Why Citi Is Backing Microsoft's Balance Sheet in the AI Debt Squeeze

CoreWeave lost 15% of its market value in a single session last week. Read the headlines and you'd assume AI demand is cracking. Citi's Tyler Radke says you'd be reading it wrong.

Radke, co-head of US software research at the bank, blames rising bond yields for the neocloud selloff, not a slowdown in orders. His evidence is pretty clean. Rental rates for Nvidia's H100 chips are sitting near record highs. That's the tell. When appetite actually softens, spot rental pricing is the first thing to roll over. It hasn't.

And yet CoreWeave's bonds were yielding close to 15%. Fifteen percent, on a company renting out the most sought-after compute on earth.

The setup is strange on its face. Record rental pricing is a bullish signal. A 15% yield is a distress signal. Both are true at the same time, which is exactly the kind of contradiction that forces an analyst to pick a side.

Neoclouds rent Nvidia silicon to AI developers, and they buy that silicon with borrowed money. Admittedly, that's the simplified version. But it's the version that matters when credit markets get twitchy. Higher yields mean a higher cost of capital, and a higher cost of capital makes the math on every new GPU cluster worse.

The Balance Sheet Advantage

That's why Radke ranks Microsoft among his highest-conviction ideas, and it has almost nothing to do with who builds a better GPU. It's about who can write the capex check without asking the bond market for permission.

Microsoft's AI infrastructure spend runs into the tens of billions a year, and almost none of it's financed at 15%. CoreWeave has to reprice its debt every time it wants to add capacity. That's the entire thesis in one sentence.

In a calm credit market, the gap is a footnote. In this one, it's the whole argument.

I'm not entirely convinced the selloff is purely a rates story, though. There's a second force at work. Skeptics are pricing real risk around chip depreciation, customer concentration, and contract duration, and none of that disappears if yields quietly drift lower. Proponents of the neocloud model will tell you those risks are overstated. Maybe they're. But the market is charging for them either way.

Here's the part Radke gets right. The question worth asking: is the market repricing neocloud risk, or is it just repricing duration? Those aren't the same thing, and the distinction matters. If H100 rental economics are intact, and the pricing data says they're, then this is a financing problem wearing a demand problem's clothes.

What to Watch

Three numbers matter from here. CoreWeave's bond spreads, the spot rental rate for H100s, and Microsoft's capex guidance on its next earnings call.

If H100 rental rates hold near records while CoreWeave's spreads keep widening, Radke's call is basically confirmed. The squeeze is about debt, not demand, and Microsoft wins by simply existing with a fortress balance sheet.

If rental rates start slipping, though, that's a different article entirely. Then the neocloud selloff was telling the truth all along, and bond yields were just the messenger.

To be fair, there's a middle path. Yields ease, CoreWeave refinances, and the whole episode looks like a bad month in credit rather than a verdict on AI infrastructure. That's the outcome the bulls are quietly hoping for. Color me skeptical, but I'd want to see spreads tighten before I buy that version.

Time will tell, though. For now, Citi is siding with the balance sheet, and the reasoning is hard to fault.