Hey Traders,

GPU rental prices just hit a record.

Rental rates for Nvidia's newest chip, the B300, rose 17 days in a row to an all-time high of about $7.45 an hour, up roughly 62% from a year ago.

Most people read that as proof AI demand is endless. It might not be.

A price goes up when buyers want more than sellers can deliver. That can happen because demand jumped, or because supply got stuck. Right now, it's the second one.

A couple years ago the hard part of AI was getting chips. Today, it's getting the power, the permits and the buildings to plug those chips into. Microsoft's CEO has said he may end up with chips sitting in inventory he can't plug in. Last month Oracle invoked force majeure on a $165 billion data center campus in New Mexico, citing delays in securing power, after the project ran into a denied energy permit and local opposition.

In the same week GPU rents hit a record, CoreWeave's 2032 bonds fell below 80 cents on the dollar, with yields above 15%. If demand alone were driving prices, companies that rent out GPUs would be flush and their borrowing costs would be falling.

Instead lenders are worried about the billions borrowed for data centers that aren't finished or powered yet. That debt costs money every month, and an unfinished building earns nothing.

The takeaway: record GPU prices aren't automatically good news for everyone in the AI buildout. If the problem is that projects can't get built, the winners are those who already own running capacity and power, along with companies selling on-premise hardware like DELL and HPE as businesses decide to own their own machines.

The ones at risk are indebted operators, lenders to stalled projects, and builders whose backlogs keep getting pushed out.

The AI demand story is intact. The buildings, the power and the debt are where the winners and losers get decided. This isn’t going to be easy.

Here for a good time… AND a long time,

Hans