Ed Zitron on Anthropic's IPO (S-1), AI Debt, and Counterparty Risk

Ed Zitron portrayed Anthropic’s finances as far from proven profitability: he cited $4.6 billion in 2025 revenue and an $8 billion loss, or $2.75 spent per revenue dollar, compared with $2.60 at OpenAI. He said adjusted-profitability claims are misleading if they exclude training, since the companies must keep investing to maintain model quality. About 47% of Anthropic’s revenue, by his account, came from Google and Amazon reselling its models, even though both compete with Anthropic. He also said 80% of Anthropic’s and OpenAI’s revenue came from 1% of customers, predominantly AI startups, challenging the idea that Anthropic has little customer-concentration risk.

Zitron cited $518 billion in Anthropic commitments, including $430 billion that cannot be canceled, and separately described $252 billion in non-cancelable computing agreements with Microsoft, Google and Amazon, plus $161.2 billion in Broadcom TPU lease obligations. He contrasted those commitments with $11.5 billion in latest-quarter revenue and warned that suppliers’ earnings expectations depend on agreements with a loss-making company. A reported $35 billion Broadcom financing deal would leave Anthropic responsible for $5 billion if it became insolvent, with Broadcom otherwise required to repay lenders. Zitron viewed Broadcom’s debt and chip-financing exposure as riskier than Nvidia’s equity investments, particularly as Broadcom’s AI-chip sales concentrate among Google, Anthropic and OpenAI.

The financial evidence surrounding an Anthropic IPO remained incomplete in Zitron’s view. He said the leaked S-1 material available to the public lacked enough detail and questioned whether everything had been released; the IPO timetable had reportedly shifted from August to September, October and November, with two delays whose reasons were unknown. He also questioned reports that Anthropic became profitable in Q2, saying a shift from an $8 billion annual loss required explanation and might involve discounted compute. He urged caution with annualized run rates: Anthropic’s $11.5 billion in quarterly revenue was paired with a reported $48 billion run rate, while Bloomberg’s reported $65 billion July figure, he said, multiplied one day’s $178 million by 365. Zitron argued for audited disclosures and SEC rules, and said 2026 results would need to improve dramatically to support the bullish case.

Zitron argued that AI infrastructure depends on recurring financing at a scale that revenue growth alone cannot guarantee. He said OpenAI and Anthropic account for 80% of data-center demand despite being loss-making startups, cited a projection that OpenAI would spend $750 billion on compute through 2030, and compared the funding needed to sustain the two companies with the $800 billion venture capital had put into AI since 2023. Anthropic might need to raise $50 billion annually in high-yield debt, he argued; he paraphrased Dario Amodei as warning that misjudging compute purchases could bankrupt the company, and cited Daniela Amodei’s explanation that public markets would give AI companies access to debt. Even with investment-grade ratings, he said, borrowing could be priced like junk and require repeated, exceptionally large financing. He saw AI investment as a financial-crisis risk concentrated in interconnected banks, private credit and insurers, while the other speaker cautioned that productivity gains may take time to show up in statistics and that investment itself counts toward GDP.

CoreWeave illustrated the gap between signed commitments and dependable cash flow. Its counterparties include Microsoft, Meta, OpenAI, Anthropic and Jane Street, but Zitron said commitments do not ensure timely payment; OpenAI’s contract includes net-360 terms, and delayed data-center openings could add interest costs. He said CoreWeave raised about $7 billion in bonds in 2025–26, estimated new borrowing could cost 13.5–14%, and cited analysts’ estimate that it might need another $100 billion over four years. In his view, 30–40% gross margins would not cover 10–13% debt costs alongside depreciation, operations and staffing. He also cited Oracle’s roughly $25 billion bond issuance, yields of about 7.1–7.5% on shorter maturities and 8.1–8.3% beyond 20 years, and the risk that a downgrade to junk could force some funds to sell. Data-center SPVs, he warned, can separate a project from its sponsors without ensuring creditors are repaid.

Zitron questioned whether demand and construction schedules would support the capacity being financed. GPU-inference providers pay for capacity whether customers use it or not, while too little capacity can undermine service; he said there was no proof demand would support even a quarter of planned data centers. Delays would compound financing costs, and he warned private-credit lenders not to treat hyperscaler commitments as assurance of repayment. He expected some private-credit debt to go unpaid and some projects to fail, with collateral potentially inadequate; he cited a Meta-backed Hyperion SPV trading at 89 cents on the dollar and an Information report that Blue Owl took 15 minutes to decide on Stargate Abilene. He predicted that within a year OpenAI and Anthropic would need an unprecedented cash-flow improvement or face outcomes including death, merger, acquisition or breakup, and warned that a downturn in 2027 or 2028 could damage technology stocks, data-center lenders and the wider economy.