US Banks Have Nearly $1 Trillion Tied to AI. Here's Where It Would Crack
The Federal Reserve's November Financial Stability Report shows the largest US bank holding companies carrying close to $1 trillion in exposure to the firms behind the AI buildout. Capital ratios look healthy. But the share of market contacts flagging AI as a risk jumped from 50% to 87% in two quarters, and that gap is the real story.
How much of the AI boom is sitting on bank balance sheets? Close to $1 trillion. That's the exposure the largest US bank holding companies carry to the firms behind the AI buildout, according to the Federal Reserve's November Financial Stability Report.
Read that again. One asset theme. One trillion dollars. Concentrated across a handful of institutions.
Here's what matters: the Fed doesn't think this is a problem yet. The report describes the banking system as sound, with high capital ratios and manageable loss expectations. All true, and all measured against today's cash flows.
The Number
The figure that should get more attention sits alongside it. 87% of market contacts surveyed by the New York Fed named AI as a risk. In the spring, that number was 50%.
The numbers tell the story. A two-quarter jump from half to nearly nine in ten isn't a shift in mood. It's a shift in positioning.
And the trillion-dollar figure deserves a closer look. Banks aren't equity investors in this trade. They're lenders, underwriters, and counterparties. The exposure runs through credit lines to data center developers, construction loans, project finance for power generation, and revolving facilities for the suppliers feeding the hyperscalers.
Different instruments. One underlying bet: that demand for compute keeps compounding at the pace of the last eight quarters.
Why the Read Is Split
From a risk perspective, the issue isn't any single borrower. It's correlation. When every lender underwrites to the same growth curve, a modest slowdown in capex gets magnified across the whole book. We watched this movie with commercial real estate. Different collateral, same mechanics.
This collateral is worse in one respect. It depreciates faster. A data center commissioned in 2026 has a shorter economic life than an office tower, because the silicon inside it turns over every few years. Lenders are pricing loans against assets that lose value on a technology cycle, not a real estate cycle.
My take: the system can absorb a slowdown. It can't easily absorb a repricing. Those are different scenarios, and only one of them gets stress-tested with real severity.
The Other 13%
So who isn't worried? Roughly one in eight market contacts. Some of that's genuine conviction about AI revenue. Some of it's the usual reluctance to call a top while the fees are still flowing.
The survey shift happened fast enough that it reads less like analysis and more like recognition. Traders are watching credit spreads on data center securitizations and the private credit funds that have quietly absorbed a lot of this paper. Neither shows real strain yet.
That's the point. Nothing cracks while the capex keeps coming.
What to Watch
Three things, in order of when they'll tell you something.
First, big bank Q4 earnings in mid-January. Listen for provisioning language around commercial and industrial loans, and any commentary on data center credit. Second, hyperscaler capex guidance on those same calls. If the four largest buyers trim their 2026 plans, the exposure math changes within a quarter.
Third, the Fed's next Financial Stability Report, expected in the spring. Watch whether the $1 trillion grows, and whether the report starts breaking out the mix of loans versus capital markets commitments. That detail matters more than the headline.
What the street is missing: this isn't a bubble question. It's a duration question. The AI buildout is real. The financing behind it's short-dated and reprices quickly. Banks are fine until the refinancing window closes, and nobody knows exactly where that window sits until it does.
