Chapter 3 · Does the AI Bubble Exist?
The strongest arguments that this is a true bubble — from the ECB's own economists to Dalio, Burry, and the credit-market tape.
The valuation case barely needs arguing: CAPE near 40x has been exceeded only in 1929, 2000 and 2021; the Buffett indicator sits at record territory; seven stocks are 32% of the index. But sophisticated bears concede those are necessary-not-sufficient conditions — bubbles need a funding mechanism and a trigger. So the serious bear case is about the credit architecture, and it comes from three directions.
First, the accounting critique. Michael Burry's central claim: hyperscalers understate chip depreciation by using 5-6 year useful lives against 2-3 year real economic lives, overstating industry earnings by ~$176B through 2028. If chips must be written down sooner — and H100 rental rates already collapsed ~75% from early-2026 peaks — then the 'real earnings' propping up valuations are themselves inflated. Unlike idle dot-com fiber, silicon rots: efficiency doubles every two years, so GPUs bought at ground-breaking run at quarter-speed by energization.
Second, the circularity critique. Follow the biggest revenue line items: ~$300B of Oracle's $638B backlog is OpenAI — a company that lost $38.5B in its last fiscal year; Microsoft says ~45% of its backlog (~$280B) is OpenAI; UBS projects labs approaching half of Google Cloud's 2027 revenue. Strip the labs' rented compute out of 'AI revenue' and the demand picture thins dramatically. Meanwhile the suppliers fund the buyers — Nvidia's commitments into its own customers crossed $500B. When the largest customer of your largest segment depends on your own vendor financing, revenue quality becomes a matter of definition.
Reported debt vs off-balance-sheet obligations across the five biggest AI builders (Nikkei investigation, mid-2026). Off-book devices include SPVs, leased data centers and purchase commitments.
CAPE ~40.4x as of 21 Aug 2026 — higher than ~99% of months since 1881 (long-run median 16.6) and just below the dot-com peak of 44.2. The ECB notes US valuations 'close to their historical peak' and argues a correction is likely even if AI enthusiasm is partly rational.
Mag7 ≈32% of S&P 500 market cap as of Aug 2026 (~13% in 2018) — highest top-7 concentration since the Nifty Fifty; top-10 ≈35-38%. A -25% move in the top-10 alone mechanically knocks ~9% off the index.
Buffett indicator (market cap/GDP) at record ~210-238% vs ~140% at the dot-com peak; S&P P/E ~32 vs long-run mean ~17.9. Goldman's Oppenheimer: the risk is not a valuation bubble but an 'earnings bubble' — expectations embedded in estimates.
2026 hyperscaler capex guidance is $720-745B (+~50% YoY; ~$835B including Oracle FY27, per TMT Finance, Aug 2026); combined OpenAI ($40B) + Anthropic ($65B) + Gemini + rest-of-industry model revenue is roughly $150B annualized. JPMorgan math (per corpus): AI must eventually carry ~$650B/yr of revenue to justify capex; Morgan Stanley sees a ~$1.5T funding hole through 2028.
Barclays' Ross Sandler: ~73% of Amazon's 2026-27 AI revenue is OpenAI/Anthropic compute spend; UBS: labs ≈48% of Google Cloud revenue by 2027; Wells Fargo: ~74% of Microsoft's AI revenues by FY27 end. Microsoft says ~45% of backlog (~$280B) is OpenAI; ~$300B of Oracle's ~$640B RPO is OpenAI.
Michael Burry: hyperscalers extend useful life to 5-6 years against 1-3 year real GPU economic life; understatement ~$176B through 2028, with $20B/year at Meta+Oracle alone by 2026-28. Shortening lives to 2-3 years would cut big-tech earnings meaningfully.
Nikkei investigation: $1.65T hidden vs $1.35T visible across Alphabet/Microsoft/Amazon/Meta/Oracle; Meta's hidden pile ~3x its reported debt; ~$900B of fresh commitments signed in a single quarter (mid-2026); Morgan Stanley pegs industry-wide off-BS exposure ~$1.8T. Blueprint: Meta's $27.3B Beignet SPV (Blue Owl) with residual-value guarantees.
Nvidia committed up to $100B into OpenAI tied to chip purchases, backs a $250B+ OpenAI compute facility and a $350B GPU-financing package; a new $500B financing vehicle was arranged with Apollo/BlackRock/Blackstone/Brookfield/GS/KKR. Nvidia's own CDS cost doubled since late May 2026 while its stock made highs.
Oracle: $638B RPO with only ~12% converting within 12 months; FY26 capex $55.7B (+162%) flipped FCF from +$26B to -$24B; rating one notch above junk; 5-yr CDS blew out to ~215bps (multi-year high); CoreWeave carries $25B debt at 7-15% coupons with two straight operating losses and depends on Nvidia backstops.
MIT study: ~95% of enterprise AI pilots show no measurable P&L impact; Gartner: 85% of projects underdeliver; S&P Global: 42% of firms abandoned most AI initiatives in 2025 (vs 17% in 2024); Apollo's Slok finds zero AI margin lift outside tech sectors.
Chinese models passed US models on OpenRouter token share (Feb 2026) and now exceed 60% of tokens; DeepSeek V4 Flash at ~$0.03/task vs ~$3.15 for Claude-class (~100x gap); OpenAI cut prices up to 80%; open-source inference runs at ~1/10 frontier cost. If inference is a commodity, $700B of capex earns utility-like returns.
Oracle 5yr CDS ~215bps (from ~145bps end-2025); SpaceX CDS +>50% in first month; record 5-yr CDS across Oracle/Meta/Alphabet/Amazon/Broadcom/Nvidia; Amazon's bond book collapsed from $62B to $41B orders; Meta's Louisiana SPV priced at 7.53% (near-junk spread); hyperscaler bond issuance $225B in H1-2026, up ~974% YoY.
US life insurers hold ~$4T bonds with ~$800B (1 in 5 dollars) in illiquid private credit/ABS/data-center financing; life insurers own ~42% of the private-credit market ($849B); BIS flags $2.1T moved to offshore reinsurers (14%→40% share); pension funds bought Meta/Blue Owl SPV bonds; SEC exempted these deals from post-2008 securitization rules.
Alphabet posted its first-ever negative-FCF quarter (-$5.8B: $44.9B capex vs $39.1B OCF); big-4 combined FCF fell from ~$210B/yr toward zero/negative in 2026; five hyperscalers issued ~$200B debt in H1 2026; BofA estimates AI capex consumes ~95% of post-dividend operating cash flow; hyperscalers flipped to net share sellers (-$147B in 2026).
Dalio: 'Yeah. Yeah. Yeah.' when asked if this is a classic bubble; points to the 1929/2000 mechanics: rising rates + record issuance meeting exhausted buyers. SpaceX IPO'd June 2026 as the largest ever and halved from peak; Anthropic/OpenAI mega-IPOs queued for late 2026/2027; Dalio's mechanic: forced selling takes $100 positions to $25 with loans still owed.
Jason Furman: information-processing equipment & software drove ~92% of H1-2025 GDP growth despite being ~4% of GDP; ex-AI the economy grew ~0.1%. Reventure: AI capex ≈49.5% of Q1-2026 GDP growth. An AI capex pause is functionally a recession trigger — policy will fight to prevent it.
Internet of Bugs: chip efficiency doubles every ~2 years, so GPUs ordered at ground-breaking run at ≤¼ efficiency when the data center energizes; ~75% of all AI chips sit unpowered in warehouses (Burry thesis via Rule #1 Investing); only ~5% of Nvidia's newest generation is deployed. Fiber could wait for users; silicon cannot.
US 10Y at 4.74%, 30Y above 5.33% (19-year high) as of Aug 2026; Japan dumped a record $66.7B of Treasuries in May during yen defense; US bought yen July 31 2026 (first since 1998); futures briefly priced Fed HIKES. CNBC frame: rising yields are the classic bubble-popper if they keep climbing.
Top 10% of households own ~90% of stocks, so capex cuts land directly on the consumption of the wealthiest; euro-area households hold ~€440B indirect Mag7 exposure (ECB); Korea showed retail margin-cascade mechanics live (300k+ accounts liquidated in hours); WaPo/political pieces warn of backlash as power bills and protests rise.
Motilal Oswal: AI could remove 9-12% of Indian IT revenue over four years; Nifty IT fell 19.5% in Feb 2026 (worst month since Sept 2008); fresher hiring collapsed ~600k/yr → ~120k/yr (-80%) in three years; June 2026 entry-level openings -44%, senior -67%, active tech demand at a 28-month low.
BofA August 2026 FMS: India Asia's least-preferred market at 32% net underweight (partly for LACKING AI exposure); CY2026 FII outflows ₹2.37 lakh crore through mid-August (> all of 2025's ₹1.66 lakh crore); rupee weakened to ₹93.7 avg (peak ₹96.9 on 23 July). Crowded-out positioning cuts both ways: no crowding = less cascade risk, but sentiment flips are sharp.
Scroll.in: correction triggers global risk-off → FPI withdrawals → rupee pressure; Indian retail holds US tech via Nasdaq-100/AI-themed funds; Indian AI-startup funding (+4x YoY in H1-2026) depends on the same global VC cycle that would freeze; semiconductor-manufacturing ambitions need foreign capital that would evaporate.
Third — and this is where 2026 added hard evidence — the credit tape started moving. Oracle's 5-year CDS spread blew out from ~145bps to ~215bps while its stock made highs; SpaceX CDS jumped >50% within weeks of listing; Amazon's bond book collapsed from $62B to $41B; Meta's Louisiana SPV priced at 7.53%, near-junk spreads for a company supposedly flush. Hyperscalers flipped from ~$190B of net buybacks in 2024 to net share SELLING in 2026 while issuing $225B of bonds in six months (+974% YoY). Alphabet posted its first negative free-cash-flow quarter since its 2004 IPO. None of this is normal late-cycle behavior for self-funding champions; it is behavior of companies that have outrun their internal cash generation and now depend on market access.
The macro overlay completes it: Jason Furman calculates that information-processing equipment and software drove ~92% of H1-2025 US GDP growth — meaning an AI capex pause IS a recession, which explains why Washington would move mountains to prevent one, and why the eventual policy response may be inflationary rather than deflationary. Add Dalio's cycle lens ('biggest bubble in American history,' with new-issue supply — SpaceX, Anthropic, OpenAI IPOs — as the classic topping mechanism), Korea's live demonstration of leveraged-retail unwinding, and the ECB's conclusion that correction is likely even under RATIONAL valuation assumptions, and the bear case stops being fringe commentary. It is now the official-sector baseline scenario.