The AI Bubble Against the Dot-com Bubble

First edition: 2026-07-02 ｜ Data refreshed through: the 2026-10-16 close (all three markets in sync; weekly refresh of 10-19; external trigger readings to 10-18 23:55 UTC; §2.5 / §2.6 are topical snapshots from the 10-15 interim refresh and keep their original date) Method: a deep-retrieval workflow (106 agents / 24 sources / 115 candidate claims → 25 adversarially verified → 22 confirmed, 3 rejected) + first-hand data across three markets (A-shares / Hong Kong / US, each on its own five-year valuation percentile) + the academic framework (Pastor–Veronesi rational bubbles, Greenwood–Shleifer–You). Date conventions: the 2000 history (§1) is fixed fact; institutional figures (IMF / Goldman / GMO etc.) keep their own report dates; this page's first-hand readings (§2.1 / §2.2 / §2.4 / §4.4) are refreshed wholesale to the data date each edition; old values are not kept. Earlier editions are in the archive at the foot of the page.

Evidence grades: ✅ = fact confirmed by adversarial verification ｜ ⚠ = institutional model estimate or opinion (source checkable, not hard fact) ｜ ❔ = accepted history not put through this round's three-vote check ｜ ❌ = rejected by adversarial verification (do not cite).

One-line conclusion: the AI trade is most likely a "real technological revolution plus a localised, layered bubble", not a 2000-style wholesale speculative mania. The timing of any break is academically unforecastable; the right posture is not to guess a date but to diagnose bubble intensity by layer (the three-market data happen to give three temperatures) and monitor trigger conditions. The most transferable lesson of 2000: the break was about 31 months of staged declines, not a one-day crash, and the bottom came about 2.5 years after the top — any response has to be scaled to the whole clearing period, not to the first leg down.

About this public edition: this page is generated by mechanically cropping a research source (source 23c7a02, 2026-10-19). It contains no action plan, positions or stock picks; it keeps only the historical comparison, the current check-up, the scenario framework and the leading-indicator dashboard, with numbers and evidence grades unchanged. Not investment advice.

1. Anatomy of the 2000 dot-com bubble (verified data)

1.1 Crash timeline and shape ✅

Source: Wikipedia (Stock market downturn of 2002 / Nasdaq Composite); annual declines re-computed from independent year-end closes and consistent.

Core conclusion on shape ✅: the break was a staged grind lower (three annual declines of −39% / −21% / −32%, with several 20–40% rallies in between), not a single crash. There was no "bottom day" — the bottom came about 2.5 years after the top. Money that went in after the first sharp drop most likely still had one to two years of pain ahead.

1.2 Valuation at the top ✅

• NASDAQ-100 trailing PE ≈ 250 at the top (⚠ trailing basis; forward about 60× — any comparison with today must use the same basis, or the gap is exaggerated). Source: Nasdaq official documents.
• S&P CAPE peak 44.2 (1999-12, Shiller monthly basis), an all-time record; the mean since 1880 is 17.6.
• ❔Accepted background (not separately verified this round): the S&P 500 fell about −49% peak to trough over the same period — the bubble was not confined to the Nasdaq (the claim that it was lost 0–3 ❌).

1.3 IPO mania: the quantifiable signature of a top ✅

Source: WilmerHale's contemporaneous 1999-12-31 report, cross-checked against Jay Ritter's IPO database.

Implication ✅: when these numbers appeared, the top was only 10 weeks away — gains and mania themselves give no timing (the stretch before a top is exactly the most profitable), but they are after-the-fact checkable "accompanying features of a top" and serve as a benchmark for today's AI issuance heat.

1.4 The world after the break ❔ (accepted history from retrieved secondary sources, not three-vote verified)

• Cisco: about −89% from peak; its first closing high above the 2000 level came on 2025-12-10 (CNBC) — about 25.7 years.
• Nasdaq round trip: about 15 years to regain 5,048 (2015-04).
• Amazon: $107 → $5.97 (2001-10), about −94%; then one of the biggest winners of the following decade and more.
• Microsoft / Intel: drawdowns of −60% to −80%, recoveries of 10+ years.
• Infrastructure glut: debt-financed telecom fibre ran at low utilisation for years (commonly cited <5%); the telecom casualty list was longer than the internet one (Telecoms crash).
• Value / defensives: value beat growth by a wide margin in 2000–2002 (Fama-French HML strongly positive) — a bubble's unwinding is a harvest season for value investing.
• Buffett: mocked for refusing tech in 1999; in 2002–2003 turned to buying mispriced credit, including Amazon junk bonds (Bloomberg 2003-04: Buffett Lauds Amazon's Courage, Then Buys Its Junk) — the best move after a break is not necessarily buying equity; it can be buying mispriced high-yield debt.

1.5 Recovery base rates: heavily right-skewed ✅

Statistics on the maximum drawdowns of 6,500+ US stocks, 1985–2024 (Morgan Stanley, Mauboussin & Callahan 2025):

• The average recovery reaches about 340% of the prior peak; the median recovery reaches only about 90% — the median stock never regains its prior high.
• Among stocks that do regain the high, the median trough-to-high time is 2.5 years.
• ⚠Note: the statistic covers all US drawdowns 1985–2024 (not 2000–2002 specifically); two claims from the same paper, "median drawdown 85%" and "54% never recover", failed adversarial verification (1–2 votes) ❌ — do not cite those numbers.

Implication: the outsized returns after a break are concentrated in a very small number of high-quality survivors (the Amazon type). "Buy whatever fell the most" is mathematically doomed to mediocrity or permanent loss; "select survivors with real cash flow" is the only way to be on the right side of a right-skewed distribution.

2. Check-up of the AI market today (mid-2026)

2.1 Aggregate valuation and concentration

Reading: aggregate valuation is at "second most expensive ever, behind only 2000", but still some distance from the 2000 peak (the IMF also judges the deviation smaller than historical peaks ⚠); meanwhile concentration and household transmission exceed 2000 — if it breaks, the path is not "a Nasdaq pocket" but "index-level plus household-wealth-level".

2.2 Capex, debt and circular deals (the part that most resembles 2000 telecoms)

• ✅Capex scale: Goldman's 2026 supply-side model — AI-related annual capex about $765bn in 2026, base case ~$1.6tn/yr by 2031; highly sensitive to chip-life / utilisation assumptions (shorter life or lower utilisation → disappointing return on invested capital). IMF: cumulative ~$3.4tn by 2029 (⚠ adopted analyst forecasts), about 70% borne by hyperscalers.
• ✅Debt-financed and starting to stratify (07-20 → 08-31 milestones; all figures first-hand or verified secondary at the time): 2026 investment-grade issuance by the big six (Amazon/Alphabet/Nvidia/Meta/Oracle/SpaceX) cumulatively ~$182bn (+1,300% YoY), all into data centres and AI infrastructure. Sequence of credit stratification: 07-20 Oracle 5y CDS 75bp (7-year high) → 07-27 ~203bp → 07-28 S&P cuts Oracle to BBB- (last IG notch), FY26 capex $55.7bn → 07-29 Oracle FY26 free cash flow −$23.7bn, total debt ~$130bn; Moody's names Oracle + CoreWeave "the weakest AI credit links" → 08-03 CoreWeave 5y CDS ~855bp (model-implied 5-year default ~50%, ⚠ single source), Q1 interest expense doubled to $536m, FCF −$4.71bn; Moody's: the six's direct debt ~$460bn + lease commitments $1.2tn → 08-06 lenders start demanding covenants; AI-related CDS volume 2026Q2 +~600% YoY → 08-10 CoreWeave's $2.6bn loan only subscribed at +100–125bp with a 1.35x DSCR maintenance covenant → 08-12 CoreWeave Q2 beats (backlog $104bn) but debt has rolled to $35bn; criterion set as "watch CDS and loan spreads for convergence, not the share price" → 08-17 total debt $35.6bn (✅) → 08-31 new $2.6bn DDTL priced at SOFR+550bp, 100bp wider than May's +450 (✅ SEC 8-K first-hand: same borrower repriced within three months). ⚠ Reading discipline: wider CDS "marks rising risk, not imminent default"; 30Y Treasuries widened ~40bp over the same period, so the net "collateral confidence" signal must be read after stripping the rate beta. This is the closest contemporary analogue to "telecoms borrowing to lay fibre in 2000", and has moved from volume to "volume + stratification + individual distress + tighter terms + repricing".
• 09-14 update (weekly): a [contrary backdrop] appeared this period, and it makes the AI-side anomaly more conspicuous, not less — ✅ broad credit extremely compressed: ICE BofA US high-yield OAS 270bp (09-10), in the richest decile of history (long-run median ~450bp). 🔴 ⚠ But this cannot fill #2's box: #2 watches AI-related credit (CoreWeave/Oracle CDS and loan spreads), where the last reading was CoreWeave's new loan at SOFR+550 (100bp wider than May, ✅ SEC 8-K). 📌 ⇒ In a market where nobody demands a risk premium, a 100bp mark-up for AI borrowers alone carries [more] information than it would during a general widening, not less. ⚠ EDGAR was not checked this period ⇒ whether CoreWeave filed a new 8-K or Anthropic's S-1 appeared: no new reading, no inference. 🔴 But the equity side weakened clearly (✅ this page's first-hand data): CRWV loss-making, ROE −43.6%, 1-month −17.4%, PB 15th pct; Nebius PE 1462.6x / ⚠99th pct. ⇒ The "equity–credit divergence" since August converged this period by [equity moving toward credit], not the other way round — historically (2000 telecoms / 2008) exactly this direction.
• 09-21 update (weekly): #2 gets its third first-hand data point, and the instrument changed — CoreWeave 09-17 8-K (Items 7.01/8.01, ✅ SEC): a proposed $3bn convertible (2033) + $0.5bn option, senior unsecured, with capped call; coupon and conversion price undisclosed ("privately negotiated"). 🔴 Reading: choosing a convertible at 675bp CDS (media-implied 5-year default 42%, ⚠ secondary) means straight debt is too expensive, so future equity is swapped for today's cash ⇒ third stage of credit tightening: May SOFR+450 (pricing) → Aug +550 (repricing) → Sep convertible (instrument change). Equity the same week CRWV −8.6% (88.99→81.36), PB pct 15→5 ⇒ the divergence keeps converging by "equity toward credit". 2026 capex need $31–35bn vs TTM revenue $6.2bn. ⚠ No new first-hand CDS reading; 42% recorded as direction only.
• 09-28 update (weekly): #2's fourth first-hand point — after the instrument change, the demand side's answer is "yes, and more": CoreWeave 09-22 8-K (Item 1.01, ✅ SEC): convertible completed and upsized — $4.2bn 2.875% convertible 2033 (announced 09-17 at $3bn + $0.5bn option → base upsized to $3.7bn, option fully exercised), senior unsecured, conversion premium 22.5% (vs 09-17 close $79.88, conversion price ≈$97.85), capped call cap $199.70 (150% premium), net proceeds $4.137bn, use "general corporate purposes". 🔴 The same 8-K incidentally priced the straight debt: outstanding senior notes carry 8.5% / 9.0% / 9.25% / 9.625% / 9.75% — the 6–7pp between the 2.875% convertible and the 8.5–9.75% notes is the market's price for "future equity". ⚖ Reading: ① last period's "straight debt too expensive, so future equity for today's cash" is confirmed by first-hand coupons; ② but the equity-linked window is not only open, it was oversubscribed by 40% — credit is tight on "debt", not shut on "CoreWeave the company"; ③ equity the same week +7.7% (81.36→87.59), PB pct 5→15, opposite to the prior week's −8.6% — down in the announcement week, up in the completion week: the market prices "can it get the money" above "with what". ⚠ No new first-hand CDS; the earlier 675bp remains direction only. 📌 2000 comparison gap: the shape of the telecom convertible window in 1999–2000 was not evidenced for this report; registered, not cited.
• 10-02 update (interim refresh): nothing new first-hand for #2 — no new CoreWeave 8-K/424B5 on EDGAR since the 09-22 convertible 8-K (checked to 10-02); equity CRWV 88.57 (week +1.1%, PB 16th pct). 🔴 Broad backdrop changed sharply: ICE BofA US HY OAS 266bp (09-21 low) → 280 (09-24) → 312bp (09-30), +46bp in nine days, +32bp in four sessions; IG OAS 77→84bp. ⚠ Same discipline as 09-14: broad spreads do not fill #2's box (#2 watches AI borrowers), but they change the 09-14 premise that "nobody demands a risk premium" — now everyone is beginning to demand one, and the AI borrowers' mark-up (SOFR+550 / 2.875% convertible vs 8.5–9.75% notes) is no longer an outlier but a leading value. #2's grade stays until the next 8-K / first-hand CDS reading.
• 10-05 update (weekly): still nothing new first-hand from AI borrowers for #2 (no CoreWeave 8-K since 09-23, checked to 10-04); equity CRWV 89.62 (1-month +10.7%, PB 19th pct); broad HY OAS keeps widening: 266 (09-21) → 312 (09-30) → 324bp (10-01), +58bp in nine sessions; IG 77→86bp. ⚠ Same discipline as 10-02: broad spreads do not fill #2's box; but the widening has now lasted two weeks, so the 09-14 premise that "nobody demands a risk premium" is formally void — the next AI-borrower pricing (8-K / CDS), landing in a widening market, is the criterion for upgrading #2.
• 10-12 update (weekly): still nothing new first-hand for #2 (no CoreWeave 8-K in October, checked to 10-11); equity CRWV 82.08 (1-month −7.9%, PB percentile 19→6, back to the level of the 09-21 convertible-announcement week); broad HY OAS 324 (10-01) → 303 (10-06) → 315bp (10-08), IG 82 — the widening has paused but not returned to the 266 starting point. ⚠ Broad spreads do not fill #2.
• 10-19 update (weekly): #2 — the equity/credit divergence is over; stocks and bonds now fall together — 🔶 CoreWeave's two unsecured issues due 2031/2032 fell below 80 cents on 10-14 (distressed territory); the 2032 notes (issued near par four months ago) traded at about 78 cents, yielding >15%; CDS-implied five-year default probability back at a record >50% (Jin10 / Sina and others citing Trace and Bloomberg, ⚠ second-hand); the shares fell 15.5% the same day and closed at 61.39 on 10-16, PB percentile 0. ✅ EDGAR: no 8-K / 424B5 in October, only two small executive Form 4 sales (10-12/10-13, about 44k shares, about $3.5m in total — not a credit signal). A CEO internal memo says backlog was $104.2bn at end-Q2 with >$25bn added in Q3, enough to cover the secured and parent-company unsecured debt. ⚠ Broad market: HY OAS 311 (10-09) → 325 (10-14) → 323bp (10-15), IG 83→84bp (✅ FRED) — the broad market widened only slightly, the spread gap between AI borrowers and the broad market is the widest of this cycle.
• ✅Circular deals have been officially quantified: the IMF estimates that of the +12pp equal-weight return of the AI circle (Amazon/AMD/Alphabet/Intel/Microsoft/Nvidia/Oracle) from 2025-09 to 12, 7pp is attributable to self-reinforcing intra-circle correlation (⚠ the IMF's own arithmetic is questionable: 7% × $2T ≈ $140B while the text says $40B — the true contribution may be understated ~3×). This is the modern form of vendor financing.
• ⚠Revenue / investment mismatch (2–1 votes, medium confidence): GMO cites estimates of 2025 AI native / application-layer revenue <$50bn vs $1tn+ invested that year; by mid-2026 OpenAI + Anthropic run-rate ~$55bn, 34 AI startups ~$80bn combined (Epoch AI); 08-06 update: Anthropic alone at a $47bn run-rate ($30bn at the start of the year, $10bn for all of 2025, driven by Claude Code) — nearly 4× in a year for one company; the mismatch is closing faster than previously estimated, the strongest new evidence on the "real technological revolution" side; but the mismatch is still an order of magnitude. Basis note: excludes the pick-and-shovel sellers (NVIDIA data centre alone annualises above $300bn).

2.3 The most important structural difference from 2000 ✅

NVIDIA fiscal 2027 Q1 (to 2026-04-26, SEC 8-K confirmed): quarterly revenue $81.6bn (+85% YoY), of which data centre $75.2bn (+92% YoY).

The 2000 top's valuations rested on loss-making dot-coms and "eyeballs"; today's AI infrastructure-layer valuations rest on realised, audited revenue and cash flow. This is the strongest reason the "78% wholesale crash" scenario cannot be transferred directly — and the core basis on which arXiv 2606.01575 (2026-05) judges "real technological revolution + localised layered bubble".

2.4 Three-market first-hand data: three temperatures of one theme (data to the 2026-10-16 close)

This page's own cross-sectional evidence — the same AI theme runs at entirely different valuation temperatures in three markets (A-share percentiles computed daily on local five-year series; 10-19 report basis, all three markets in sync on 10-16):

Four key marginal changes on 10-16:

① 🔴 #9 upgraded to "rotation confirmed" (pre-registered criterion triggered): style gap over three periods +7.8→+9.3→+18.4pp, core 1-month median −13.9%; crowding 41.6% / 59th pct still 🟢.

② 🟢 The compute chain's hot layer has largely cleared: Hygon's percentile 54→26; the only names still high are WUS (87) and TFC (86). ⚠ Opposite in direction to ① — valuation says "already cheap", structure says "still rotating".

③ 🔴 AI-theme stock pool down four weeks running (489→425): AI-theme stocks fall below the market-cap threshold and drop out mechanically — the size measure points the same way as #9.

④ 🔴 #2's primary-market consequence has arrived: CoreWeave's 2031/2032 unsecured notes below 80 cents, the 2032 notes yielding >15% (several outlets citing Trace quotes), equity PB percentile 0 — the equity/credit divergence is over; stocks and bonds fall together.

2.5 The periphery breaks first: Korea's circuit breaker (interim refresh, 10-15)

Facts (🔶 Xinhua / Sina / Korean media; index and single-stock closes cross-checked against Yahoo): on 10-15 the KOSPI closed at 5,554.22, −7.57%; an 8% intraday fall triggered a 20-minute circuit-breaker halt (the tenth this year), and the KOSDAQ halted the same day. Seven straight down days total −20.7% (from 7,004), a −39.1% drawdown from the 06-22 high of 9,114.55, yet the year is still +31.8% (2025 close 4,214.17). Samsung −9.48%, SK hynix −11.06%. Foreign investors net sold about ₩3.64 trillion (about $2.7 billion) of KOSPI stock on the day (closing basis; Pinpoint News trading table ₩3.635 trillion, consistent with Field News and others). From the start of the year to 10-12, foreigners net sold ₩199.1 trillion of KOSPI stock, 87% of it in Samsung plus SK hynix (Chosun Ilbo citing the Korea Exchange, via The Paper). Korean power-equipment names fell hard as well (Hyosung Heavy −14.7%, LS Electric −12.1%, HD Hyundai Electric −10.3%). ⚠ Corrections: ① the widely circulated "Samsung −9.9% / SK hynix −11.3%" were pre-close prices; the closes are as above; ② the first 10-15 version of this section said "foreign selling ₩4.82 trillion, ₩16.1 trillion over eight days", taken from Wall Street CN; it does not match Korea's own closing statistics and has been withdrawn, as has the first version's claim that "₩3.6 trillion is wrong" (pointed out by cross-review).

The US periphery on the same day (10-14 closes):

Macro the same day: September CPI 3.7% (August 3.4%), core 2.6% (August 2.45%), core month-on-month +0.33% (✅ BLS first-hand); 10Y closed at 5.19%, HY OAS 311→325bp (10-09→10-14, ✅ FRED). ⚠ "10Y at 5.32% intraday" and "CoreWeave bond yield 14.8%" appear only in Wall Street CN; unverified, not used.

Five implications for this study (this page's judgement): 1. "The bubble sits in the supply chain" is borne out, and the break runs from the outside in: the deepest falls are in the most levered, most cyclical periphery (neoclouds, Oracle, memory), while the demand-side core, Nvidia, is only −7.5% from its high — consistent with §2.4's layered call since July that the heat is on the supply side. 2. What fell was valuation and funding, not earnings: Samsung's preliminary Q3 operating profit was ₩107.4 trillion (🔶 Samsung's 10-07 preliminary results, via Kyunghyang, Dong-A and others; the first Korean company to top ₩100 trillion in a quarter), about 8.8× the ₩12.2 trillion of a year earlier, and the shares fell after the release. Foreign outflows + the end of roughly ₩55 trillion of combined buybacks at the two companies + shrinking retail leverage add up to a hole in the bid, not a collapse in demand — the same family as TSMC's July "capex raise read as bad news": good news is no longer rewarded. 3. #2 credit and #6 inflation deepened on the same day: CoreWeave −54% over 52 weeks, HY spreads +14bp in three days; headline and core CPI both rose, raising the risk of another hike at the 10-28 FOMC. 4. #5 is still not lit: leader earnings still beat; #1 and #5, the two most reliable triggers, are both unlit ⇒ a system-level break is still not confirmed. 5. "AI power" is no safe haven in Korea: power equipment fell with the chips. The defensive end of A-share #9 is green-power operators (utility-like), which is different, so the #9 logic is unaffected for now.

To be tested: ① whether the KOSPI retakes its 200-day average within two weeks (⚠ Chosun Ilbo's count: of stays below the line of ≥20 trading days, about 64% became declines of three months or more); ② the 10-28 FOMC; ③ whether Nvidia / TSMC catch down — if the core catches down, "the periphery clearing out" becomes a systemic break.

2.6 China's A-share + Hong Kong AI chain: peaked before Korea and fell further (interim refresh, 10-15)

Basis: 10-15 moves are Tencent quote closes; historical ranges, P/E and five-year percentiles come from our own five-year percentile series (to the 10-14 close; percentiles sampled weekly, P/E > 0 only). This section is computed on the spot.

On 10-15 (against KOSPI −7.57%): Shanghai Composite −1.59%, CSI 300 −1.23%, ChiNext −2.92%, STAR 50 −1.50%, Hang Seng −1.03%, Hang Seng Tech −2.72%.

Equal-weight A-share compute basket (Cambricon / Hygon / Zhongji Innolight / Eoptolink / TFC / WUS / Victory Giant / Foxconn Industrial Internet / Montage / Envicool, chained daily pct_chg): peaked on 06-25, almost together with the KOSPI's 06-22 high; drawdown to 10-15 −41.0%, −20.2% over the last 20 trading days, −11.8% over the last 5.

Hong Kong

Readings (this page's judgement): 1. China's AI chain was not dragged down by Korea: the compute basket peaked in late June together with the KOSPI, but had already completed a deeper fall (−41% vs −39%, with the median name about −48% from its high), so the 10-15 shock was comparatively small. 2. The A-share valuation peak has largely cleared: Hygon's percentile fell from 100 in July to 35; the only basket names still high are TFC (87) and WUS (88) — the #8 micro hot layer in §4.4 has cooled markedly. 3. Memory moved with Korea; equipment and foundries went the other way: GigaDevice −67% from its high, CXMT −25% over a month; NAURA and SMIC A closed up on the day — the domestic-substitution logic was not hit by the Korean event. 4. Hong Kong is split to the extreme at both ends: the model companies are about 80% off their highs; Lenovo, +184% on the year, fell 14.5% in a day (the Hong Kong name that most resembles Korean memory); the platforms Tencent and Alibaba are already down about 30% on the year and edged up over the last five days. 5. ⚠ Cambricon's P/E percentile shows 0 and is unusable: it was loss-making for most of its history, so there are too few positive samples.

3. AI bubble vs. dot-com bubble: differences and transferability

4. Probability and timing of a break: an honest framework

4.1 The academic floor: timing is unforecastable ✅

Greenwood–Shleifer–You (JFE 2019; US 1926–2014 + international 1985–2014):

1. A sharp industry run-up on average does not predict low subsequent returns — "it rose too much" alone cannot call a top (the stretch before a top is exactly the most profitable; see the 1999 IPO data).
2. But sharp run-ups raise crash probability markedly: after a 100% two-year rise, the probability of a 40%-class drawdown within two years is about 53% — that precondition (~100% in two years) already holds for the AI leaders.
3. Accompanying features (rising volatility, rising turnover, heavy new issuance, an accelerating price path) help predict the eventual crash.

arXiv 2606.01575 (2026-05) argues explicitly that bubble frameworks are for diagnosis, not deterministic prediction — markets can correct years after showing bubble features, or digest valuations through earnings growth. Any "forecast" naming a specific date is false precision.

4.2 The spectrum of institutional views (mid-2026) ✅

Not one of them gives a date for a break.

4.3 Scenario framework and subjective probabilities (⚠ this report's judgement, not confirmed fact)

Based on the GSY base rate (53% / two years / 40%-class), the layered evidence and the three-market cross-section, subjective scenario probabilities:

If "when" must be answered: the window with the highest probability density is 2027 ± 1 year — not because a date can be computed, but because three evidence lines "hand in their papers" then: ① the chip-life / utilisation assumptions in Goldman's capex model start to be adjudicated by actual depreciation and returns; ② the IMF's "EPS needs 30–35% CAGR" reversion test falls in 2027; ③ if issuance / debt heat is extrapolated at the current slope, it approaches 1999-style readings around 2027. Treat this as scenario reasoning, not a forecast — per §4.1 the timing is inherently unforecastable, and any response hangs on triggers, not on the calendar.

Appendix: ranking of detonators for a crash before 10/31 (worked out 2026-08-22; ⚠ all conditional probabilities — conditional on "a crash occurs", not predictions of a crash)

Framework: timing is unforecastable (§4.1), so this is only a "pre-written post-mortem". Background pattern = "denominator relief (#6 turning dovish) + supply extreme (#3 peaking)", i.e. the 1999Q4 pattern; the 2000-03 top was detonated by the issuance side, not the earnings side. The six calendar-densest weeks in the window = late September to late October (S-1 flip end-August → roadshow September → FOMC 9/15–16 → new fiscal year 10/1 → Anthropic IPO pricing mid-October → FOMC 10/27–28 → big-tech Q3 earnings late October).

Composite judgement: if a crash occurs, the most likely post-mortem is one chain — NVDA short of a "big beat" (8/26) → compute narrative cools → Anthropic roadshow cold / priced at a discount (October) → AMZN/GOOGL revaluation reverses + CRWV spreads blow out → earnings and credit light up together → late-October Q3 earnings confirm. The single highest-weight event = Anthropic's IPO pricing day (a liquidity, accounting and narrative event at once).

Scoring rules, pre-registered (2026-08-24; not to be revised after a trigger): 1. Adjudication date: 2026-11-01; 2. Crash definition (equity side, mechanical threshold): NDX close ≥15% below its highest 2026 close, on any trading day up to and including 10/31 ⇒ "occurred". NDX is chosen because this cycle's AI concentration is highest there and it maps directly onto the chain above; one threshold, no discretion. (Crypto-side counterpart: BTC Binance perpetual close < $58,605, registered separately); 3. Detonator "hit" criterion: within 7 calendar days before the start of the crash's initiation leg, at least one of that detonator's warning indicators has fired and it is the earliest to fire among all detonators. Initiation-leg start (equity side) = during the fall to −15%, the first trading day with a single-day close ≤ −3.5%; absent such a day (slow-grind form), the first trading day the drawdown closes beyond −10%. Equity and crypto post-mortems are written independently then reconciled; disagreements kept as dual records; 4. Three-branch scoring: occurred → score ranking quality + warning lead; did not occur but ≥2 classes of warnings fired → score a false alarm and grade each indicator; did not occur and quiet → examine which counter-argument (the counter section of this §4.3 appendix) did the work; 5. The probability numbers themselves are not scored — a single event cannot calibrate probabilities (the table's probabilities are subjective declarations, calibration history n=0, not to be adjusted after the fact); only the ranking and warning lead are scored, which is all that n=1 can honestly test. 6. Reading rules for adjudication (added 2026-08-30: thresholds untouched, only the reading pinned; not a character of the thresholds has changed since registration): ① denominator basis: "highest 2026 close" = the rolling high as seen from the adjudication date (on any adjudication day t only the highest close in [2026-01-01, t] may be used — if a new high is made before adjudication, the base moves up with it, a natural property of a drawdown definition, not drift, no discretion); ② adjudication series pinned: NDX closes are Nasdaq's official index closes (channel ^NDX; index closes differ from futures settlements and are not revised after the fact, so no T+2 wait; in the rare event of a vendor correction, the corrected value governs with both values recorded); ③ numerator and denominator from the same series on the same basis — no futures/ETF numerator against an index denominator.

⚠Boundary statement on the calendar argument: "September–October is calendar-densest" in this appendix describes event concentration; it is not a basis for action — all-green warnings with the calendar arriving constitute no reason to do anything; the calendar answers only "if it breaks, roughly when", never "will it break".

🔴Scoring the first link of the chain (2026-08-27; NVDA Q2 FY2027 results in):

Result = big beat; the first link did not fire. Revenue $96.2bn (+106% YoY, +4.4% above the $92.2bn consensus and through the top of the $93–95bn whisper range); adjusted EPS $2.22 (expected $2.10); gross margin 75%, no erosion; data centre +117%; Q3 guidance $108bn ±2% (expected $104.2bn, +3.6% above); the call guided FY28 revenue growth ~70%; ⚠ guidance counts China data-centre revenue as zero — any easing of controls is pure upside.

Precise path of the event shock (Binance NVDAUSDT stock perpetual, 5-minute bars, 08-26 UTC): | Time | Price | Event | |---|---|---| | 19:00 | 211.45 | pre-release | | 20:20 | close 208.05, wick low 203.80 (volume 589k = 20× normal) | sell-the-news at release, −3.6% | | 21:00 | 214.87 | call begins | | 21:20→21:50 | 218.69 → peak 220.45 | FY28 +70% guidance ignites | | 23:20 | 218.44 | holds the high |

Reading: ① net event shock +3.2 to 4.3%; the results were not priced in beforehand — but the real incremental information was not the Q2 numbers (absorbed by the whisper, hence −3.6% at release) but the call's FY28 outlook; ② #5 (leader earnings-miss chain) clearly did not light this round: the "expectations above guidance = tiny margin for error" risk was actively repaired before results (sell-side estimates fell back to $91–92B), and the actual result far exceeded even the repaired expectations; ③ by the §4.3 ranking, detonator ③ (NVDA short of a "big beat", ~15% conditional) did not hit — the chain's starting point failed, so ② (credit) and ① (Anthropic IPO) lose this fuse, but their independent paths are unaffected.

⚠Same-day macro headwind: July core PCE +0.2% m/m (below expectations) but 3.3% y/y, headline PCE 3.7% y/y (the expected fall to 3.6% did not happen); Q2 GDP 1.5% not revised down; real consumption ~0% m/m. Fed funds 3.50–3.75%, unchanged in 2026; September cut odds ~1%, hike 33–40%, cumulative hike odds for the year ~70% (largest window 12/8–9) — #6's hawkish undertone is stronger than the mid-August "turning dovish" reading; the trigger must be re-assessed in the next report.

4.4 Leading-indicator dashboard (triggers, ranked by historical reliability)

Dashboard status snapshots (rolling)

5. Limits of the evidence and open questions

1. Rejected claims (do not cite): NDX rose 5× in 1998–2000 and the bubble was confined to the Nasdaq (0–3); median single-stock drawdown 85% / median 2.5 years to bottom (1–2); 54% of stocks never regain their prior high (1–2).
2. Coverage gaps (outline items that did not get three votes, marked ❔ in the text): the Fed's 1999–2000 hiking path, the Microsoft ruling's timing, specific single-stock peak/trough figures, fibre utilisation, the 2000–02 value factor, details of Buffett's behaviour, a quantitative review of the 2002-10 / 2003-03 double bottom, current credit spreads and neocloud debt detail.
3. Basis traps: NDX 250x is trailing (forward ~60x); "AI revenue <$50B" refers to the native layer only and is dated (mid-2026 run-rate $55–80B); the IMF's circular-deal attribution arithmetic is questionable (may understate ~3×).
4. Timeliness: IMF GFSR 2026-04, Goldman 2026-05, NVIDIA results 2026-05, GMO 2026-01 — AI data iterate extremely fast; this page's valuation cross-section is as of 2026-07-02, and trigger readings need updating each weekly edition.
5. Open questions: real-time item-by-item readings of the GSY accompanying features (especially AI issuance share vs 1999's 60%); where responsibility for continuous monitoring of neocloud debt and spreads sits; clarification of the IMF attribution arithmetic.
6. Unmonitorable jump risk: the Taiwan Strait (added 2026-08-10) — this dashboard deliberately admits only "readable" leading indicators (percentiles / thresholds / daily frequency); a Taiwan conflict is an unforecastable jump, and including it would create an illusion of "monitorability", so it stays out of §4.4; its magnitude (external estimates of a full conflict: about −9.6% of global GDP in the first year) is a tail risk outside this framework and must be treated explicitly outside the dashboard.