The AI Bubble is Bleeding Cash, Part 2: Anthropic's Receipts | Ed Zitron

Zitron argues that AI companies’ reported profitability can obscure the recurring investment needed to train and maintain models: he cites Anthropic spending $2.75 to generate a dollar of revenue, questions profitability after an $8 billion annual loss, and compares model investment to oil companies’ drilling costs. He also says rising revenue may not resolve Anthropic’s cash-flow needs because costs rise with revenue, while OpenAI is projected to spend $750 billion on compute through 2030.

Zitron describes concentrated counterparties on both sides of Anthropic’s business: about 47% of its revenue comes through Google and Amazon reselling its models, while 80% of Anthropic’s and OpenAI’s revenue comes from 1% of customers. Against that concentration, he cites Anthropic’s $518 billion in compute commitments, including $413 billion that are non-cancellable, as significant counterparty and financing risk relative to its $11.5 billion latest-quarter revenue.

Zitron questions headline AI revenue run rates because they can annualize a short period or non-recurring activity: he says OpenAI’s reported $70 billion run rate rose from $40 billion in a month, estimates its full-year revenue at $36 billion, and describes Anthropic’s $65 billion July figure as based on one day’s revenue multiplied by 365. He says unconfirmed online estimates show Anthropic’s run rate flattening while OpenAI grows and takes market share back.

Zitron argues that data-center economics may be strained by both financing costs and uncertain demand: he says even 30%–40% gross margins may not cover 10%–13% debt costs, and that OpenAI and Anthropic account for 80% of demand. He also cites CoreWeave’s estimated 13% financing cost on repriced bonds and analysts’ estimate that it may need to raise another $100 billion over four years.

Zitron points to a disconnect between AI-linked market enthusiasm and practical constraints: Oracle performed well after OpenAI’s reported $70 billion annualized run rate, despite its Port Washington data center lacking needed power. He also says Oracle is close to losing investment-grade status, which could force some funds to sell its bonds and make refinancing debt raised at 2.8% cost 5%–8%.

Zitron says debt markets may be less receptive to AI infrastructure financing than stock-market excitement suggests: he identifies debt as one source of planned AI capital expenditure and notes that an Amazon bond sale was subscribed 1.6 times, below the three to four times he says is generally desired. He predicts some bond sales could fail or require uneconomic rates, and warns that private-credit debt tied to volatile, low-margin data-center projects could go unpaid, identifying Blue Owl as heavily involved.