AI Bubble: "Something profound has changed"

A 47-minute interview, uploaded on 16 October 2026, in which Isaac Pound of The Tech Report talks to Ed Zitron during the year’s third tech sell-off. Zitron’s claim is in the title: something has changed. The industry’s usual trick of putting the toothpaste back in the tube has stopped working.

He watches the debt markets, which he finds less emotional than stocks. CoreWeave’s bonds and loans were yielding roughly 15–16% in mid-October, so borrowing for it and every smaller neocloud now costs double digits. Gartner, which a year earlier was telling clients to embrace neoclouds, now says 80% of them will cease independent operations by 2030, and that AI data-centre owners need $8.5 trillion in cumulative revenue through 2030 to earn a 12% return. (That 80% includes firms that are bought or merged.) He cites a Goldman Sachs estimate of the AI revenue hyperscalers need to break even on 2026–27 capex as $360B; the reported figure is about $300B a year. His picture is a slow boil rather than an explosion: missed payments, debt that can’t be raised, depreciation on chips not yet in service.

Then Anthropic. A leak to investors ahead of the IPO showed $16.8B in third-quarter revenue, about 45% up on the quarter, with an adjusted gross margin of 85%, an adjusted operating margin above 11%, and no new annualised run rate. The last one, from July, was $65B, and investors had expected $100B or more by year end. He adds that 5–7% of revenue came from Chinese labs distilling Claude (since cut sharply), and that by outlets’ reckoning from the prospectus about $413B of commitments rests on Anthropic growing, with Google, Broadcom and smaller data-centre firms exposed. Ramp’s data shows 80% of business spending on OpenAI and Anthropic comes from 1% of firms, mostly subsidised AI startups. His general claim is that the bubble was inflated on made-up numbers, and every real number that arrives punctures it.

The hyperscalers, he says, hold a wolf by the ears. They can’t stop spending without being asked why they started, and they have no other growth story. Jefferies has said the most likely outcome is “massive capital destruction”. His big call is that Oracle dies. He dismisses Anthropic’s new self-replication research as a distraction. It showed small open-weight models, not Claude, copying themselves onto cheap devices under deliberately easy conditions. He ends on inflation: AI demand drives up memory prices, which interest rates can’t reach.