Chapter 4 · Does the AI Bubble Exist?

The Verdict: Yes — With Confidence Level and Dates

Weighing both cases into an explicit call, four probability-weighted scenarios, and the triggers that tell you which world you're in.

Bubble exists?Yes (~80%)valuation + financing excess on real earnings core
Major repricing by 2028~75%≥25% index drawdown, worse in AI infra
Most likely windowLate 2026 – mid 2027capex deceleration + IPO supply + yields
Systemic 2008-style bust~15%requires credit-layer contagion

Four scenarios, probability-weighted

Slow Deflation / Rotation (Base Case)

6–18 months40% probability

No single crash. Capex growth decelerates from 2027, second-tier names and neoclouds reprice hard, GPU depreciation reality dents big-tech EPS, and money rotates toward value, EM ex-China and defensives. Index drifts sideways-to-down in nominal terms while the bubble deflates out of the credit layer first.

Triggers to watch
  • Hyperscalers guide 2027 capex flat or down
  • Neocloud refinancing at visibly worse terms
  • GPU price cuts / write-downs hit earnings
Playbook pointers
  • Favor diversified large-caps over AI-concentrated funds
  • IT services becomes interesting again on valuation + relief
  • Avoid leveraged plays on the neocloud/SPV complex

Melt-Up Continues

3–12 months25% probability

The financing machine keeps working: securitization absorbs supply, Anthropic/OpenAI IPOs succeed, earnings keep beating and capex guidance keeps rising. The bubble extends — historically bubbles run 2+ years past the point where valuation-based bears capitulate. Higher highs before any resolution.

Triggers to watch
  • Blockbuster lab IPOs absorbing $100B+ without breaking liquidity
  • Capex guides raised again into 2027
  • Productivity evidence finally shows in non-tech earnings
Playbook pointers
  • Participation justified with position sizing discipline
  • Pre-commit exit rules for AI-heavy allocations now
  • Keep staggered rebalancing triggers rather than all-or-nothing calls

Contained Crash (Rate-Cut Rescue)

any time; most likely 202720% probability

A discrete shock — a failed mega-refinancing, an IPO pulled, a lab solvency scare — takes the Nasdaq down 25–35%. The Fed cuts hard, hyperscalers backstop their ecosystem, and markets stabilize within quarters. Dot-com-style wealth destruction in the most exposed layer, but no banking crisis.

Triggers to watch
  • OpenAI or Anthropic funding round/IPO fails or reprices sharply
  • Oracle/neocloud debt raise pulls or defaults cascade
  • 10Y yield breaks above 5% forcing de-risking
Playbook pointers
  • Hold dry powder / staggered buy lists ready
  • Do not catch falling neoclouds or SPV-linked paper
  • Use the dislocation to add quality IT, banks, consumption

Credit-Led Systemic Bust

2027–202915% probability

The bear case in full: circular financing unwinds — OpenAI's obligations crack, Oracle's RPO evaporates, SPV paper inside insurers/pensions marks down, private-credit losses surface — producing a 2008-flavored deleveraging, not a dot-com equity-only washout. Recession follows because AI capex has been carrying GDP growth.

Triggers to watch
  • Major lab insolvency or forced restructuring
  • Insurer/pension losses force asset sales across markets
  • Japan/yen-carry break compounds global tightening
Playbook pointers
  • Capital preservation mode: gilts/fixed income, gold allocation, cash ladder
  • Avoid all leverage; reduce equity beta systematically
  • Prepare shopping list of compounders for the aftermath

S&P 500 Shiller CAPE — Where We Sit in History

Cyclically-adjusted P/E at selected market peaks and today. Only 1929, 2000 and 2021 were more expensive. Live values move daily — use the link for the current reading.

Why this distribution rather than something more dramatic: the corpus's own strongest bears (ECB included) describe correction LIKELY, not certain; the financing layer adds genuine delay capacity; and the earnings core is real, which historically caps index-level damage short of 1929-style collapses unless credit contagion takes hold (the 15% tail). Conversely, pure melt-up-forever requires either productivity miracles showing up in non-tech earnings quickly, or indefinite credit tolerance for negative-FCF champions — possible, not probable. The middle scenarios dominate because that is where the evidence currently points: slowing second derivative, repricing edges, intact core.

A note on intellectual honesty: these probabilities are judgment, not measurement. They synthesize 157 sources whose median stance is bearish-but-timing-humble (63 of 88 video commentators lean bear; professional articles split closer to even, weighted by institutional weight toward 'correction likely'). Treat the numbers as a structured way to update as triggers fire — the watchlist in Chapter 7 tells you exactly what to monitor.

Sources cited on this page

    View the full research corpus (157 sources) ↗