The AI bubble's fatal flaw.

Not a hot take. That is the view of Jim Chanos, the man who actually exposed Enron. He is looking at the AI industry right now and calling it the diamond or platinum level of fraud. Worse than the dot-com bubble. His reasoning is hard to dismiss: at least during the telecom boom, the companies buying the equipment were profitable. The companies driving today's AI spending are burning cash and betting everything on what these models might become. Not what they do today.

OpenAI and Anthropic are heading for IPO. The numbers being floated are extraordinary. But before anyone invests, pivots, or restructures their business around this narrative, one question is worth sitting with: is this a valuation, or a prophecy?

The Enron parallel nobody wants to say out loud

Chanos did something simple in 2000. He read the financial statements. What he found was that Enron was losing money on every dollar it borrowed, while Wall Street had decided it was a limitless tech innovator.

Sound familiar? Enron said it had reinvented energy. Today's AI giants say they are reinventing intelligence. In both cases the vision did most of the work. The actual financial mechanics were secondary. Until they were not.

The scaling story is breaking down

The entire trillion-dollar thesis rests on one assumption: feeding more data and compute into large language models will eventually produce superintelligence. Bigger in, better out. Indefinitely.

Ilya Sutskever, the co-founder of OpenAI and the person most responsible for proving scaling worked in the first place, said in late 2025 that the age of scaling is over. The easy gains are gone. These models can pass brutal academic exams and still struggle to fix a simple bug. They generalise dramatically worse than people do in real conditions.

That is not exponential progress. It is incremental progress in an exponential costume.

There is no moat

If scale is not the answer, the story shifts to defensibility. These frontier labs must have something competitors cannot replicate.

A recent MIT study says no. Free, open-source models perform at about 90 percent of the capacity of closed frontier models on release and close the gap within months. At a fraction of the cost. A free Chinese model called Kimmy is already in the same performance tier as GPT and Claude, beating flagship models from Meta and Amazon outright.

Even Microsoft, OpenAI's biggest backer, is exploring free models to manage its own compute costs. When your largest investor starts shopping for cheaper alternatives, the moat argument gets hard to make with a straight face.

Three things worth watching

Bubbles do not pop because the technology fails. They pop when paper wealth has to become real money. If this peaks after the IPOs, early investors cash out and the public holds the bag.

Does the scale story hold? Next-generation models should be genuine leaps, not incremental improvements with better press releases. If it is the latter, the valuation is a prophecy.

Does the moat hold? If open-source stays within months of frontier performance at a fraction of the price, there is no defensible position for companies charging frontier premiums.

Can the businesses support the spending? The debt taken on to buy rapidly depreciating AI chips is real. Watch for stress there.

There is an old Wall Street line: those who live by the crystal ball are destined to eat broken glass. The IPO window is opening. Look at the actual numbers before you walk through it.

By Michael Lennard Gnaedinger. © 2026 Gnaedinger Consultancy. All rights reserved.