CoreWeave's Plunge Shows How Rising Yields Test Debt-Funded AI Compute

Demand for data centre space across North America is close to record highs, and CoreWeave's shares and bonds fell sharply anyway. Vacancy in major hubs has been running unusually low. The customers are there. What investors started pricing was the cost of the money underneath the machines.

That gap is the useful part of this story. A business can sell everything it builds and still struggle if its debt costs more, for longer, than its assets can earn back. Rising bond yields test that math for anyone running or backing AI compute, and for a property investor in Edmonton carrying a rental against a variable rate. The question in each case is whether the asset produces enough cash, fast enough, before the loan comes due or reprices.

Demand risk and financing risk are two different bets

CoreWeave grew because the big cloud providers couldn't get enough Nvidia chips into customers' hands during the AI rush. It filled that gap by borrowing heavily, including a $7.5 billion debt facility in May 2024 to buy hardware. The company has used its chip inventory as collateral for billions in financing.

Borrowing to grab market share before competitors arrive is a deliberate move, and it can work. It also means the company is making two bets at once: that demand holds, and that credit stays affordable. The recent selloff hit the second bet. When you look at any AI infrastructure name, score these separately. A company can be right about customers and still be squeezed by its lenders.

The collateral has a short clock

A house or an office tower used as collateral tends to hold much of its value for decades. A GPU does not. When a faster, more efficient generation ships, the older chips lose resale value and earn less per hour of rented compute. A lender holding those chips as security knows that, and prices the loan accordingly.

So the debt has to be paid down quickly or rolled into new financing. Here is where yields bite. As an illustration only: refinancing $1 billion at 7% costs about $70 million a year in interest. The same $1 billion at 9% costs about $90 million. That extra $20 million has to come from compute revenue on hardware that's worth less each quarter. Higher-for-longer rates don't change how fast the chips age. They raise the cost of every refinancing that ageing forces.

The lesson travels well: match the life of the loan to the useful life of the asset. Short-lived assets financed by debt that must keep rolling over are exposed to whatever rates happen to be when the roll arrives.

Test at the rate you might pay

Canadian mortgage rules offer a tidy model here. Under OSFI's stress test, in place since 2018, a borrower must qualify at the higher of their contract rate plus 2 percentage points or 5.25%, even though they actually pay the contract rate. The point is to prove the payment still works if rates move against you.

AI infrastructure has no regulator forcing that discipline, so investors and operators have to apply it themselves. Run the model with financing costs a couple of points higher than today's. Shorten the assumed useful life of the hardware by a year. If the project only works under the original assumptions, the market will eventually find that out at a worse moment than you would.

This matters close to home too. Alberta's cheaper land and energy supply make it a natural candidate for data centre projects as other hubs strain for power. The projects that get built and last here will be the ones whose financing was tested against harder conditions than the ones in the pitch.

Investors are now asking AI companies to show sustainable debt management alongside growth. Before you back any compute build, ask how the numbers look at two points higher and one year shorter.