Cheap GPUs Are Great News, Unless You Financed Them
GPU rental rates have fallen hard, squeezing the AI hosting companies that borrowed billions against their chip fleets. A new set of hedges is emerging to manage the pain, and lenders finally have a reason to care.
What happens when the machines you financed earn less each month than the debt you used to buy them?
That's the question sitting in every AI hosting boardroom right now. The companies building AI apps don't have to own hardware anymore. They rent it by the hour. And the hourly rate keeps sliding.
Here's what matters: cheap compute is a gift to whoever's buying it and a headache for whoever's selling it. Both things are true at once, and the market has only priced one of them.
The numbers tell the story
Spot rental rates for Nvidia's H100 have fallen to roughly $2 to $3 an hour. A year ago the same chip cleared $8 in tight markets. Some brokers now quote under $2 for long-term reserved capacity.
Meanwhile the debt stacked against those chips keeps growing. CoreWeave has taken on more than $7 billion in financing, much of it underwritten with the GPUs themselves as collateral. Other neoclouds have followed with smaller versions of the same trade.
Do the math. If a host borrowed assuming $6 an hour and now collects $2.50, the interest doesn't care. The gap isn't measured in basis points. It's measured in whether the loan gets refinanced.
Why the math broke
This is a classic capital cycle, and it's playing out faster than most people expected. High rents in 2023 pulled in enormous supply. New chips shipped. New data centers came online. Supply caught demand, and price did what price always does.
Depreciation is the other half of it. Nvidia talks about four to six year useful life. Anyone who lived through the last two generations will tell you it's closer to two or three when the software keeps moving. On the risk side, that's the whole ballgame for a lender.
Who wants to lend against an asset that loses half its rental value in 18 months? Fewer people than six months ago, frankly.
The hedges arrive
So the market is building tools to manage it. GPU rental futures and index-linked swaps are showing up at a handful of desks. The idea is simple. A host locks in a forward rate on its capacity, the way an airline hedges jet fuel.
Traders are watching these contracts closely because they'd give lenders a real mark for GPU collateral. That's a big deal. Right now the marks are guesswork. The reality is nobody can agree what a two-year-old H100 is actually worth on a balance sheet.
What the street is missing: a functioning hedge market doesn't stop prices from falling. It just moves the pain from the operator to whoever took the other side.
What to watch
Three things. First, spot H100 pricing. If it breaks below $1.50 an hour, watch for restructuring headlines within a quarter. Second, Blackwell delivery volumes through mid-2025. Every new chip that lands puts more pressure on the older fleet. And third, hyperscaler capex guidance on the next earnings calls. If Microsoft, Amazon, and Google hold spending steady, the demand side is fine and this is just a repricing.
If they blink, the hosts are on their own.
My conviction is that cheaper compute wins for the application layer and the users. It always does. The pain lands in the middle, on the balance sheets that bet on yesterday's rates. That's not a crash thesis. It's a repricing, and the hedges are how the smart money is positioning for it.
