AI Bubble Watch

The timeline

The AI basket index (100 = ChatGPT's launch on 30 Nov 2022), and the model's window when it has one. "Historical scenarios" adds two past Nasdaq declines, applied to the basket's highest level so far. They show what those episodes would look like on the same scale; they are not forecasts, and they don't estimate the likelihood, timing or size of a future fall.

Company share-price returns

Share prices rebased to 100 at the chosen start (log scale, price only, dividends excluded). The six largest companies are shown first; click a name in the legend to add or remove it.

Korea, Taiwan and Japan

Not part of the basket or any score. Seoul, Taipei and Tokyo trade before New York, so a fall there often shows first. The six companies here are part of the AI supply chain but are listed only at home, so the basket can't include them on the same data. Prices are in local currency, so currency moves aren't included. The chart is rebased to 100 a year ago, with the AI basket (in US dollars) for comparison.

Loading…

Company financials

Select a company to see its share price, revenue and capital spending.

Capital spending and Nvidia's revenue

The bars are five large technology companies' trailing-12-month capital spending: Microsoft, Alphabet, Amazon, Meta and Oracle. The line is Nvidia's trailing-12-month revenue. Together they show the scale of the build-out and of one supplier's sales. They don't show whether the spending earns a return; Funding the build-out looks at how it is paid for. Capital spending here is company-wide cash spending on property and equipment, including non-AI spending and excluding equipment acquired under finance leases, and the two series don't match dollar for dollar. The spending-to-revenue ratio is tracked separately as "Capital spending intensity" in the signals.

Funding the build-out

The spending chart shows how much is going in. This section asks how it is being paid for, and what has been promised for later. AI data centres create obligations that can outlast the spending boom: companies fund them from their own cash flow, by borrowing, and through partnerships and separate project companies, while committing to leases, construction and guarantees. If demand or prices disappoint, those payments may stay due even as the facilities lose value. How much that matters depends on the contracts, what the buildings and chips would fetch, and the sponsor's ability to pay. This section is not part of the Bubble Pressure score: the companies define these items differently, and there is no tested way to weight them yet.

Can they pay for it from what they earn?

Capital spending as a share of operating cash flow, trailing 12 months, from SEC filings. Above 100% means a company is spending more on property and equipment than its whole business generates in cash, and has to cover the gap from savings, borrowing or new shares. That isn't new or automatically a warning: Amazon was above 100% in 2021–22 during its warehouse build-out. What matters is how long it lasts and how it is financed. Capex here excludes equipment acquired under finance leases.

What they have committed to pay later

Leases that have been signed but not yet started are not on the balance sheet. They are disclosed in the notes. Figures as reported in each company's latest filing that states them; dates differ, so treat the total as a rough scale.

Other commitments are disclosed separately and shouldn't be added to the leases, because some overlap: Meta reports a further US$454.5 billion of non-cancelable contractual commitments (mostly cloud capacity, servers and energy), and Amazon US$883.6 billion of total contractual commitments including US$294.3 billion of debt principal and interest. Oracle describes the risk in its own annual report: if it overestimates demand, or key customers can't pay, it "could be locked into multi-year commitments for excess data center space and related capital expenditures, as well as associated financings, without receiving corresponding revenue."

A worked example: Meta's data-centre ventures

In October 2025 Meta formed a venture with funds managed by Blue Owl Capital to build its Hyperion campus in Louisiana. Blue Owl's funds own 80% and Meta 20%; the venture raised US$27.3 billion of bonds through Beignet Investor LLC, which S&P Global Ratings rated A+ (S&P has published its reasoning only in a webinar). Meta doesn't consolidate the venture because, it says, it doesn't direct the activities that most affect the venture's results; its auditor flagged that judgment as a critical audit matter while giving an unqualified opinion. Meta will lease the buildings (from 2029, initial four-year terms renewable to 20 years) and has guaranteed part of their future value if it walks away. In July 2026 it set up a similar venture in Texas, and the same month announced an expansion of Hyperion from about 2 GW to 5 GW, at over US$50 billion.

This is the most Meta could lose, not debt it owes or a loss it expects: it combines an investment already made, lease payments, funding promises and guarantees that pay only if Meta ends a lease and the buildings are worth less than a set threshold. Meta says those guarantee payments aren't probable, so it records no liability for them. The rise from December largely reflects adding the Texas venture, not a deterioration in Louisiana. The jump in Hyperion's cost, from an early estimate of US$10 billion to US$27 billion to over US$50 billion, mainly reflects a larger project (about US$13 billion per gigawatt in 2025, at least US$10 billion per gigawatt for the 5 GW plan), not cost overruns. Arrangements like this deserve scrutiny, and they move risk around rather than removing it; their existence alone doesn't show misconduct or prove a bubble. Sources: Meta 10-Q, Sep 2026 (Note 5, p. 17–18) · Meta 10-K, 2025 (Note 5 and auditor's report) · Meta announcement, Oct 2025 · S&P Global Ratings: "Understanding our A+ rating" (webinar; registration may be required) · CNBC, Jul 2026.

What would signal stress

• Capital spending staying above operating cash flow for several quarters, funded by new debt or shares.
• Credit-rating downgrades, or the cost of borrowing rising faster than for other companies with similar ratings.
• Impairments (write-downs) of data centres or servers, or useful lives of servers being shortened.
• Documented cancellations, delays or downsizing of data-centre projects and leases.
• Guarantee payments becoming probable, so that a liability is recorded.

Corporate credit spreads

The extra yield investors demand to lend to US companies rather than the government, for investment-grade and high-yield (junk) bonds. Rising spreads would mean lenders are pricing in more risk. FRED publishes only about three years of these series (a licensing limit on the ICE BofA indices, not the length of their full history), so there is no long history to compare with here.

Against the dot-com bubble

The Nasdaq Composite around its 10 March 2000 peak, set against the AI basket around its highest level so far. Both are rebased to 100 at that point, with days measured from it. The basket's high may not be its final peak, and because the basket is today's winners, its run-up before the high is overstated relative to the Nasdaq's.

What drives the gauge

Five signals are scored 0–100 as a percentile of their own history up to that day (never using later data): a score of 90 means the reading is higher than on 90% of earlier days. Where a longer public record exists, that is the history used: the 10-year yield and the VIX since 1990, and capital-spending intensity since 2010. The rest have only the basket's own history (the "History from" column). Bubble signature is not a percentile: it is the share of the model's fits that meet the bubble criteria. Bubble Pressure is the simple average of these six scores. Relative strength is shown but not counted, because a basket chosen in 2026 beats the Nasdaq partly by construction. The ticks on each bar mark 50 and 70, where the gauge turns amber and red. Equal weights are a design choice and the combination hasn't been validated as a predictor.

Other lenses

Readings that help interpret the score but are not part of it, because there is no tested way to weight them in yet.

How the burst forecast works

A burst here means the basket closing at least 50% below its highest level so far. The basket has never done that: its worst fall since 2017 was 41%, in 2022. So the estimate borrows two long records that contain the two great modern bubbles: the Nasdaq Composite since 1971 (dot-com) and Japan's Nikkei 225 since 1949 (1990). Each week of those records is compared with today on three things: how far the index was below its high, whether that high was set within the past year, and how much it had risen in the three years before it (at least 150%, at least 100%, or any, whichever tier the basket is in now). Seventeen concentrated stocks swing more than a broad index, so the comparison is scaled by volatility. Matching weeks less than six months apart count as one episode, each episode counts once, and the probability is the share of episodes followed by a burst within 12 or 24 months. A standard small-sample adjustment (a Jeffreys prior) keeps a handful of events from reading as 0% or 100%, and the 80% range comes from the same calculation.

Past episodes like today

"Burst within" is the share of the episode's matching weeks that were followed by a burst within that time. Falls and rises are in basket terms, after the volatility scaling.

How much the answer depends on the choices

The same calculation with a different run-up tier, and with each index on its own.

Readings the daily judgement watches

Only the basket's own fall feeds the probability (it decides which episodes match). These are context for the judgement, not inputs.

Scorecard

Each day's probabilities go into the forecast log and are scored when their 12 or 24 months have passed.

What it can't tell you

• Two bursts are a thin record. The real uncertainty is wider than the 80% range, which only reflects how few episodes there are, not whether these two indices are the right comparison.
• It moves in steps: as the basket's fall deepens or recovers, different past episodes match, and the number can jump.
• It knows nothing about what is different this time: the capital spending, how it is funded, the companies' profits. Those are in the sections below and in the daily judgement.
• It isn't investment advice.

How the model works, and how it has done

The dated windows come from the Log-Periodic Power Law Singularity (LPPLS) model developed by Didier Sornette and colleagues. The idea is that in a bubble, investors pile in after each other, so prices rise at an accelerating rate with oscillations that speed up. The model fits that shape and estimates a critical time, when the run-up is most likely to end. A falling market can show the mirror-image pattern, a "negative bubble", whose critical time marks a possible rebound.

This implementation is experimental. Each run fits the model to 26 windows ending on the same day, from 125 to 750 trading days long (about 6 months to 3 years). The page reports how many of those fits meet standard qualification criteria (exponent m between 0.01 and 0.99, angular frequency ω between 2 and 15, at least 2.5 oscillations, damping of at least 0.8) and, when some do, the 10th to 90th percentile spread of their critical times. That count is often called the LPPLS "confidence indicator", but it is a share of fits, not a calibrated probability of a crash or a rebound. The spread of critical times is not a statistical confidence interval either, and it doesn't identify the period of highest risk. Fitting details: the nonlinear parameters are searched with Nelder–Mead from 24 random starting points (the best 4 refined), critical times are limited to within half the window length ahead, fits that fail are excluded from the count, and weekly history uses each week's last trading day. Research on the method: Filimonov & Sornette (2013), Sornette et al. (2015), Zhang, Zhang & Sornette (2016), and on the instability of critical-time estimates, Filimonov, Demos & Sornette (2016).

Backtest on the Nasdaq Composite, re-run on 1 Nov 2026 with data cut off on each date shown. The rule, set before scoring: a bubble call is a hit if the Nasdaq's highest close in the next 12 months falls inside the window and is followed by a fall of 20% or more within a year. A bubble call followed by such a fall whose peak is outside the window is a timing miss; a bubble call with no fall of 20% or more within 12 months is a false alarm. A no-signal call is correct if there is no peak-to-trough fall of 20% or more in the next 12 months, and a missed fall otherwise.

Five dates, chosen by hand to cover known peaks and calm periods, and two of them concern the same event. That is far too small to show the model is reliable; the walk-forward test below scores every week since 1980 instead, counting misses and false alarms. An earlier version of this table called the 2025 row "No false alarm"; under a stated rule it is a missed fall (corrections).

Walk-forward test on the Nasdaq Composite, every week since 1980

The five dates above were chosen by hand. This test fits the model every week using only the prices up to that week, on an index nobody chose, and scores every week with the same rule. Weeks overlap (a single bubble produces many calls in a row), so results are also counted by episode, and compared with how often a 20% fall happened anyway.

Shaded: weeks followed by a Nasdaq fall of 20% or more within 12 months. A reading above the dashed line is a call.

Walk-forward rerun on the AI basket

These weekly runs were computed retrospectively, in one batch, with each run using only prices up to its own date. The first came more than two years before the high, far too early to act on; the chart below also shows the same model on the Nasdaq Composite, which has no hand-picked companies. They are not forecasts that were published at the time, and the index they use is built from today's basket and share counts. Live, dated forecasts start with the forecast log.

Bubble signature over time

Forecast log

Each day's forecast is logged as it is published and never rewritten. SpaceX joined the basket on 1 Nov 2026, so earlier entries are from the 16-company basket. Entries from 1 Nov 2026 also record the model version, the time they were logged, the companies in the basket, and a hash chained to the previous entry: anyone who saves a copy of forecast-log.json can recompute the hashes (SHA-256 of each entry without its "hash" field, keys sorted, no spaces) to check that no earlier entry has changed.

* Logged before 1 Nov 2026 with 16 companies and the earlier index method, so its drawdown and Bubble Pressure differ from what the same day reads now. Entries are never rewritten; see corrections.

Data and method

• The basket: Nvidia, Microsoft, Alphabet, Amazon, Meta, Apple, TSMC, Broadcom, Oracle, ASML, AMD, Intel, Palantir, Arm, Super Micro, CoreWeave and SpaceX (owner of xAI and X). What each one does.
• How they were chosen: listed companies whose value depends substantially on AI, as a model builder, a chip or equipment maker, an AI cloud or server supplier, or one of the largest spenders on AI data centres. The list was chosen in 2026, so it favours companies that are prominent now (hindsight and survivorship bias). Private labs (OpenAI, Anthropic) have no share price. Companies are counted at full market value even where most of the business isn't AI (Apple, Amazon, Alphabet, SpaceX).
• Prices: daily closing prices from Nasdaq, adjusted for splits but not dividends, so returns are price-only. Nasdaq's price history adds a session about six hours after the close, so the latest session's official close is taken from Nasdaq's quote within an hour of the close and replaced by the history on a later run. TSMC and ASML are their US-listed shares; Alphabet is its Class A shares.
• The index: cap-weighted and chained daily: each day's change is the market-value-weighted average of the companies' price changes, so a company that lists later (Arm in 2023, CoreWeave in 2025, SpaceX in 2026) joins without causing a jump. The equal-weighted index averages the companies' daily price changes, which amounts to rebalancing daily.
• Market value: price times shares. Today's share count is implied by Nasdaq's market value, and scaled back through time using the share counts on SEC filing cover pages, dated to the period they cover. Counts are adjusted for share splits. Multi-class companies don't report a single cover-page count, so for them the quarterly diluted share count is used instead.
• Revenue and capex: from SEC EDGAR XBRL filings (10-K, 10-Q, 20-F). Revenue includes non-AI business and sales between basket companies, so it isn't a measure of final demand for AI. Capex is cash spent on property, plant and equipment from the cash-flow statement; it excludes equipment acquired under finance leases, which is large for some cloud companies. Each figure is used from the date it was filed, and the first filing of each period is kept, so later restatements aren't applied. Trailing-12-month figures are summed from quarters. Only annual figures are machine-readable for TSMC and ASML, though both report quarterly. TSMC's recent revenue is summed from its monthly revenue reports (Form 6-K); ASML's uses its latest annual report, and is dropped from price/sales once more than 15 months old. Foreign-currency figures are converted at the exchange rate on the period's last day (FRED), an approximation for flows earned across the period.
• Funding: operating cash flow from the same SEC filings (trailing 12 months, summed from quarters), compared with cash capex. The commitments table is entered by hand from each filing's notes and dated; the page flags it when a newer filing appears.
• Macro: Nasdaq Composite, VIX, the 10-year Treasury yield, credit spreads and exchange rates from FRED (Federal Reserve Bank of St. Louis). FRED publishes about a day after the market, so the latest session's Nasdaq Composite, VIX and 10-year yield come from Nasdaq, Cboe and the US Treasury (the sources FRED itself uses) until FRED catches up. The ICE credit spreads are only published a day late. Each series' latest observation date, and where a same-day value came from, is in the table below.
• Commentary is written daily by an AI model (Anthropic's Claude). It sees only the figures on this page, formatted as the page shows them, is told to give both the bull and bear case and to say what the model can't tell you, and is labelled with its date, model, data snapshot and prompt version. Nobody edits it.

All data sources, with their latest observation dates

Download the data behind this page: site.json (prices, index, signals, model history, sources), forecast-log.json, commentary.json.

Corrections and editorial responsibility

Edited by Brendan Faulds, who is responsible for this page. Spot an error? Get in touch. Corrections are listed here and earlier forecasts are never rewritten.

• 6 Nov 2026, morning share prices: the 5 Nov update said the morning run uses each share's official close. It doesn't: Nasdaq's quote gives the last trade at 4pm New York time, and the official closing price, set in the closing auction, can differ. On 5 Nov they were up to 0.4% apart (Nvidia 182.29 against 181.60), so the forecast log's entry for 5 Nov reads 15.43% below the high and Bubble Pressure 50.72, where the official closes give 15.6% and 50.70. That entry stays as logged. Morning prices are now labelled provisional, a day's forecast is logged only once every official close is in (normally the 4:45pm Sydney run), and the commentary is rewritten then.
• 5 Nov 2026, update schedule: the page updated once a day, Tuesday to Saturday, in the Sydney afternoon. By then its newest prices were already a day old, and the Nasdaq Composite, VIX and 10-year yield (from FRED) were often a day older again. The commentary and judgement were weekly. It now updates every day at 9am and 4pm Sydney time. The 9am run takes the previous US session's official closes from Nasdaq, Cboe and the US Treasury. The commentary is rewritten for each new session and the judgement every day. The page now says so if it hasn't refreshed for more than 24 hours. (On 6 Nov the runs moved to 8:45am and 4:45pm, once it was measured that every same-day source has its close by 8:31am.)
• 1 Nov 2026, scoring change (after a third review): five of the seven signals were scored only against the basket's own history since October 2017, so anything at a decade high scored near 100. The 10-year yield of 4.98% scored 99; against the record since 1990 it scores 67. Rates and the VIX are now scored against their records since 1990, and capital-spending intensity against the five spenders' filings since 2010 (it still scores 100: the current level is a record). Relative strength is no longer counted, because a basket chosen in 2026 beats the Nasdaq partly by construction. Bubble Pressure for 30 Oct 2026 is 54 under the new method; it read 63 under the old one, and the forecast log keeps 63 for that day. Each new log entry records the scoring method. Also: the funding table's "left over" column is now worked out from the figures as shown, so every row adds up as printed (Oracle's showed −US$28.7B against −US$28.8B from its own columns).
• 1 Nov 2026, index error (found while checking a second external review): the basket index was chained on day-to-day changes in market value, and SEC share counts weren't adjusted for share splits. Amazon's 20:1 split and Nvidia's 10:1 split therefore appeared as one-day jumps of +17% (20 Jul 2022) and +21% (23 Aug 2024), and historic market values were far too low (Nvidia's end-2020 value showed as US$8 billion instead of about US$320 billion). The index now chains on price changes weighted by market value, as the method always said, and share counts are split-adjusted. The basket's rise since ChatGPT's launch was overstated: +243% before the fix, +222% after. Bubble Pressure, the price-to-sales history and the bubble model's history were recomputed. The forecast logged for 30 Oct 2026 is kept as published; it used the old index and the 16-company basket (−10.9% from the high, against −11.7% now).
• 1 Nov 2026 (second external review): the forecast log now records each entry's basket and a hash chain; the commentary is checked figure by figure against its data, and may no longer present the retrospective weekly runs as a track record; the model's history is tested every week since 1980 on the Nasdaq Composite; Alphabet's, Meta's, Palantir's and CoreWeave's market values use diluted share counts instead of holding today's count constant; the S&P A+ rating on Meta's Hyperion bonds is now cited to S&P.
• 1 Nov 2026 (after an external review): company profiles corrected and given source links (Nvidia was called the largest company in the basket, but Apple was larger; Microsoft's OpenAI stake, Amazon's Anthropic investment, Arm's own chip, CoreWeave's services, Oracle's funding, Apple's Gemini deal, Nvidia's OpenAI investment). The weighting caption described the gap between the two indices the wrong way round. "Relative strength" was described as a ratio of gains; it is a ratio of index levels. The signal description said all seven signals are percentiles; one isn't. The backtest's 2 Jan 2025 row said "No false alarm"; it is a missed 24% fall. ASML was described as reporting annually; it reports quarterly. The weekly reruns were labelled "out-of-sample"; they were computed retrospectively.
• 1 Nov 2026: SpaceX (which owns xAI) added to the basket.

Not financial advice. This is an independent, automated analysis for general information. It doesn't know your circumstances, it will sometimes be wrong, and nothing here is a recommendation to buy or sell anything. Talk to a licensed adviser before making investment decisions.