability to lend to the rest of the economy" [6].
Leverage sets how fast that happens. With debt at as much as 90 percent of project cost, the loan sits within about ten points of the asset's value, so a tenth off a finished project exhausts the sponsor's equity and everything past it belongs to the lender [4][21]. "You do not need a big shock to the demand for that debt to be in trouble," Van Nieuwerburgh said [5].
KeyCorp is the lender the American Banker account puts numbers to [22]. Utility loans are 5.3 percent of its $191 billion of assets [7][19]. The almost-40-percent rise over three years works back to roughly $7.2 billion in 2023, about $2.9 billion added since [20]. KeyCorp declined to comment for the story [10].
Foran does not think the utility loans are the danger. Utilities have historically been a low-risk sector, he said, because their revenue is regulated and a certain amount of demand is guaranteed, though removing the AI element would likely take away the explosive growth [14]. "I want to say they'd be okay," Foran said. "You never really know until the tide goes out" [15].
Permission is the other route to a loss, and it has nothing to do with AI revenue. "I think many of the technology companies have been deeply surprised by the level of opposition and the speed at which that opposition has mobilized," Jeremy Fisher, principal advisor for climate and energy at the Sierra Club, said [18]. "If there's a clamp-down on new data-center construction, but you've already started, and you've already issued the debt to finance it, now you're stuck," Van Nieuwerburgh said [17].
In my view the utility leg is the sturdier half of this: a rate-regulated borrower with guaranteed volume is a different credit from a single-purpose project financed at 90 percent of cost [14][4]. The counter-argument is Foran's own, that taking AI out of the picture removes the explosive growth the utility books were built on [14]. Two things would show the gloomier reading wrong. Per-bank disclosure putting data-center books well inside the concentration limits Van Nieuwerburgh describes [3], and project structures where equity runs well above a tenth of cost [4]. Until then the exposure is one professor's sector-wide aggregate of mortgages, syndicated loans and project debt [1], and he expects it to keep rising as the buildout continues [2].
What to watch
- Whether banks start breaking out data-center and utility concentration in quarterly disclosure; KeyCorp declined to comment on its book.
- Whether local opposition hardens into construction moratoria that hit projects where the debt is already drawn.
- Whether cheaper Chinese AI alternatives compress pricing enough to test the 10 percent equity layer Van Nieuwerburgh describes.