Ken Griffin Says AI Chips Are Like Airline Seats. He Might Have a Point.
Citadel's Ken Griffin thinks a glut of AI chips could send GPU pricing into freefall, and he's using an airline seat analogy to explain why. The comparison is sharper than it sounds, and it raises real questions about utilization, depreciation, and who survives the next AI downturn.
I nearly scrolled past Ken Griffin's latest comments. Then I hit the airline seat comparison and stopped.
Griffin, the Citadel chief, said a glut of AI chips could send pricing into freefall. His analogy: airline seats still earn money even when they're sold cheap, because an empty seat earns nothing at all. GPUs, he implied, work the same way. Once you've bought the chip, the only sin is leaving it idle. So you rent it out at whatever the market will bear.
That's a sharper observation than it first sounds.
The Airline Seat Problem
Here's the mechanics most AI bulls skip past. A GPU isn't an asset that appreciates. It's a depreciating machine with a useful life measured in a handful of years, and its value is almost entirely a function of how many hours it runs at capacity. At the peak of the shortage, an H100 reportedly changed hands for around $40,000, and cloud rental rates for the same card ran north of $8 an hour. Those numbers only hold up when demand outruns supply.
The moment supply catches up, the math flips. Operators who've already sunk capital into the hardware have every incentive to fill idle hours at marginal cost, exactly like an airline dumping unsold seats the night before a flight. Pricing doesn't glide down. It falls through the floor, because the alternative is earning zero on a machine that's losing value every day it sits dark.
Griffin didn't predict a collapse. He said bubbles almost certainly exist within AI, and he's right that they probably do. But he also made the point that companies emerging from a bubble tend to lead their markets for decades. Both things can be true at once. That's what makes the quote interesting rather than just bearish.
Who Survives a Bubble
The dot-com wreckage is the obvious precedent. Amazon survived. Pets.com didn't. The difference wasn't the thesis. It was unit economics and staying power.
So what does a GPU glut mean for everyone downstream? For the hyperscalers, cheaper compute is a gift, since it lowers the cost of training and inference. For the smaller labs renting capacity, it's oxygen. For Nvidia, it's the first real test of whether its margins are a moat or a cyclical peak. And for the startups that built their pitch on access to scarce chips, a glut removes the scarcity and the pitch with it.
The question worth asking: if a GPU is only as valuable as its next hour of utilization, what does that say about the depreciation schedules the big cloud providers are running? I'm not entirely convinced those numbers are as conservative as investors assume.
What I'd Watch
Look, I've covered enough hype cycles to be careful about calling tops. The AI buildout is real, and the demand for compute isn't going away. But the chip bulls are pricing in permanent scarcity, and that's a bet I wouldn't make.
Three things I'd track from here. First, hourly GPU rental rates on the secondary market, which are the cleanest read on real demand. Second, utilization disclosures from the hyperscalers, if they ever offer them. Third, Nvidia's gross margins over the next few quarters, because that's where a glut shows up first.
Color me skeptical, but the airline seat analogy is doing more work than most people realize. Seats fill planes. They don't hold price.
