Steve Eisman: Here's What Anthropic Isn't Telling You

In progressEd Elson (Prof G Markets), with Steve Eisman and Dan Chiolanpodcast2026-10-28read planted ai llmsbusiness

Synopsis — AI-drafted from Dan's notes

Two days after Anthropic’s S-1, Ed Elson asks Steve Eisman what jumped out. Two things, Eisman says. First, the long list of future compute commitments, $800 to $900 billion depending on what you count. Second, and more telling, the filing’s numbers stop at the second quarter. He tracks two figures per quarter, sequential revenue growth and revenue divided by compute cost, through the first half of 2026, when companies told engineers to use AI around the clock regardless of cost (“token maxing”). By his figures revenue per compute dollar climbed from below one late in 2025 to about 2.1 in the second quarter. Third-quarter figures reported since, from outside the filing, show growth slowing to about 45% and the ratio slipping to about 1.7 as buyers grew cost-conscious and Chinese open-weight models took share. He has also heard, without confirming it, that Anthropic cut prices by about half in October; if true, he says, it shows the business has no moat but price. Elson notes the S-1’s charts leave out the third quarter; Eisman calls that very intentional. Would he invest? No.

On governance, Eisman has little to say: public shareholders have no real rights, and that is disclosed. The structure gives one Class F share, held by the seven co-founders, 50.1% of the vote, while the Long-Term Benefit Trust’s Class T share elects most of the board. He compares it with Meta, whose stock has done fine anyway, and calls it tech’s founder worship. He thinks the filing raised investor confidence because few people noticed it ends in June. The other thing to watch, he says, is CoreWeave’s debt: with its bonds trading at yields near 15%, borrowing to build data centres, its whole model, looks prohibitive unless Nvidia steps in.

A second segment, with Prof G researcher Dan Chiolan, surveys autonomous vehicles. Waymo leads, with paid rides in 15 US cities and about 500,000 a week; Tesla is second but scaling fast, and its Cybercab costs far less to build. China leads on manufacturing, the US on commercial rides, and Europe and the Middle East are where the two will compete. US regulation is state by state, and public opinion of self-driving cars has soured along with opinion of AI. His picks-and-shovels ideas are US-made lidar, fleet servicing (where Hertz has set up a business), and trucking. Elson closes with three points on the S-1: the $80 billion loss is mostly a non-cash charge on convertible notes; first-half growth of about 1,300% has since slowed; and narrowing operating losses will not last once more than $800 billion of commitments, and new capital spending, reach the income statement.

Connections