Bankers Report Business Loan Demand Swung 28 Points Negative as AI Capex Pulls Credit to Nonbanks

This is the first round since early 2025 in which banks reported weaker C&I demand across every borrower size. For large and middle-market firms the net balance went from +16.1% to -11.7%; for small firms from +3.6% to -11.9%. A 27.8-point reversal in a single quarter is not soft-data noise. It is the size of swing that turned 2026's loan recovery into a stall.

Banks did not say they lacked willing credit. The two reasons they gave for weaker demand were declining customer investment in plant and equipment and customer borrowing shifting to more attractive bank or nonbank funding sources. The second reason is the new one: the last time the survey showed softening commercial loan demand, in the second quarter of 2025, banks did not cite migration to other lenders at all.

Here is the tension in the data. Equipment investment in the national accounts grew at a 20.9% annualized rate in the third quarter, adding more to GDP growth than in any quarter in six years. That is the same three months in which banks reported weaker demand for the loans that fund plant and equipment. Both statements can only be true if the capex boom and the bank borrowers are different companies.

For the part of the economy that genuinely depends on bank credit, the numbers are close to stagnant. Community banks — which still hold 37% of the industry's $717 billion of small business loans on just 14% of total industry loans — grew small-dollar C&I loans only 0.1% in 2025, down from 0.4% in 2024. Commercial and industrial lending is 56% of that small-dollar book. This is the credit that funds payrolls, dealerships and contractors, and it is not being repriced out of existence; it is barely growing.

Every displaced borrower that ends up in private credit is, at the end of the chain, still a bank liability. Banks' lending to nondepository financial institutions grew 35.2% in 2025, compounding at 22.7% a year since 2010, and the Fed's own Financial Stability Report says loan commitments to private credit funds and BDCs kept rising through the end of 2025. Meanwhile the October survey shows banks tightening standards on that NDFI lending for a specific reason: deterioration in NDFI borrowers' financial strength.

That channel has a fault line. Perpetual-life BDCs and interval funds hold $425 billion of gross assets and $241 billion of net assets — about 20% of private credit assets under management — and their managers have been capping redemptions after outflows exceeded inflows in the first quarter of 2026. Banks are extending wholesale credit into a private credit system that is simultaneously being asked to hand cash back to its own investors.

This is where the crowding-out story becomes precise, and more useful. Bank C&I commitments to AI-adjacent industries have grown from about 9% of total commitments — roughly $250 billion — in 2015 to 13%, or about $450 billion, in late 2025. On the balance sheet that looks contained: average outstanding exposure to AI-adjacent industries is roughly 0.8% of total bank assets.

The stress shows up in capital, not assets. Across large banks, committed AI-adjacent C&I exposure runs at roughly 25% of tier-1 capital, while outstanding exposure averages about 9%. Every dollar of AI-adjacent commitments therefore ties up more than twice as much regulatory capital as a dollar of drawn loan, and capital — not funding — is what caps how much a bank can lend to everyone else.

Which produces the counterintuitive conclusion. The AI build-out is financed mostly where banks are not lenders: AI issuers' share of investment-grade bond issuance jumped from 2% in 2023 to 10% this year, and the $330 billion issued year to date is ten times the whole of 2023. Direct bank data-center lending is small — $14.9 billion in a year — and the Kansas City Fed still describes bank data-center construction loans as a modest share of CRE lending. AI capex does not crowd out bank loans by consuming bank balance sheets. It crowds them out by pricing the risk-free rate higher, by pulling the best corporate credits into thin-spread bond markets, and by consuming the capital that would otherwise fund everyone else.

The policy backdrop makes this worse rather than better. The Fed held its target range at 3.75% to 4.00% in October after a quarter-point hike in September, with inflation at 3.7% year over year in September, up from 3.4% in August, and markets put a 76% chance on another quarter point at the December 8–9 meeting. A hike into softening credit demand and tightening standards prices every borrower that just lost its bank channel higher, at the same moment the alternative channels are gated by redemption caps.

The longer-run risk is the one the survey cannot show. If private credit's semi-liquid channel stays redemption-gated, displaced borrowers have fewer destinations than the SLOOS implies and return to banks just as bank standards on non-AI credit tighten. And if the AI capex cycle disappoints, the assets that unwind first are the levered ones: neoclouds carrying 85% debt-to-assets and project vehicles that can borrow 90% of their assets, funded in a bond market where AI issuers now take 10% of investment-grade issuance.