Anthropic IPO: “This is a dog of a company”
Dan PetersonSynopsis — AI-drafted from Dan's notes
A 38-minute interview, with Will Guyatt hosting, recorded as Reuters was publishing its stories on Anthropic’s leaked IPO prospectus. Zitron’s starting point is that the filing ends a story. Anthropic was supposed to be the stable lab, and in 2025 it lost $8 billion on $4.6 billion of revenue, about as bad as OpenAI at the same scale. What struck him most is the compute: about $518 billion committed over seven to ten years, roughly 80% of it non-cancellable.
He then goes through where the money comes from. Two customers made up a quarter of 2025 revenue; he guesses GitHub Copilot and Cursor. Nearly half of revenue is Amazon and Google reselling the models. Most of the paying customers, in his telling, are AI startups that lose money and are funded by venture capital, so the only new money entering the system is venture capital and debt. Outside the two big labs he can find about $22 billion a year of demand for rented GPUs. He puts all AI revenue at the big cloud companies at about $183 billion, against a Goldman Sachs estimate that they need $308 billion a year just to break even on two years of spending.
Two puzzles follow. Why do the cloud companies keep buying chips they cannot plug in? He says nobody has told him, and guesses that Nvidia ties future supply to buying now. And how does the debt get paid? Components cost more, so builders borrow more, which raises the price of borrowing, with the ten-year Treasury above 5% as the floor: a doom loop. Oracle is a notch above junk, and a downgrade would force funds to sell its bonds.
On the IPO itself he says a $2 trillion listing leaves buyers little room, that the company would need $50 billion to $100 billion a year in new debt, and that an investment-grade rating would be a scandal. His verdict gives the episode its title.
The prospectus figures, the customer concentration, the resale share and the planned November listing all match the reporting. Some of the rest does not. Google went public in 2004 at about $23 billion, not the “hundred billion” he gives. Anthropic’s compute deal with SpaceX is up to $84.5 billion, not $48 billion. Oracle is one notch above junk at S&P only. The $183 billion and $22 billion are his own sums from analyst notes, and the Goldman figure is reported elsewhere as “about $300 billion”. He flags two of his claims as unconfirmed himself: GPUs failing within six months, and the Nvidia ultimatum.
Connections
Links
- https://www.youtube.com/watch?v=XLXn0Ut8adM
- https://www.investing.com/news/stock-market-news/anthropic-prospectus-shows-2-trln-ipo-ambition-despite-huge-loss-reuters-reports-4921435
- https://www.implicator.ai/anthropic-ipo-prospectus-518-billion-compute-commitments/
- https://www.calcalistech.com/ctechnews/article/r8sh14ux4
- https://www.wheresyoured.at/dead-money/
- https://www.investing.com/news/stock-market-news/hyperscalers-need-300-billion-in-annual-ai-revenue-to-break-even-goldman-4917423
- https://www.cnbc.com/2016/08/19/google-alphabet-ipo-anniversary-growth-since.html
- https://fred.stlouisfed.org/series/dgs10