Jim Chanos says AI is "much worse" than dot-com. The reason is a dollar going in a circle

Jim Chanos owns this warning. He's the short-seller who called Enron's collapse before the rest of Wall Street had a clue, spent years pressing Tesla, and he keeps telling anyone who will listen that the AI trade is not just a bubble but the opening of a "diamond or platinum" level of fraud, worse than the dot-com bust. Read that headline in a nervous market and it's easy to land on one of two reactions: it's a perma-bear crying wolf again, or it's a signal to get out. Both miss the point. The warning isn't a crash prediction. It's a description of an accounting loop that is running through the AI trade right now — and once you know what a circle looks like, you can see it yourself.

Call it the AI circular economy. The same money, going around and coming back as revenue.

The dollar that circles back

Here's the loop. A chipmaker like Nvidia, or a cloud giant like Microsoft, invests billions into a hot but money-losing AI startup — an OpenAI, a CoreWeave, a Nebius. As a condition of that investment, the startup commits to spending the capital on the investor's chips or cloud. The investor then books that internal spend as revenue, the startup points at the partnership to justify a higher valuation, and the investor marks up its own equity stake. Chanos put it bluntly on the circular-financing claim he and Michael Burry have been hammering: Nvidia, he said, is "putting money into money-losing companies in order for those companies to order their chips."

Nothing about that loop is illegal on its face — related-party transactions happen all the time. What makes Chanos reach for the Enron and Lucent comparison is what the loop does to the accounting. Nvidia denies the scheme, sending analysts a seven-page memo insisting it doesn't offer the long-term vendor financing that sank Lucent and noting its customers pay within about 53 days. Burry isn't convinced, and neither is Chanos. The point isn't that either of them has proved fraud. The point is structural: when a seller's demand is financed by the seller, "demand" is no longer evidence of end customers wanting the product. It's a bookkeeping loop that keeps the headline growth coming until the financed buyers can't pay — which is exactly how Lucent, and Enron itself, ended.

Revenue today, cost tomorrow

The second half of the trick is timing, and it's where the numbers get tangible. Chip suppliers recognize their revenue the moment they sell. The hyperscalers on the buying side capitalize — they carry the cost as an asset on the balance sheet and spread the write-off over years instead of expensing it today. So you get a portrait of a booming industry painted in front-loaded revenue while the bill quietly piles up out of the income statement's sight.

The most visible version is the finance-lease carve-out. Under modern lease accounting, principal repayments on finance leases land in financing activities rather than the investing or operating lines, so a company can pile on massive infrastructure obligations while its headline free cash flow stays high. Microsoft's finance-lease liabilities alone jumped to about $46 billion in fiscal 2025 from roughly $27 billion the year before — nearly $20 billion in infrastructure spending that never hit the expected capex line. Oracle booked around $4 billion in finance-lease liabilities in a recent quarter after sitting near zero a year earlier. This is the plumbing under the "AI boom": revenue recognized now, the cost pushed into the future, on a balance sheet someone will eventually have to answer for.

The fragility underneath is easy to measure. One MIT report Chanos cites found only about 5% of AI pilot programs generated immediate revenue; the other 95% fell short of expectations. Read that against the scale of the buildout and you see why he frames the whole thing as a pre-fraud environment rather than a verdict. Cheap capital lets everyone finance everyone else's purchases and push the reckoning down the road. It works — briefly. It worked at Lucent; it worked in the years before Enron's special-purpose vehicles stopped holding water.

What you do with it, not just what you believe

None of this tells you when anything breaks, and Chanos — who has been warning about the "most dangerous months" of August, September, and October — would be the first to admit he doesn't have a stopwatch. Short-sellers make their living listening to their own delivery dates, and this one is a fraud-cycle man, not a soothsayer. His core framework is that extended bull markets erode skepticism, so fraud follows financial success with a lag. That's a lens, not a forecast.

The same eye, by the way, is why he turned up in crypto — not to short Bitcoin itself, but to short the packaging. He ran long bitcoin against short MicroStrategy (now Strategy), betting that the stock's premium to the net asset value of the bitcoin it held couldn't hold, calling these "bitcoin treasury" structures SPAC-style risk. He's not reflexively against the asset; he's against paying a marked-up price for a wrapper around it. That distinction is the entire lesson.

So here's the takeaway that survives the name. When you look at an AI growth story — or any growth story, crypto or otherwise — the single most important question is not "is this exciting" but "who is ultimately paying for the capex?" If the answer is a genuine third party with real cash flow, the demand is real. If the answer is the sellers financing their own buyers in a loop, then the "growth" is partly an accounting artifact, and you're holding the bag on a timing game with someone else's stopwatch. Chanos's warning is valuable not because he's about to be right next quarter, but because he's giving you the specific entry to check before you buy. Check that one, and you'll out-skeptic most of the crowd without ever shorting a share.