In April a data center bond backed by a Microsoft lease priced at 5.7%. In August the same issuer's next bond, same tenant, priced at 7.2%. Credit markets are pricing a risk that equity markets are not.

Nvidia reported last night. Every AI conversation this morning is about that print.

Almost nobody is talking about this: in April, a QTS data center bond backed by a Microsoft lease priced at 5.7%, three times oversubscribed. In August, the same issuer's next bond priced at 7.2%, junk-like territory, same tenant. Oracle just got cut to BBB- after Wisconsin regulators demanded a $7 billion collateral requirement on one of its data center campuses. KKR and Blackstone have both walked from data center deals over insurance gaps that big insurers can't yet price.

None of that is speculation. It's from a Financial Times investigation published yesterday, August 26, into how Wall Street is financing the AI buildout, and it's the more important story.

The AI buildout runs on take-or-pay compute contracts, financed today against the assumption that demand keeps compounding. That's structurally a loan collateralized by future cash flow, and loans like that don't deflate slowly when the assumption wobbles. They snap, the way subprime did in 2008. CreditSights already projects AI compute supply outstripping demand after 2029, with no playbook yet for unwinding a distressed data center.

This doesn't mean the trade is wrong. Power and grid capacity are a real bottleneck, and Nvidia's revenue visibility into 2030 is backed by actual contracts. But the trade being right and the financing under specific projects being wrong can both be true at once. Not every railroad built in the 1870s boom was a bad investment. The ones funded with mismatched-duration debt are the ones that got stranded.

Credit markets are already pricing this. Equity markets aren't yet. Watch which hyperscaler-backed bond prices next, and whether it prices tighter or wider than QTS did in August.

Thoughts?