行业

Gary Marcus 用 Claude 和 Gemini 交叉分析 AI 泡沫规模

精选理由

有人让 Claude 和 Gemini 一起算 AI 泡沫有多大,连循环融资和退休金敞口都算了,还拿 2008 年次贷做对比,数据挺全。

Gary Marcus 转发 Luke Hutchison 的做法:让 Claude 和 Gemini 协作分析 AI 泡沫,再由 Claude 交叉核验并汇总报告。报告估算 AI 相关股票约 30 万亿美元,占美股约 40%,高出长期趋势约 12-14 万亿美元。IMF 统计 AI 公司之间的循环融资含 3000 亿美元股权加 6500 亿美元债务。报告还对比 2008 年次贷危机:次贷债务 2.1 万亿美元,AI 含承诺的债务规模约为其 2.4-3.3 倍,并估算美国退休金约 10 万亿美元投向 AI 相关股票。

原文 · Gary Marcus

If you don’t believe me and @edzitron maybe you will believe Claude and Gemini. Whether you do or not, the outcome will be the same. Luke Hutchison @LH I had Claude and Gemini work together on analyzing the size of the AI bubble (including the circular-funding circus), the exposure of retirement savings to AI-linked stocks, and a comparison to the size of the 2008 subprime mortgage crisis -- then had Claude cross-check everything and write a combined report. Claude's executive summary: -- The AI bubble by the numbers (Nov 6, 2026) SIZE • AI-linked stocks: ~$30T, ~40% of the US stock market • Value above long-run trend: ~$12-14T (central estimate; range $2-22T) • Funded AI debt: ~$1-2T, plus ~$4-5T of purchase and lease commitments BUBBLE INDICATORS • S&P 500 CAPE: 41 (dot-com peak: 44) • US market cap/GDP: ~229%, near its 240% all-time record (dot-com peak: 144%) • Top 10 stocks = 40% of the S&P 500 (16% in 2016) • AI revenue covers at most 20% of the $1T+ spent on AI infrastructure in 2026 • Big Tech capex ($1T in 2026) exceeds operating cash flow ($804B); hyperscaler free cash flow projected negative by early 2027 • Circular financing: IMF counts $300B equity + $650B debt flowing between AI firms, tied to $1.4T of revenue commitments back to the investors • Suppliers guaranteeing their customers' debt: ~$260B of explicit backstops (Broadcom alone up to $43B) • Profits flattered: $160B of $285B in Q2 profit at 5 megacaps came from marking up private AI stakes; longer server depreciation lives add billions per year • Credit stress: Nvidia CDS hit a record in July; neocloud borrowing now costs ~10% • Nov 5: Nasdaq -5.5% into correction; chip stocks -12%, worst day since March 2020 RETIREMENT SAVINGS EXPOSURE • Total US retirement assets: $51.2T (IRAs $19.9T, 401(k)/DC plans $15.0T, pensions $13.6T, annuities $2.7T) • Invested in AI-linked stocks: ~$10T (range $6-16T), or 12-32% of all retirement assets • Retirement accounts and pensions hold ~1/3 of all US stocks • An S&P 500 index fund puts ~40 cents of every dollar into the top 10 stocks, mostly AI-linked • 80%+ of 401(k) plans default savers into target-date funds; the 10 largest active 401(k) funds hold an average 38% in tech and communication stocks • Stocks are 34% of US household wealth (dot-com peak: 26.7%) • Estimated retirement losses: Magnificent 7 fall 25%: ~$1.7T. AI stocks return to trend: ~$4-5T (range $0.7-7.3T). Broad market reversion: ~$12-15T VS. THE 2008 SUBPRIME CRASH (today's dollars) • Subprime debt outstanding: $2.1T. AI funded debt: ~0.5-1x that. Including commitments: ~2.4-3.3x • Housing bubble excess value: $13.4T. AI excess value: ~$12-14T, about the same • Household wealth lost 2007-09: $22-28T. AI stocks returning to trend: ~$12-14T lost. A broad market reversion: $35-45T KEY DIFFERENCE: 2008 was leveraged household mortgage debt held by banks. The AI boom is mostly equity, so losses land on shareholders and retirement savers; default risk is concentrated in neoclouds, AI labs and their off-balance-sheet financing vehicles. -- Full report with sources: claude.ai/artifact/AFTSW… 🔗 View Quoted Tweet 💬 3 🔄 2 ❤️ 13 👀 1423 📊 3 ⚡