A May 22 conversation between investor Steve Eisman and AI researcher Gary Marcus, resurfaced on June 10, gives a falsifiable framework for whether GenAI's demand model holds, and which firms get hit first if it does not.
The AI industry's current growth story rests on a specific, testable bet: that customers will keep paying for tokens at prices high enough to cover the cost of running the models. A May 22 conversation between investor Steve Eisman and AI researcher Gary Marcus, which Marcus published on Substack on June 10 under the title "Breaking news, and how the end might begin", lays out why that bet might not survive the next twelve months of earnings.
Eisman, the investor who predicted the 2008 subprime crisis and the basis for Steve Carell's character in The Big Short, spent the interview asking Marcus the question he says he is now asking of every GenAI executive: can the providers actually charge for tokens, and reach profitability, within roughly a year. Marcus, a long-standing skeptic of the current AI scaling thesis, treats that as the right question. The two disagree on probability, but agree on mechanism. The break, if it comes, will not arrive as a sudden technical failure. It will arrive as a credit event in the customer base.
The subprime parallel does analytical work, not decorative work. In 2007, mortgage credit quality did not collapse because houses stopped existing. It collapsed because the marginal buyer stopped buying the paper, and the leveraged suppliers found themselves holding inventory priced for a market that no longer existed. The AI analog, as Marcus frames it, is when the marginal enterprise customer decides that token pricing no longer justifies the use case, and the highest-burn model providers run out of patient capital before they can cut prices enough to win the demand back.
That second clause matters more than the first. The hyperscalers, Microsoft, Google, Amazon, and Meta, have the balance sheet to absorb a pricing reset. They can keep running inference at a loss while they wait for unit economics to catch up, or they can drop prices and force smaller competitors out of the market. The capital-hungry GenAI labs do not. Marcus argues that OpenAI, by his accounting, has the thinnest cushion of any major AI firm, and that a stumble there would be the kind of event that makes a credit-quality question go from theoretical to urgent for the rest of the sector.
A few caveats are worth naming. The "breaking news" Marcus alludes to in his post title is not specified in the excerpt that has been published. Marcus flags the interview as "now supremely relevant" without naming the trigger, and a portion of the first snippet is paywalled in a separate video. The Eisman quotes that have circulated are a single-witness secondary citation through Marcus, so any specific phrasing should be read as his transcription of a conversation rather than as a direct Eisman statement. Marcus's framing of OpenAI's "stumbles" is also his own characterization, not a confirmed industry event. None of this invalidates the framework. It does mean the framework should be tested against upcoming earnings, capex guidance, and token-pricing decisions, rather than treated as confirmed fact.
The framework itself is simple enough for a reader to apply. The questions to ask on the next round of AI earnings are: which providers raised prices, which held, and which cut. Which companies announced new capex commitments, and how they plan to fund them. Which enterprise customers disclosed AI unit economics, and whether the gross margins were the ones the bulls modeled. If the answers to those questions trend the wrong way for long enough, the marginal buyer in the AI market does what the marginal buyer did in 2007. Stops buying.
The interview is one informed voice, not a forecast. But it identifies the mechanism the AI bulls are not yet modeling, and the calendar on which that mechanism would start to show up in the numbers. That is more useful than a prediction, and it is what the next quarter of AI earnings will quietly be judged against.