Weighing It
Conclusions: Where This Probably Goes
This page applies the frameworks from the research to the evidence on the rest of the site. It offers pattern-matches and watchable criteria, not forecasts, and it leads with the reason for that humility.
Updated
The historical base rate is that even correct bubble calls ran one to three or more years early, at large opportunity cost. Greenspan's 1996 warning preceded roughly 300% of further Nasdaq gains; housing calls from 2002-03 preceded three to four more years of rising prices; and manias have typically run another 12-24 months even after warnings became mainstream consensus (DQYDJ (opens in a new window), NBER (opens in a new window); full treatment on the Case Against page). Nothing below is a claim about dates, and history suggests you should distrust anyone who offers one.
1. The scorecard, as the research grades it
These statuses restate the Warning Lights dashboard exactly as the underlying research assesses them (readings 2025 through mid-2026), neither upgraded nor downgraded:
| Indicator | Status | One-line reading |
|---|---|---|
| Market concentration & breadth | Flashing | Top-10 index weight at 37.6% to 39.1% (fund-holdings data, Aug 2026); beyond prior-era norms. The site's earlier 40.7% figure is dropped for lack of a locatable methodology. |
| Margin debt & leverage | Amber | Record $1.502T (Jun 2026), then -5.65% to $1.417T (Jul), the largest one-month drop on record, +38.6% YoY (FINRA (opens in a new window)). |
| New issuance | Amber | Public IPOs below 1999 levels; the $50-80B private-credit figure is Bloomberg's own estimate, no located methodology, not scored as a number (Bloomberg (opens in a new window)). |
| Central-bank policy | Mixed | Easing after aggressive hikes; historically compatible with melt-up or plateau (Fed H.15 (opens in a new window)). |
| Capex vs operating cash flow | Amber / red in spots | No consolidated Big-5 aggregate exists; Oracle's negative FCF to 2029 is a forecast, not an observed reading (Calcbench (opens in a new window)). |
| Credit spreads | Quiet (sector amber) | Broad spreads tight; mild widening in AI-infra credits only (Bloomberg (opens in a new window)). |
| Credit-to-GDP gap | Quiet | Negative since 2021, at -11.5 points (2025 Q4); the calmest light on the board. |
| Insider selling | Not scoreable | Late-cycle pattern at AI winners; no aggregate metric exists against any defined baseline, so not scored (MarketWatch (opens in a new window)). |
| Valuations | Not scoreable | No locked vendor, timestamp, or earnings basis for a forward-P/E reading, so not scored (Britannica (opens in a new window)). |
Summary: sector-level excess (concentration, insiders, pockets of capex strain) without the economy-wide credit boom that marked 2008. Closer to 1998's board than to 2007's.
2. The circularity clock
Across five historical episodes, circular financing extended booms roughly 2 to 4 years beyond organic demand before a disclosure or liquidity event, a write-down, a cash call, a forced consolidation, exposed the loop (CNET/McKinsey (opens in a new window), Odlyzko (opens in a new window)). The major AI circular structures, Nvidia's customer stakes, the Microsoft-OpenAI commitments, the SPV and GPU-debt complex, date mostly to 2023-2024 (Bloomberg (opens in a new window)). If the historical pattern held, the exposure window would fall roughly in 2026 through 2028. That is a pattern-match, not a prediction: the sample is five episodes, the structures differ in real ways, and nothing obliges history to keep its own schedule. What the pattern does justify is watching for the specific exposure events listed under the scenarios below.
3. The closest analog: telecom 2000, with two honest differences
Of every episode in the catalog, the telecom and fiber bubble matches the AI boom most closely on structure: massive infrastructure overbuild plus vendor-style financing plus a genuinely transformative underlying technology (7GC (opens in a new window)). Two documented differences pull in opposite directions, and both deserve to stand:
- Today's spenders are profitable incumbents, funding the build mostly from operating cash flow rather than debt, which argues for resilience: they can absorb a failed bet without failing (Fortune/BofA (opens in a new window); the full argument).
- GPUs [depreciate](/glossary#depreciation) far faster than fiber, losing most of their value in 3-5 years versus decades for glass, which argues for fragility: the window for demand to catch up is compressed to a single hardware generation (ServerBuyback (opens in a new window); the depreciation evidence).
4. Three scenarios, and the evidence that selects between them
The research frames three futures, without probabilities. Attached to each are the specific, measurable 12-24 month criteria the research identifies (InnovaPath review (opens in a new window); criteria catalog in the Case Against research), so you can watch the evidence yourself rather than trusting anyone's vibes, including this site's.
The clock did not stop in March 2026. OpenAI's $122 billion raise (opens in a new window) that month refuted, on its own timetable, the existential-funding prediction its most prominent critic, Ed Zitron, made in 2023-24 (the full record). He has not dropped the thesis; he has re-dated it. His most recent post (opens in a new window) predicts OpenAI will need to raise again, near the same scale, by March 2027, and repeatedly through 2030. That dated prediction is now itself part of what these scenarios track: watch whether it lands, the same way you would watch any other criterion below.
Scenario A: Correction and consolidation
The pattern the bubble literature calls most likely for a real general-purpose technology: valuations correct meaningfully, weak and over-leveraged players (some neocloudsA newer, smaller cloud company built specifically to rent out AI computing power (GPUs), often financed with debt secured against the chips themselves.Full definition in the glossary, thin-margin wrappers) fail or get absorbed, the giants write down some stakes, and the technology keeps compounding under new ownership structures. You are watching this happen if: AI-linked equities correct without a 70-80% sector collapse; defaults stay contained to the weakest borrowers; hyperscaler capex growth slows toward cash flow; and AI revenue keeps growing through the shakeout.
Scenario B: Sector bust, telecom-2001 style
The loop breaks the way it broke in 2001. Confirmation criteria from the research: hyperscaler AI segments missing revenue targets by wide margins while aggregate free cash flow goes negative; credit eventsThe extra interest a risky borrower pays compared to the safest borrower (the US government). Widening spreads mean lenders are getting nervous.Full definition in the glossary among neoclouds, data-center SPVsA separate legal company created to hold one project and its debt, keeping both off the parent company's books.Full definition in the glossary, or GPU-leasing firms; an AI-heavy index suffering a 70-80% drawdownHow far a price falls from its peak to its lowest point, in percent. The dot-com crash was a 78% drawdown for the Nasdaq.Full definition in the glossary; vendors writing down equity stakes in customers or tearing up offtake agreementsA long-term contract to buy a project's future output, signed before it's built. Lenders rely on these promises, so the buyer's health becomes the project's foundation.Full definition in the glossary (the Lucent moment); used-GPU prices collapsing to 20-30% of list within 3-4 years alongside persistently low utilizationHow much of the time expensive equipment is actually being used. Low GPU utilization means capacity was built ahead of real demand.Full definition in the glossary; or a macro shock (oil, Taiwan, the grid) that AI-linked assets absorb disproportionately.
Scenario C: Demand catches up
The build-out is roughly right, and the bubble label was wrong. Falsification criteria for the bubble thesis: hyperscaler AI revenue scaling from tens of billions toward $200-300 billion or more annually by 2028 at margins comparable to cloud software; enterprise AI returns turning broadly positive (more than half of organizations reporting a positive EBIT impact from AI within 12-24 months, up from McKinsey's current 37% reporting any EBIT impact (McKinsey (opens in a new window))); capexMoney a company spends on long-lived physical things: buildings, machines, chips, data centers. Spent now, paid back (hopefully) over years.Full definition in the glossary stabilizing below 70-80% of operating cash flow with free cash flow turning up and less reliance on private creditLoans made by investment funds instead of banks, with little public disclosure. Tens of billions of it now finances AI data centers.Full definition in the glossary; and no major AI credit event even through macro shocks.
Where that leaves the question
The honest synthesis of everything on this site: the AI boom shows several classic late-stage bubble signatures (record concentration, circular financing, capex far ahead of counted demand) layered on a real, fast-monetizing technology funded mostly by unusually strong balance sheets, with the best-tested economy-wide crisis indicators still quiet. History's most common resolution for that combination is overvaluation, correction, consolidation, with the value ultimately accruing to later owners and users rather than peak-price builders, on a timetable that has embarrassed everyone who claimed to know it. Watch the criteria, not the commentary.