The Giant AI Bubble That Couldn't Pop

I. The Man Who Needs a Confession

In November of last year, Michael Burry closed the door.

A week after a regulatory filing revealed his last big position — put options on Nvidia and Palantir, around $1.1 billion in notional value1 — he deregistered Scion Asset Management with the SEC, the regulatory equivalent of switching off the lights and walking out. He returned his investors' money, told them his sense of value had fallen out of step with the market, and resurfaced days later as a paid Substack under the name Cassandra Unchained. The man Michael Lewis turned into a folk hero, the one who saw subprime when the rest of the room saw a housing market, had looked at artificial intelligence and seen 2008 again. Then he went quiet… seemingly.

It was a beautiful piece of theater, and I want to be careful here, because Burry isn't a clown and this isn't a takedown. It's an object lesson, and the lesson is larger than the trade. His central accusation is correct. He says the hyperscalers depreciate Nvidia GPUs over five and six years when the chips are commercially stale in two or three, that this understates depreciation by something like $176 billion between 2026 and 2028, and that the understatement inflates earnings across the industry.2 The practice is real; the filings are public. Amazon went the other way in early 2025, shortening some server lives and taking a $700 million hit — a company that looked at the same chips and declined to tell itself the comfortable story.3 The setup rhymes with the telecom overbuild of the late 1990s; Nvidia is Cisco rather than Enron, a supply boom mistaken for permanent demand. He may be right about every brick in the building.