Dead Money (Ed Zitron)
Connections
Ed Zitron continues his “Where’s Your Ed At” newsletter series on the AI bubble. This piece is the direct follow-up to Where’re All The AI Chips (Ed Zitron) (2026-09-22), which established that hundreds of billions in NVIDIA GPUs are warehoused and uninstalled. “Dead Money” escalates the argument to the full financial picture: hyperscalers, AI labs, and AI data center debt are all structurally insolvent at current capex levels. Key claims: Nvidia GPU sales are largely speculative purchases by hyperscalers (Microsoft, Google, Amazon, Meta (Company)) that sit in warehouses for years; OpenAI and Anthropic account for 64%+ of all hyperscaler AI revenue; the AI data center debt doom loop (more capex → higher debt costs → higher hardware costs → more debt needed) is accelerating; and Anthropic’s leaked S-1 shows $8B+ loss on $4.6B revenue in 2025. CoreWeave faces particularly dire refinancing risk. Connects to Kyle’s Capital Allocation and AI Infrastructure threads.
Summary
Fidelity’s Jurien Timmer called AI “dead money” for three-plus months (flat/declining GPU lease rates, H100/A100 rental rates flat to down). Zitron builds the full financial case:
Hyperscalers need $2–3 trillion in annual AI revenue to justify capex through 2030. Current estimates: ~$183B in combined AI revenue for Google, Microsoft, Amazon, Oracle, and SpaceX — of which at least 64% ($118B) comes from Anthropic and OpenAI. To break even on just 2026-2027 capex, hyperscalers need $308B annually; for a 10% ROIC they need $417B. They are $125–243B short.
AI companies would need ~$400B in annual revenues for hyperscalers to break even (assuming 10% operating margins). OpenAI estimates $36B in revenue for 2026; Anthropic ~$40B; even doubling both, total AI startup revenues barely reach $157B — $268B short of break-even.
The AI data center doom loop: more capex → more competition for RAM/HBM/copper/labor → higher hardware prices → more expensive debt → more capex needed. Morgan Stanley estimates $570B in AI-related debt issuance in 2026; Goldman estimates hyperscalers will fund 30%+ of AI investments with debt in 2027 ($400B). 10-year Treasuries at ~5.24% as of publication (up from 4.6% in July).
Oracle and CoreWeave as barometers: Oracle’s Nov 2025 bonds (~$18B) would be 41.5% more expensive if issued today, adding $6.89B in interest; S&P downgraded Oracle to BBB (one above junk). CoreWeave faces 11–13% yields on refinancing vs. 9% coupons it already pays. The $18B Project Jupiter SPV has entered “distressed” territory (89–91 cents on the dollar) after a force majeure notice on the New Mexico data center.
Anthropic’s S-1 (leaked to Reuters): In 2025, Anthropic lost $8B+ on $4.6B revenue. Compute costs: $7.33B. Spending $2.75 to make $1 — worse than OpenAI ($2.60 to make $1). Zitron alleges the “65B annualized revenue run rate” from July 2026 used July 31 × 365 rather than trailing-28-days × 13, calling it “sub-Enron.”
Long-term GPU rental rates are far below spot: B300 long-term contracts (5-year) pricing at $3.30–$4.30/hour vs. SiliconData’s spot index of $5.78/hour for B200. At those rates, a $1.1B data center barely pays off in five years with zero operating expenses and no debt.
Conclusion: ~$800B in global VC investment in AI since 2023; Zitron estimates at least $300B is dead money. Outside Anthropic and OpenAI, ~$22B in global non-lab AI compute demand. The doom loop breaks when data center debt becomes untenable to raise, most likely at 11%+ rates or when a major data center construction project collapses.