Michael Burry has released a few of his recent newsletters for free. And boy of boy is one of them about AI. This line comes toward the end of the newsletter, right before he embeds a clip of himself as portrayed by Christian Bale in The Big Short where Christian Bale as Burry says he wants to short the housing market:

This is the exact structure of the residential mortgage-backed securitizations I shorted two decades ago.

I suppose if you are the bear of bears, you kind of just keep looking for more of what you know, and in this newsletter, Burry really hammers it:

Reliance on only a few customers:

Cisco never had even one 10% customer. Cisco needed a more correlated pull-back to be hurt. NVIDIA is hurt badly if just one customer pulls back, or even if that one customer does not grow its orders as fast as NVIDIA needs it to.

Then a lengthy aside about the etymology of the term "the bezzle":

Alone among the various forms of larceny [embezzlement] has a time parameter. Weeks, months or years may elapse between the commission of the crime and its discovery. (This is a period, incidentally, when the embezzler has his gain and the man who has been embezzled, oddly enough, feels no loss. There is a net increase in psychic wealth.) At any given time there exists an inventory of undiscovered embezzlement in—or more precisely not in—the country’s business and banks.

This inventory—it should perhaps be called the bezzle—amounts at any moment to many millions of dollars. It also varies in size with the business cycle. In good times, people are relaxed, trusting, and money is plentiful. But even though money is plentiful, there are always many people who need more. Under these circumstances, the rate of embezzlement grows, the rate of discovery falls off, and the bezzle increases rapidly. In depression, all this is reversed. Money is watched with a narrow, suspicious eye. The man who handles it is assumed to be dishonest until he proves himself otherwise. Audits are penetrating and meticulous. Commercial morality is enormously improved. The bezzle shrinks.

And a Charlie Munger quotes on this:

Galbraith coined the “bezzle” word because he saw that undisclosed embezzlement, per dollar, had a very powerful stimulating effect on spending. After all, the embezzler spends more because he has more income, and his employer spends as before because he doesn’t know any of his assets are gone. But Galbraith did not push his insight on. He was content to stop with being a stimulating gadfly. So I will now try to push Galbraith’s “bezzle” concept on to the next logical level.

But it's not like Burry thinks AI is the all a charade or something, just the demand:

The bezzle is not that AI is fake (though I may argue it is inadequately defined). Rather, the bezzle is the temporary demand – benchmarking, trace-harvesting, leaderboard hopping – that is being financed and tallied as if it were permanent demand, or an indication of even greater demand.

He thinks that the current demand for AI compute is already peak demand, and it's unlikely to see much growth from here, for many reasons, one of which is that people just...like...aren't going to do very much useful with it:

Four decades of Microsoft Excel and it is on nearly every PC in the world, yet people very rarely use even 1/100^th of its capabilities even though it is sitting right there in front of them, again, for decades. The same fate awaits AI compute. Vast capacity, with humans only caressing the possibilities.

Which is a problem if you are banking on charts that go up and to the right because you are otherwise more like a ponzi than not:

And then there's this whole part how a lot of the funding for the great data-center build out is actually kinda being packaged up and rolled out in a manner that looks a little risky, like housing in 2008:

The question at the end of the day is whether those loans are good loans, and I just do not see a way that they are good. They are structured with the hyperscalers as sponsors who provide a loose backstop for the loans on their data centers. Sponsor support agreements, however are not strong backstops. There are outs for construction delays, power-delivery issues, and even short outages can trigger nonpayment of rent.

To sum it up, Burry suspects that the current AI market has got the wrong signal:

The bezzle is not that AI is useless. The bezzle is that the market is capitalizing the most expensive phase of AI adoption as if it were normal and indicative of future demand. Benchmarking is not production, training is not performance. And tokenmaxxing is not a dependable infrastructure demand signal.

Which means:

So, yes, son, they are flying those airplanes around empty. The only time that made sense was at the beginning of flight. When pilots need hours, engineers need data and regulators need time. That benchmarking phase, all lusty fire and brimstone, needed at the very beginning of anything at all worthwhile.

The letter is long and I'm sure people with a better background in finance will do justice to it where I have only pulled out the glamorous bits. I will say that Burry has felt a bit like a broken clock hoping for his time to be right. But of course this doesn't mean he isn't right.