Bubble mechanics, operationalised for the AI boom

Last updated: 2026-10-21 · Author: Claude (Fable 5.1 subagent) · Status: reviewed · Reviewer: Codex (gpt-6-astra) cross-check, applied 2026-10-17 · Updated 2026-10-21 from signals/2026-10-21.md

Summary

• Five complementary frameworks organise the questions: Minsky on debt-service capacity, Kindleberger on credit and market stages, Perez on technology installation and deployment, Dalio on six bubble indicators, and Shiller on valuation and narratives. Their overlap is not independent confirmation of a bubble; Perez's fetched AI commentary does not predict a current crash.
• The dated evidence shows expensive US equities and a large infrastructure build. September CAPE was 40.58 versus a 17.8 long-run mean; September margin debt was $1.480 tn, up 31.4% year on year but below June's record. Oracle and CoreWeave reported backlogs of $664 bn and $103.7 bn at different quarter ends. Commitments alone do not establish inadequate returns (claims 1, 4, 15–16, 33, 36).
• Internal investment funding headroom is shrinking: Microsoft, Amazon, Alphabet and Meta spent 86.1% of operating cash flow on cash capex in Q1 2026 and 96.1% in Q2. The original 94% Q1 figure included Oracle as a fifth company. Individual firms differ, and capex/OCF alone cannot assign Minsky debt-service postures (claims 30–31, 41).
• Forward P/E of 19.0x is slightly below its ten-year mean and provides a counterweight to high CAPE. Shiller's published excess CAPE yield is positive, while the note's differently constructed TIPS proxy is negative. Neither contrast proves that earnings growth must persist indefinitely (claims 7–10, 44).
• Higher rates and issuance are potential pressure channels: the September hike, October Treasury high and record Q2 IPO proceeds are documented. Issuance was concentrated and Q3 weakened; financing can also improve liquidity. These observations do not establish an unavoidable bust or its date (claims 18–22, 27–29, 51).
• Financing risk varies by issuer. CoreWeave's cash flows show refinancing dependence, not demonstrated Ponzi finance; September disclosures show completed financing and demand. October 19–20 quotes put its unsecured yields near 15–16% and Oracle CDS at 315/322 bp; these are issuer stress readings, not defaults or synchronized closes. S&P warns of opaque liabilities without a new rating action; Elice’s weak debut and divergent executive views qualify a uniform risk-appetite story. The original rubric sums to 20/26, but its composite assessment remains withdrawn pending consistent rules and calibration (claims 23–26, 42–43, 46–50, 59–64).
• No current numerical Dalio aggregate reading was located in this research. The central unresolved tests are sustainable customer cash generation, contract quality, debt maturities and recoverable asset values. The May Fed report also records systemic resilience alongside leverage vulnerabilities (claims 39–40, 52).

Key claims

Type: primary / secondary / tertiary. Grade: A–D (see report/methodology.md). A grades for company guidance, consensus estimates and institutional assessments verify what was published, not whether the forecast or interpretation is true. Arithmetic is B; unverified underlying forecasts retain their lower confidence. The existing CSVs contain recent Shiller and FINRA observations, not the full histories used for sample statistics. Full-history calculations cite the fetched workbooks. New quarterly figures use cash capex, excluding newly signed leases; year-to-date cash flows are differenced to obtain individual quarters.

Analysis

1. Minsky: stability breeds fragility

Minsky's original paper distinguishes hedge finance, where operating cash can meet contractual interest and principal, from speculative finance, which can pay interest but needs refinancing for principal, and Ponzi finance, which requires borrowing or asset sales even to meet interest. Prolonged stability can encourage a shift toward fragile financing; monetary constraint can expose that fragility (source 25). These are debt-service tests, not tests of net profit or discretionary growth capex.

• Hyperscalers: the named four generated $150.727 bn OCF against $129.750 bn cash capex in Q1, leaving $20.977 bn, rather than no free cash flow (claim 30). Q2 cash capex/OCF rose to 96.1% (claim 41). Amazon and Alphabet spent more than operating inflows in Q2, while Microsoft and Meta retained surpluses. This compresses investment funding headroom, but the aggregate cannot classify each firm's contractual debt-service posture. Epoch's historical series includes Oracle; its trough and subsequent quarterly readings are shown separately in claim 31. Its extrapolated crossover is a scenario, not an observed transition to speculative finance.
• Oracle: negative investment FCF and completed equity issuance (claim 33) show external funding of expansion. They do not establish inability to repay maturing debt from operating cash, nor a need to borrow for interest.
• CoreWeave: the missing cash test is available. H1 OCF was +$3.663 bn after $806 m of operating cash interest. Cash generation before that interest was $4.469 bn, below $6.201 bn of principal and total cash interest paid (claims 42–43). Deferred revenue contributed $1.365 bn to OCF. This supports speculative financing exposure and a test of customer prepayment sustainability; it does not establish present Ponzi finance. Net interest expense/revenue in claim 35 is not cash-interest coverage.

The September rate increase is a potential Minsky pressure channel, not proof that a transition has occurred. The newer Fed record also includes views that policy remained only mildly restrictive or not restrictive (claim 46); contractual maturities and borrower-specific cash flows must show the transmission.

CoreWeave's September financing and demand disclosures qualify the distress reading. The September 18 base pricing of $3.7 bn plus a $500 m option was followed by a September 22 closing of the full $4.2 bn at 2.875%, due 2033. Net proceeds after capped calls were $3.5708 bn before offering expenses, versus the earlier base-offering estimate of $3.1457 bn (claims 47–48). This older closing, already recorded in the October 17 signal, corrects the note's settlement uncertainty; it is not new October funding. It shows market access and potential dilution, not cheap comparable straight debt. Its September 17 ATM authorises up to 35 m shares without establishing any completed sales (claim 49). The same day's pricing update reports higher-priced short contracts and >$25 bn of early-Q3 net commitments outside June backlog (claim 50). Management's disclosures counter a simple demand-collapse story without proving customer solvency or turning commitments into recognised revenue.

2. Kindleberger: stages, and credit as fuel

The displacement, boom, euphoria, profit-taking and panic sequence (source 26) provides hypotheses to test, not a unique stage inferred from proxies.

Renaissance's record Q2 proceeds were concentrated: SpaceX supplied 71.6%, leaving $29.8 bn, still strong by its assessment (claim 51). Q3 proceeds excluding SK hynix were only $6.2 bn, with postponements (claim 27). Rounded headline and mega-deal totals need not subtract to that separately reported exclusion figure. IPO dollars do not count new investors; existing investors can buy new issuance. Equity portfolio weights can rise with prices alone.

Credit remains the mechanism to investigate. The debt-issuance forecast and Futuriom growth inputs (claims 32, 37) need primary confirmation. Conditional on claim 32 (arithmetic in claim 56), capex growth was about 3.6 times revenue growth in 2025 and the 2026 plan implied 5.2 times. Ratios of growth rates do not measure funding dollars, and a plan is not a second realised year. Those conditional inputs cannot establish the direction or scale of the financing gap by themselves.

3. Dalio's checklist, item by item

Dalio asks six questions about unsustainable prices (sources 20, 35). His published past readings and later verbal comments are in claims 39–40. No current numerical aggregate reading was located in this research. Rate increases and issuance are potential pins, separate from the six measures. This note's proposed proxies and thresholds are not Bridgewater's gauge.

1. Prices high versus traditional measures. September CAPE is near the 98.9th percentile of the January 1881–October 2026 sample, including partial October, below the 1999 peak and August 2026 (claims 1–4, 45). The Fed's valuation warning is consistent with expensive equities (claim 11). The draft CAPE threshold is retained in §7 for audit.
2. Prices discounting unsustainable conditions. Dalio's own measure estimates growth needed for equity returns to exceed bond returns (claim 57; source 35). A CAPE/forward-P/E percentile gap is not that calculation. No matched forward-P/E distribution was supplied. The 19.0x multiple is slightly below its ten-year mean (claim 8); a mean does not establish a median or percentile. Illustratively, reducing growth from 32.8% to 16% gives 19 × 1.328 / 1.16 = 21.8x at unchanged prices and a common earnings base (claim 53). CY2026 growth and NTM EPS cover different periods, so this is not exact NTM repricing or proof of unsustainable expectations.
3. New buyers entering. Issuance and household equity weights are indirect proxies (claims 27–28, 38, 51). Concentrated mega-deals and an unverified allocation statistic cannot establish broad new participation. Private-lab listing dates remain reported, unverified forecasts (claim 29).
4. Broad bullish sentiment. Bears exceed bulls by 2.9 percentage points; FMS cash is 3.9% (claims 25–26). These readings do not satisfy the draft's full-score thresholds. Survey sentiment does not uniquely locate late euphoria or profit-taking; alternative interpretations belong in Conflicts.
5. Leverage-financed purchases. September margin debt was up 31.4% y/y or 25.0% since October, but remained below June's record (claims 15–16). The 4.5% debt/GDP figure mixes a debt stock with annualised output and reference dates; it is not AI-specific (claim 17). The FSR's hedge-fund warning spans several strategies, not just AI (claims 12, 52).
6. Extended forward purchases. Oracle's backlog is about 34 quarters of current revenue, or 8.6 times annualised revenue; CoreWeave's is 10.068 times using unrounded quarterly revenue (claim 55). These are static ratios, not actual delivery duration. CoreWeave expects recognition over the schedule in claim 36. High capex/OCF (claims 30, 41) shows exposure to future demand, but neither metric alone proves purchases motivated by scarcity or inability to earn a return. CoreWeave's top-five-seller status was not established for the draft threshold.
7. Tightening policy, a potential pin. The hike and Treasury high are documented (claims 18–22); one hike does not establish an unavoidable bust. Yield increases have several drivers in the latest Fed record (claim 46). Issuance can finance productive expansion and improve liquidity.

The tertiary concentration proxy distinguishes ten security lines (38.39%) from ten companies (39.89%), with Alphabet combined. Those are not certified index weights (claim 13, grade D); Apollo's broader concentration assessment is separate evidence (claim 14).

4. Perez: installation, frenzy, and what survives

Perez's installation/deployment framework separates technological progress from investor returns. The accessible introduction to her March 2024 article places AI within the ongoing ICT revolution (claim 58; source 45). The earlier note relied on Wladawsky-Berger's quotations for her broader mid-diffusion and finance/production argument (source 27); the cross-check did not verify the full paywalled text. No fetched Perez source calls the current build a frenzy or predicts its crash. Kedrosky's interpretation emphasises late installation and surviving infrastructure (source 28); these are competing interpretations, not Perez's current-cycle forecast.

The proposed operational test compares investment growth, revenue and financing. Conditional growth-rate ratios of 3.6x in 2025 and 5.2x in the 2026 plan (claims 32, 56), and the observed increase in cash capex/OCF (claims 30, 41), motivate scrutiny. They do not establish the draft rule's two realised years plus debt funding of the marginal dollar. That missing financing test prevents a full-score frenzy finding. Useful infrastructure may survive losses, but utilisation, operating costs and compute obsolescence determine its recoverable value; physical survival does not guarantee recovery.

5. Shiller: CAPE and narrative

September CAPE was 40.58, with partial October at 40.71, versus a 17.8 sample mean (claims 1–4). The earlier historical-exclusivity claim was wrong: August 2026 was higher (claim 45). September/October CPI is estimated, and October's price and nominal yield use October 2 observations, not an October 17 market snapshot (claims 2, 44).

The TIPS-based proxy is −0.32 percentage points using September CAPE and October 15 real yields (claim 7). Shiller's published excess CAPE yield is +0.82% for September and +0.56% for partial October under a different construction (claim 44). The opposite signs materially change the draft negative-yield scoring interpretation; the measures cannot share an unlabeled threshold. The review did not establish an all-history real-yield novelty claim under a consistent real-rate definition, so none is asserted here.

Shiller's narrative lens asks how stories spread. Griffin versus Fink/Sewing provides dated evidence of disagreement, not a broad sentiment measure (claim 63). Yardeni’s October 21 meter separately reports high bubble conditions and low triggers (claim 64). Its historical two-year bear-market frequency is a model backtest with a different event definition; it does not calibrate this note’s withdrawn rubric or supply a crash clock. No quantitative AI narrative index was fetched; earnings-call mentions remain a possible research measure.

6. Technology bubble vs asset-price bubble, and the role of debt

An asset-price reversal can destroy marked-to-market wealth and transmit losses through spending, collateral and financing. It need not be a zero-sum wealth transfer. Technology overbuild can leave useful infrastructure, but recoverable value depends on utilisation, operating costs and obsolescence. Trace losses through equity, contractual claims, collateral and guarantees.

The AI case combines expensive equity claims with a large physical build. The four-company funding ratio rose to 96.1% in Q2; Oracle and CoreWeave's combined $767.7 bn backlog joins August and June snapshots, not same-date near-term cash requirements (claims 41, 55). Equity ordinarily absorbs losses before creditors at Oracle and CoreWeave as well as at hyperscalers. CoreWeave's carrying debt is distinct from scheduled principal (claim 34); creditor recoveries require a claim-by-claim map of security, recourse and guarantees. The SPV/private-credit layer remains unsized here.

The May FSR highlights weaker debt service at riskier floating-rate borrowers and some BDC redemption pressures, alongside high hedge-fund leverage. It also describes moderate debt/funding vulnerabilities, declining aggregate debt/GDP and strong bank capital (claims 11–12, 52; source 4). This qualifies an economy-wide fragility inference and AI attribution. Broad HY and IG OAS cannot establish that AI risk is unpriced; the newer sector evidence in claims 46 and 59–60 changes that conclusion. S&P’s new opacity warning reinforces the need for that map, while its decision not to change ratings qualifies an immediate-downgrade reading (claim 61). The key test is whether customer cash arrives before debt and investment commitments fall due, and what assets are recoverable if it does not.

7. Draft scoring rubric

The original draft used 0 = absent, 1 = present, 2 = historical extreme. Those labels were not historically calibrated. The original assignments below are retained for audit, with corrected readings and explicit defects; all composite scoring and stage inference are suspended. This is not a revised numerical assessment.

The displayed assignments sum to 20/26, not 19/26. Overlapping valuation proxies, inconsistent thresholds and unverified inputs preclude a defensible mechanically corrected assessment. Before any revised score, specify current versus recent records, direct versus proxy participation, a matched valuation test, and realised growth plus marginal funding; then calibrate against history. Neither this rubric nor Dalio's own timing caveat (source 35) supports a crisis clock. At unchanged prices, lower expected earnings raise forward P/E; the observed multiple can rise or fall depending on the price response. No claim that spreads or forward multiples necessarily move last is established here.

Conflicts and open questions

• Valuation and the cross-check's judgement. The original note read high CAPE and normal forward P/E as proof that permanently high growth was priced in, and treated the discrepancy as bubble confirmation. The review is less confident about unsustainable pricing: the forward multiple and Fed's earnings-led-rally assessment provide a genuine competing reading (claims 8–10, 46). The bear hypothesis is that infrastructure spending supports suppliers' earnings without sustainable end-user cash generation; the bull hypothesis is that earnings growth earns the investment return. Matched EPS periods and customer cash economics are needed to distinguish them. The unfetched Goldman commentary in the draft is not evidence here.
• Debt-service classifications. The original author called hyperscalers hedge units, Oracle speculative and CoreWeave speculative-to-Ponzi. The review rejects those categorical labels from capex deficits/net losses: CoreWeave's disclosed cash test supports refinancing dependence but not present Ponzi finance (claims 42–43). The original concern about continued market access remains a risk hypothesis. Resolve it with maturity schedules, cash available for interest/principal, and prepayment/contract durability.
• Credit pricing. The original note inferred unpriced AI risk from low broad OAS. The review is more concerned about differentiated credit risk, citing already elevated hyperscaler spreads in the minutes. Broad indices and sector spreads measure different exposures; retain both (claims 23–24, 46). Neither establishes the fair spread for each borrower.
• Stage and triggers. The original author favoured late euphoria/early profit-taking with the policy pin already in. The review finds the stage and timing inference unsupported: bearish sentiment, financing access and a single hike permit several readings. Falling customer cash conversion, failed refinancing and broad forced selling would strengthen distress; durable customer cash growth, funded maturities and backlog conversion would weaken that diagnosis. Corporate issuance alone is not insider selling.
• Shiller yield definitions. The original negative TIPS proxy motivated full bubble points; the review adds Shiller's positive published ECY (claims 7, 44). Both are retained with their definitions and dates; a score depends on which construction is justified, not a silent sign substitution.
• Dalio attribution and limits. The original note treated the absence of a current numeric gauge as the biggest uncertainty and asserted it was unpublished. The review only failed to locate one. Past readings and verbal comments remain in claims 39–40; the rubric is the author's, not Bridgewater's. Proprietary percentiles would not resolve customer cash, maturities or recoverable asset values, nor automatically validate the rubric.
• Lamont's four conditions. Fortune (source 23) reports overvaluation, bubble beliefs, new entrants and issuance, with issuance still missing in February 2026. The original author said Q2 completed the framework; the review's participation/concentration objection prevents that deduction. The underlying Acadian post was not reached: the attribution is grade C, and the conditions need direct, dated evidence.
• Perez. The original author used capex growth and debt forecasts to diagnose frenzy while acknowledging that Perez did not make that call. The review emphasises her placement of AI within ICT and the unverified marginal-debt/two-realised-year tests. Wladawsky-Berger and Kedrosky offer differing current-cycle interpretations (sources 27–28, 45). Infrastructure survival and adequate economic recovery are separate propositions.
• Sentiment versus leverage. The draft explained bearish AAII as a lagging retail survey or leverage concentrated elsewhere. The review treats those as hypotheses, not a stage deduction; sentiment can also have contrarian interpretations. Margin debt remains below June's peak. Direct participation, positioning and leveraged-product flows would help resolve it.
• Concentration and evidence grades. The draft's 38.4–39.9% range mixed securities with companies; the review reconciles the units and grades the tertiary proxy D pending S&P confirmation (claim 13). It also downgrades the unparsed household-allocation claim to D (claim 38). By contrast, FactSet and Renaissance are primary for their own tallies/estimates, not secondary merely because they are private publishers.
• Unverified inputs and loss mapping. Futuriom growth estimates, the Morgan Stanley debt forecast and private-lab listing dates retain their reported/forecast status (claims 29, 32, 37). The off-balance-sheet SPV and private-credit layer is unsized. This gap and the collateral/guarantee map matter directly to who bears losses.
• Margin debt's July drop. Claim 54 establishes the largest absolute-dollar fall in the downloaded history, not the largest percentage decline. Forced deleveraging versus reporting/reclassification remains open; FINRA cautions that reporting-method changes can affect monthly balances.
• October credit discipline versus an imminent-crash reading. New CoreWeave and Oracle quotes make the earlier qualitative sector-repricing description materially incomplete (claims 59–60), but S&P sees no current need to change ratings despite its opacity warning (claim 61). Lloyd calls the CoreWeave selloff excessive while owning its bonds (source 50); that interested view does not invalidate prices. Elice raised capital despite a weak debut; Griffin’s warning and Fink/Sewing’s competing views prevent a consensus sentiment inference (claims 62–63). Yardeni’s low-trigger backtest is also counterevidence to a mechanically imminent bust, not a validated probability for this report (claim 64). Test funding access and cash conversion directly.

Signals to watch

Sources

1. Federal Reserve, "Federal Reserve issues FOMC statement", 16 Sep 2026. https://www.federalreserve.gov/newsevents/pressreleases/monetary20260916a.htm (accessed 2026-10-17)

2. Federal Reserve, "Federal Reserve issues FOMC statement", 29 Jul 2026. https://www.federalreserve.gov/newsevents/pressreleases/monetary20260729a.htm (accessed 2026-10-17)

3. Federal Reserve, "Open Market Operations" (policy-rate history). https://www.federalreserve.gov/monetarypolicy/openmarket.htm (accessed 2026-10-17)

4. Federal Reserve, "Financial Stability Report, May 2026: Overview" (data as of 23 Apr 2026). https://www.federalreserve.gov/publications/2026-may-financial-stability-report-overview.htm (accessed 2026-10-17)

5. FINRA, "Margin Statistics". https://www.finra.org/rules-guidance/key-topics/margin-accounts/margin-statistics (accessed 2026-10-17)

6. Robert J. Shiller, "Online Data", ie_data.xls, Data sheet, file version 1791297298731. https://img1.wsimg.com/blobby/go/e5e77e0b-59d1-44d9-ab25-4763ac982e53/downloads/1449043e-a1ed-4e2e-8e9c-7b9a2a064493/ie_data.xls?ver=1791297298731 (downloaded and parsed 2026-10-17; October price/GS10 from Oct 2; September/October CPI estimated)

7. multpl.com, "Shiller PE Ratio", 16 Oct 2026. https://www.multpl.com/shiller-pe (accessed 2026-10-17)

8. FRED, "10-Year Treasury Constant Maturity Rate (DGS10)". https://fred.stlouisfed.org/series/DGS10 (accessed 2026-10-17; full CSV via fredgraph.csv)

9. FRED, "10-Year Treasury Inflation-Indexed Security (DFII10)". https://fred.stlouisfed.org/series/DFII10 (accessed 2026-10-17)

10. FRED, "ICE BofA US High Yield Index Option-Adjusted Spread (BAMLH0A0HYM2)". https://fred.stlouisfed.org/series/BAMLH0A0HYM2 (accessed 2026-10-17)

11. FRED, "ICE BofA US Corporate Index Option-Adjusted Spread (BAMLC0A0CM)". https://fred.stlouisfed.org/series/BAMLC0A0CM (accessed 2026-10-17)

12. FRED, "Gross Domestic Product (GDP)". https://fred.stlouisfed.org/series/GDP (accessed 2026-10-17)

13. FactSet, John Butters, "S&P 500 Earnings Season Update: October 16, 2026". https://insight.factset.com/sp-500-earnings-season-update-october-16-2026 (accessed 2026-10-17)

14. AAII, "Investor Sentiment Survey", week ending 14 Oct 2026. https://www.aaii.com/sentimentsurvey (accessed 2026-10-17)

15. interactive investor, "Fund manager optimism faltering" (BofA Global Fund Manager Survey, October 2026), 13 Oct 2026. https://www.ii.co.uk/analysis-commentary/fund-manager-optimism-faltering-ii540567 (accessed 2026-10-17)

16. Renaissance Capital, "US IPO Market: 3Q 2026 Quarterly Review", 1 Oct 2026. https://www.renaissancecapital.com/review/3Q26USReview_Public.pdf (accessed 2026-10-17)

17. Epoch AI, "Hyperscaler capex vs cash flow", 16 Jun 2026. https://epoch.ai/data-insights/hyperscaler-capex-vs-cash-flow (accessed 2026-10-17)

18. Apollo Global Management, Torsten Slok, Rajvi Shah, Shruti Galwankar, "The extreme weight of AI in the S&P 500", Jan 2026. https://www.apolloacademy.com/wp-content/uploads/2026/01/011326-MarketConcentration.pdf (accessed 2026-10-17)

19. stockanalysis.com, "S&P 500 Stocks List". https://stockanalysis.com/list/sp-500-stocks/ (accessed 2026-10-17)

20. Ray Dalio, "Are We in a Stock Market Bubble?", LinkedIn, 22 Feb 2021. https://www.linkedin.com/pulse/we-stock-market-bubble-ray-dalio (accessed 2026-10-17)

21. Ray Dalio, "Are We in a Stock Market Bubble?" (March 2024), republished by The Acquirer's Multiple at Hedge Fund Alpha, 6 Mar 2024. https://hedgefundalpha.com/?p=2162956 (accessed 2026-10-17)

22. Fortune, "Ray Dalio says AI is in the early stages of a bubble, so watch out for 2026", 6 Jan 2026. https://fortune.com/2026/01/06/ray-dalio-says-ai-is-in-the-early-stages-of-a-bubble-so-watch-out-for-2026/ (accessed 2026-10-17)

23. Fortune, "Ray Dalio: AI bubble, 1929, 2000, IPOs, 'wealth is not money'", 4 Aug 2026. https://fortune.com/2026/08/04/ray-dalio-ai-bubble-1929-2000-ipos-wealth-is-not-money/ (accessed 2026-10-17)

24. Bloomberg Law, "Dalio Sees AI Bubble Bursting as Wealth Is Converted Into Money", 3 Jun 2026. https://news.bloomberglaw.com/financial-accounting/dalio-sees-ai-bubble-bursting-as-wealth-is-converted-into-money (accessed 2026-10-17)

25. Hyman P. Minsky, "The Financial Instability Hypothesis", Levy Economics Institute Working Paper No. 74, May 1992. https://www.levyinstitute.org/pubs/wp74.pdf (accessed 2026-10-17)

26. Charles P. Kindleberger, Robert Z. Aliber, Robert N. McCauley, Manias, Panics, and Crashes: A History of Financial Crises, 8th ed., Palgrave Macmillan, 2023. https://link.springer.com/book/9783031160073 (accessed 2026-10-17)

27. Irving Wladawsky-Berger, "What Is AI's Place in History?" (quoting Carlota Perez, Project Syndicate, March 2024), 24 Dec 2025. https://blog.irvingwb.com/blog/2025/12/what-is-ais-place-in-history.html (accessed 2026-10-17)

28. Paul Kedrosky, "Why AI Capex isn't a Bubble: A Perez-ian Perspective", 9 Oct 2025. https://paulkedrosky.com/why-ai-capex-isnt-a-bubble-a-perez-ian-perspective/ (accessed 2026-10-17)

29. Oracle, "Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues", 10 Sep 2026. https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Q1-Results-Driven-by-Triple-Digit-Growth-in-Cloud-Infrastructure-Revenues/default.aspx (fetched 2026-10-17; replaces the draft's secondary mirror)

30. CoreWeave, Inc., Form 10-Q for the quarter ended 30 Jun 2026, SEC EDGAR. https://www.sec.gov/Archives/edgar/data/0001769628/000176962826000366/crwv-20260630.htm (accessed 2026-10-17)

31. Allwork.space, "Morgan Stanley forecasts $570 billion AI debt wave as hyperscalers race to fund growth", 10 Jun 2026. https://allwork.space/2026/06/morgan-stanley-forecasts-570-billion-ai-debt-wave-as-hyperscalers-race-to-fund-growth/ (accessed 2026-10-17)

32. Futuriom, Mary Jander, "Hyperscaler AI Spending Doubts Rising", 6 Apr 2026. https://www.futuriom.com/articles/news/ctp-is-hyperscaler-ai-spending-sustainable/2026/04 (accessed 2026-10-17)

33. thetrading.tools, "Household equity allocation record" (Fed Z.1 derived), 14 Aug 2026. https://www.thetrading.tools/chart-of-the-day/2026-08-14-household-equity-allocation-record (accessed 2026-10-17)

34. Apollo, Torsten Slok, "Consensus just revised up capex for hyperscalers", 28 Jul 2026. https://www.apollo.com/insights-news/insights/daily-spark/consensus-just-revised-up-capex-for-hyperscalers (accessed 2026-10-17)

35. Bridgewater Associates, Ray Dalio, "Stock Market Bubble?", 2 Feb 2021. https://www.bridgewater.com/research-and-insights/ray-dalio-stock-market-bubble (fetched 2026-10-17; original six measures and timing caveat)

36. Federal Reserve, "Minutes of the Federal Open Market Committee, September 15–16, 2026", released 7 Oct 2026. https://www.federalreserve.gov/monetarypolicy/fomcminutes20260916.htm?trk=article-ssr-frontend-pulse_little-text-block (fetched 2026-10-17)

37. Renaissance Capital, "US IPO Market: 2Q 2026 Quarterly Review", 1 Jul 2026. https://www.renaissancecapital.com/review/2Q26USReview_Public.pdf (fetched 2026-10-17)

38. Microsoft, FY26 Q3 cash flows (quarter ended 31 Mar 2026), released 29 Apr 2026: https://www.microsoft.com/en-us/Investor/earnings/FY-2026-Q3/cash-flows; FY26 Q4 earnings release (quarter ended 30 Jun 2026), 29 Jul 2026: https://www.microsoft.com/en-us/Investor/earnings/FY-2026-Q4/press-release-webcast (fetched 2026-10-17). The review's SEC company-facts endpoint, https://data.sec.gov/api/xbrl/companyfacts/CIK0000789019.json, was unavailable on re-fetch; issuer cash-flow tables verify the same inputs.

39. Amazon, Form 10-Q, quarter ended 30 Jun 2026. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm (fetched 2026-10-17; Q1 = H1 less Q2). Review reconstruction endpoint: https://data.sec.gov/api/xbrl/companyfacts/CIK0001018724.json (unavailable on re-fetch; filing used).

40. Alphabet, "Alphabet Announces First Quarter 2026 Results", 29 Apr 2026: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000043/googexhibit991q12026.htm; Form 10-Q for quarter ended 30 Jun 2026: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm (fetched 2026-10-17; Q2 = H1 less Q1). Review reconstruction endpoint: https://data.sec.gov/api/xbrl/companyfacts/CIK0001652044.json (unavailable on re-fetch; filing/release used).

41. Meta, "Meta Reports First Quarter 2026 Results", 29 Apr 2026: https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-First-Quarter-2026-Results/default.aspx; Form 10-Q for quarter ended 30 Jun 2026: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm (fetched 2026-10-17; Q2 = H1 less Q1). Review reconstruction endpoint: https://data.sec.gov/api/xbrl/companyfacts/CIK0001326801.json (unavailable on re-fetch; filing/release used).

42. CoreWeave, "CoreWeave Prices Upsized $3.7 Billion Convertible Senior Notes Offering", 18 Sep 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Prices-Upsized-3-7-Billion-Convertible-Senior-Notes-Offering/default.aspx (fetched 2026-10-17)

43. CoreWeave, "CoreWeave Announces At-the-Market Offering Program", 17 Sep 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-At-the-Market-Offering-Program/default.aspx (fetched 2026-10-17)

44. CoreWeave, "CoreWeave Continues to Contract New Compute Capacity at Higher Prices", 17 Sep 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx (fetched 2026-10-17)

45. Carlota Perez, "What Is AI’s Place in History?", Project Syndicate, 11 Mar 2024. https://www.project-syndicate.org/magazine/ai-is-part-of-larger-technological-revolution-by-carlota-perez-1-2024-03 (accessible introduction fetched 2026-10-17; full text not independently verified)

46. FINRA, "Customer Margin Balances", historical workbook. https://www.finra.org/sites/default/files/2021-03/margin-statistics.xlsx (downloaded and parsed 2026-10-17)

47. Oracle, Form 10-Q, quarter ended 31 Aug 2026. https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm (fetched 2026-10-17; exact cash-flow inputs)

48. CoreWeave, "CoreWeave Reports Strong Second Quarter 2026 Results", 11 Aug 2026. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/ (fetched 2026-10-17; distinguishes net interest expense from cash paid)

49. CNBC, “CoreWeave bond yields stay near 15% as AI stocks rebound,” 19 Oct 2026, updated 20 Oct. https://www.cnbc.com/amp/2026/10/19/coreweave-oracle-ai-credit-stress.html (full AMP body re-fetched 2026-10-21; Monday quote dates explicit).

50. CNBC, “This Janus Henderson fund manager owns CoreWeave debt — here's why he's sticking with it,” 20 Oct 2026. https://www.cnbc.com/amp/2026/10/20/coreweave-debt-janus-henderson-lloyd.html (full AMP body re-fetched 2026-10-21; quotes are secondary snapshots).

51. Financial Times via Financial Post, “AI debt worries spark surge in default insurance trading,” 20 Oct 2026. https://financialpost.com/financial-times/ai-debt-worries-spark-surge-in-default-insurance-trading (full syndication re-fetched 2026-10-21; Oracle quote dated Monday Oct 19).

52. Reuters via Boursorama, S&P/Mark Habib on AI liabilities, 19 Oct 2026. https://www.boursorama.com/bourse/actualites/les-principaux-hyperscalers-pourraient-avoir-un-endettement-lie-a-l-ia-plus-important-qu-on-ne-le-pense-actuellement-met-en-garde-s-p-global-83db047e4cfe79ccfdc7bad05d42e61b (full French syndication re-fetched 2026-10-21; secondary attribution, no primary transcript located in linked record).

53. S&P Global Ratings, “AI Infrastructure Investment To Exceed $1.3 Trillion By 2027,” 27 Aug 2026. https://press.spglobal.com/2026-08-27-AI-Infrastructure-Investment-To-Exceed-1-3-Trillion-By-2027,-S-P-Global-Ratings-Says (primary release re-fetched 2026-10-21; older scenario already verified in the Oct 18 signal).

54. The Next Web, “Korean AI data centre firm Elice sinks 32% in a record-worst IPO debut,” 20 Oct 2026. https://thenextweb.com/news/elice-ipo-32-percent-debut-korea-ai-data-centre-neocloud (full body re-fetched 2026-10-21; Reuters/Bloomberg reporting, not independently fetched exchange tape or prospectus).

55. Bloomberg via The Edge, “Finance leaders spar over AI bubble after warning by Griffin,” 21 Oct 2026. https://theedgemalaysia.com/node/822715 (full syndication directly re-fetched 2026-10-21; attributed remarks, not primary event transcript).

56. Yardeni Research, Morning Briefing, 21 Oct 2026, “AI I–III” bubble-meter sections. https://archive.yardeni.com/morning-briefing-2026/ (full primary archive directly re-fetched 2026-10-21; verifies publication of readings/backtest, not independent replication).

57. CoreWeave, 8-K, 22 Sep 2026, completed convertible offering. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000432/crwv-20260917.htm (re-fetched 2026-10-21 with generic research-bot User-Agent; older closing already in the Oct 17 signal).

Review applied: Independent Codex cross-check, 2026-10-17. Its scope was selected claims and Analysis arithmetic, not independent verification of every original claim. New figures above were re-fetched from originating sources; original unverified inputs retain their stated limitations.