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

Episode Summary

Executive Summary: The episode centers on Ed Zitron’s critique of Anthropic, OpenAI, and the broader AI boom as financially fragile, opaque, and increasingly dependent on leverage, reseller channels, and speculative infrastructure. He argues that reported run-rate and profitability claims are misleading, while massive compute commitments and private-credit funding create systemic risk that could spill into hyperscalers, bonds, and the wider economy.

Main Topics: Anthropic’s financials and profitability claims (Priority: 5/5): Zitron argues Anthropic’s leaked 2025 financials show a worse business than OpenAI, with heavy losses, high burn, and questionable profitability metrics that exclude training or use adjusted accounting. Revenue concentration and reseller dependency (Priority: 5/5): A large share of Anthropic’s revenue reportedly comes indirectly through Amazon and Google reselling models, creating counterparty and concentration risk rather than true diversified demand. Run-rate and ARR as misleading metrics (Priority: 5/5): The discussion attacks annualized run rate as a marketing number that can exaggerate revenue by extrapolating from short periods or even single days, obscuring real business performance. Compute obligations, debt, and balance-sheet risk (Priority: 5/5): Zitron highlights massive non-cancelable compute commitments, including obligations tied to hyperscalers and Broadcom, arguing these liabilities are far out of proportion to current revenue. Private credit, bonds, and data center financing (Priority: 4/5): The conversation broadens to how AI infrastructure is being funded through risky debt markets, private credit, and SPVs, with rising rates threatening refinancing and repayment. Consumer AI products and limited real-world utility (Priority: 3/5): Muse, Dots, and similar agentic products are presented as marginally useful but not transformative, with weak monetization prospects and serious privacy/safety concerns. Bubble risk and macroeconomic spillovers (Priority: 5/5): The episode concludes that if AI growth slows or the bubble pops, the fallout could hit tech valuations, GDP growth, shadow banking, and potentially the broader economy.

Key Arguments: Anthropic’s 2025 financials suggest it was a weaker business than OpenAI, burning more cash relative to revenue and relying heavily on resellers. Reported profitability is suspect because it often excludes training costs or uses adjusted metrics that do not reflect true economic profit. Run-rate figures like $70B or $65B annualized revenue are misleading because they can be derived from a month or even a single day of activity. A significant portion of Anthropic’s revenue comes indirectly via Google and Amazon, meaning its “customers” include powerful counterparties that can also compete with it. Massive compute commitments, including non-cancelable obligations, could overwhelm even a profitable AI lab if growth slows or credit markets tighten. The AI buildout is being financed through increasingly fragile debt structures, private credit, and special-purpose vehicles, raising systemic risk. Consumer AI agents may have some utility, but they are not compelling enough to justify the hype, costs, or privacy risks. The AI narrative has shifted from useful marketing to a liability, especially as public skepticism, safety warnings, and weak economics collide.

Data Points: Anthropic 2025 revenue: $4.6 billion - Leaked financials discussed as part of the Reuters/IPO filing analysis. Anthropic 2025 loss: $8 billion - Reported loss cited as evidence of weak economics. Anthropic burn multiple: $2.75 spent to make $1 - Zitron’s comparison of cost efficiency. OpenAI burn multiple: $2.50 spent to make $1 - Used as a benchmark showing Anthropic slightly worse. Anthropic revenue via resellers: 47% - Portion of revenue reportedly coming through Google and Amazon reselling models. Customer concentration: 80% of revenue from 1% of customers - Claim that most revenue comes from a tiny set of mostly AI startups. Anthropic compute obligations: $518 billion - Total obligations discussed from leaked S1-style reporting. Non-cancelable obligations: $413 billion - Subset of Anthropic’s obligations described as non-cancelable. Non-cancelable hyperscaler obligations: $252 billion - Obligations across Microsoft, Google, and Amazon. TPU lease obligations: $161.2 billion - Broadcom TPU lease obligations discussed as part of Anthropic’s liability stack. Broadcom deal structure: $35 billion - Referenced as a financing arrangement with limited Anthropic insolvency exposure. Anthropic latest quarter revenue: $11.5 billion - Used to underscore mismatch with huge obligations. OpenAI run rate: $70 billion - Annualized revenue figure discussed skeptically. OpenAI prior run rate: $40 billion - Referenced as the prior month’s figure before a sharp jump. Anthropic run rate example: $65 billion - July annualized run rate cited as derived from a single day’s revenue. Anthropic single-day revenue extrapolation: $178 million x 365 - Example of how the run-rate figure was allegedly constructed. CoreWeave debt: ~$25–30+ billion - Described as a heavily leveraged AI infrastructure provider. CoreWeave bond raise: $7 billion - Straight bonds raised in 2025/2026 discussed as a reference point. CoreWeave implied reprice yield: ~13% - Estimated cost if current bonds were issued today. Oracle data-center bond pricing: 7.1%–8.3% - Estimated current borrowing cost after selloff for various maturities. Oracle debt trade level: 89–91 cents on the dollar - Used to show weakness in Oracle bond pricing. 10-year Treasury yield: ~5.287% - Cited as the benchmark worsening AI financing costs. AI capex projection: $750 billion through 2030 - OpenAI spending projection used to show scale of future financing needs. VC funding into AI since 2023: $800 billion - Used to argue that private capital may already be exhausted. Microsoft AI capacity: 2 gigawatts vs 12 gigawatts claimed - Example of the gap between claimed and actual data-center capacity. NVIDIA fiscal 2028 implied revenue need: $670+ billion - Projected revenue required to support AI buildout assumptions. AWS/Amazon oversubscription example: 1.6x - Mentioned as weaker-than-usual demand for a bond sale.

Pivotal Quotes: "Anthropic, at least in 2025, was a worse business than OpenAI." — Ed Zitron: His central assessment after reviewing the leaked financials. "The myth of Anthropic has been since the very, very beginning, that this was the more profitable, more sustainable, more stable company." — Ed Zitron: Used to challenge the popular narrative around Anthropic versus OpenAI. "Give me your goddamn financials, you worms." — Ed Zitron: Angry demand for transparency amid safety warnings and slow-rolled reporting.

Implications: If these economics hold, AI could face a financing squeeze, lower valuations, failed data-center projects, and pressure on hyperscalers and private credit. The sector’s narrative may not survive contact with audited numbers and higher rates.