AI to crash or not to crash?
PDF article with references:
The technology can change the world while its shares still fall
Everyone seems to have an opinion. Wall Street executives discuss an AI correction, and so do taxi drivers in Tel Aviv. That alone tells us very little about timing. It does tell us that AI is no longer a niche technology trade. It is a story about stocks, semiconductors, data centres, electricity, construction, corporate debt and, the wider economy.
So here is my honest answer up front: a further sharp correction in AI-related assets is entirely plausible, but the evidence available today does not confirm that a systemic market crash is imminent. Valuation and concentration make the market vulnerable. Strong earnings and still-contained credit spreads argue against a simple countdown. The most important thing to watch is whether disappointment travels from share prices into quality of earnings, capital expenditure, credit and forced selling.
That distinction matters because an AI investment bust and an AI technology bust are not the same thing. Railway mania did not make railways useless. The dot-com crash did not switch off the internet. A fall in asset prices can be brutal while the underlying technology continues to spread.
What past crashes actually tell us
There is no single crash indicator. Different episodes have different transmission mechanisms, although the same ingredients keep appearing: expensive assets, a compelling story, crowded positioning, leverage, weak liquidity and a trigger that forces everyone to reassess at once.
Episode | What broke | Main lesson for today |
|---|---|---|
1929 and the Great Depression | Valuations, leverage and the financial system, followed by a severe economic contraction | A market fall becomes far more dangerous when banking and policy failures reinforce it |
Dot-com bust, 2000 to 2002 | Valuations and the gap between promised and realised profits, alongside excessive communications investment | A genuine technological revolution can still produce terrible investment returns from the wrong starting price |
Global financial crisis, 2007 to 2009 | Housing credit, bank leverage, funding liquidity and confidence | Credit plumbing matters more than the headline stock index when losses become systemic |
Rate shock, 2022 | Valuation multiples as interest rates rose, without a comparable banking collapse | Expensive growth assets can fall sharply even when the financial system keeps functioning |
1929 and the Great Depression | Valuations, leverage and the financial system, followed by a severe economic contraction | A market fall becomes far more dangerous when banking and policy failures reinforce it |
The scale of these events should prevent glib reassurance. The Dow fell about 89% in the Great Depression, the Nasdaq about 78% in the dot-com bust, and the S&P 500 about 57% in the housing crisis.[1] The Nasdaq Composite’s exact closing-price fall from 10 March 2000 to 9 October 2002 was 77.9%. It did not close above its old peak again until 23 April 2015, more than 15 years later. That is a nominal price recovery, before inflation and without dividends.[2]
The Nasdaq took 15 years to regain its dot-com peak
This is why I would not tell a client simply to relax because “markets move in waves”. They do, but waves are not a promise of prompt recovery. Diversification can reduce the chance of permanent loss. It cannot guarantee a good outcome over every investor’s time horizon, and individual companies can fail even while an industry thrives.
S-curves and log-normal models are useful mental pictures, but they need similar care. An adoption S-curve describes how a technology spreads. It does not tell us what price investors should pay for it. Likewise, a tidy return distribution can badly understate fat tails, volatility clustering and the human tendency to sell at the worst moment. Use these models as maps, not laws.
The current dashboard: amber, not a uniform red
The confusing feature of today’s market is that different valuation measures are telling different stories.
The Shiller CAPE ratio stood at 40.13 on 30 October 2026. That is above September 1929’s 32.56 and November 2021’s 38.58, though below December 1999’s 44.19.[3] Yet FactSet put the S&P 500’s forward 12-month price-to-earnings ratio at 18.7 on 23 October 2026, below its five-year average of 19.8 and ten-year average of 19.1. Only 24% of companies had reported at that point, so the associated earnings-season figures were still preliminary.[4]
Valuation is stretched, but it is not a countdown clock
This is not a contradiction that needs to be argued away. CAPE compares prices with inflation-adjusted earnings over the previous decade. A forward P/E uses expected earnings for the next year. The gap between them places a heavy burden on those expectations. If earnings arrive, today’s forward multiples can look reasonable. If they do not, the apparently cheaper measure can rise even as prices fall.
Here is the broader dashboard.
Indicator | Current reading | What I take from it |
|---|---|---|
Long-term valuation | CAPE of 40.13 on 30 October 2026[5] | Expensive and vulnerable, but useless as a precise timing tool |
Near-term valuation | S&P 500 forward P/E of 18.7 on 23 October, close to recent averages[6] | Earnings upgrades have relieved some valuation pressure |
Earnings | Preliminary blended Q3 earnings growth of 29.5% and revenue growth of 12.7%; 83% of reporting companies beat EPS estimates, but the aggregate surprise was only 2.1%[7] | Strong results, with less room for disappointment than the headline beat rate suggests |
Concentration | The ten largest companies were more than 36% of the S&P 500 in March 2026, against 23% in 2000 and 41% in 2025[8] | Owning the index is not the same as having ten independent sources of return |
Earnings breadth | Micron and Nvidia accounted for 36% of expected S&P 500 earnings growth in Citadel Securities’ October analysis; the ten largest contributors supplied roughly two-thirds[9] | Aggregate growth can conceal considerable dependence on a small group |
Capital expenditure financing | The Bank of England cited estimates of about $450 billion of global AI-related debt issuance by early September 2026, and $4.1 trillion of debt-financed AI capital expenditure during 2026 to 2030[10] | The cycle is becoming more credit-sensitive; these are third-party estimates, not realised totals |
Credit | The US high-yield option-adjusted spread was 3.26 percentage points on 29 October 2026[11] | Credit was not confirming systemic stress, although spreads can move quickly |
Banks and leverage | The Federal Reserve described banks as broadly sound and business and household debt vulnerabilities as moderate, while hedge-fund leverage was near all-time highs[12] | A market accident can be amplified without requiring a weak banking system |
Rates and liquidity | No current yield-curve inversion, but tighter financial conditions remain a risk |
Two points in that table deserve a closer look.
First, concentration cuts both ways. It makes the index dependent on a small number of companies, but those companies can also produce unusually strong earnings. Citadel Securities reported that, since 30 June, AI-infrastructure forward EPS estimates had risen 25% while forward P/E ratios had fallen 22%, with the group’s forward P/E moving from 21.8 to 17.0.[15] In plain English, part of the market has already become cheaper because expected profits grew faster than prices. That is not what a pure, earnings-free bubble looks like. The catch is that excluding the ten largest contributors reduced expected S&P 500 earnings growth to roughly 11% in the same analysis.[16]
A broad index can still carry concentrated AI exposure
Second, the financing mix is changing. Early in the AI boom, the strongest hyperscalers could fund much of their investment from operating cash flow. That remains an important difference from many dot-com era companies. But the scale is becoming extraordinary. Meta, for example, reported 2025 capital expenditure of $72.22 billion and guided to $115 billion to $135 billion for 2026, including finance-lease principal payments.[17] The Bank of England also cited a Morgan Stanley estimate that about $700 billion of data-centre capital expenditure during 2026 to 2028 could be financed by private credit.[18]
Debt does not make an investment bad. It does make the consequences of a forecasting error less forgiving. A data centre financed largely from corporate cash flow can produce disappointing shareholders. A highly leveraged one can also produce refinancing problems, creditor losses and forced asset sales.
Second, the financing mix is changing. Early in the AI boom, the strongest hyperscalers could fund much of their investment from operating cash flow. That remains an important difference from many dot-com era companies. But the scale is becoming extraordinary. Meta, for example, reported 2025 capital expenditure of $72.22 billion and guided to $115 billion to $135 billion for 2026, including finance-lease principal payments.[19] The Bank of England also cited a Morgan Stanley estimate that about $700 billion of data-centre capital expenditure during 2026 to 2028 could be financed by private credit.[20]
Debt does not make an investment bad. It does make the consequences of a forecasting error less forgiving. A data centre financed largely from corporate cash flow can produce disappointing shareholders. A highly leveraged one can also produce refinancing problems, creditor losses and forced asset sales.
Why this bust would reach beyond the stock market
The user’s instinct here is right: an AI downturn could have a broader physical footprint than the dot-com bust. The AI build-out is not only software and advertising. It is chips, servers, networking, land, buildings, grid connections, generation equipment and power.
The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global consumption. Its base case projects about 945 terawatt-hours by 2030, just under 3% of global electricity use.[21] A serious investment reversal would therefore be felt by suppliers, contractors, utilities and local development plans, not only by holders of technology shares.
There is still a difference between a painful investment cycle and an economic depression. An IMF downside scenario reported by Reuters estimated that a steep AI-investment decline could reduce global GDP by 1.1% by 2028 and US output by 1.8%, while IMF officials said this would still be unlikely to constitute a global recession.[22] That is a scenario, not a forecast, but it usefully separates a large macroeconomic shock from an outright collapse.
We have already had a small stress test
In July 2026, AI and semiconductor shares fell sharply. The Bank of England said leveraged position unwinding amplified the move, but there was no spillover to core markets. It also warned that a sharper correction remains possible if earnings expectations weaken.[23]
The broad Nasdaq Composite fell about 9.8% from its 2 June close to its 29 July close. By 29 October it was approximately 5.1% below its 6 October closing high.[24] This is useful evidence in both directions. Crowding and leverage can accelerate a sell-off, yet a sharp sector decline does not automatically become a systemic crisis.
So, is a crash coming soon?
My answer is: a correction is plausible; a date is not knowable; a systemic crash is not the conclusion supported by the current dashboard.
A crash predicted for six months from now can arrive in two years, after the forecaster has been squeezed out of the position. Or the market can fall before the prediction becomes fashionable. Valuation tells us about vulnerability and prospective long-term returns, not next Tuesday’s direction.
The IMF’s October 2026 Global Financial Stability Report captured the current tension well: markets had been resilient, but risks were elevated, and weaker or less durable AI investment and profitability could produce a correction amplified by leverage and concentrated exposures.[25] The IMF’s latest World Economic Outlook still projected global growth of 3.1% in 2026 and 3.4% in 2027, while listing a reassessment of AI expectations among the downside risks.[26]
Rather than asking “bubble or no bubble?”, I would monitor three scenarios.
Scenario | What would confirm it | Likely character |
1. Valuation reset | Multiples fall while earnings estimates and capital-expenditure plans remain broadly intact; credit spreads stay contained | Painful for crowded shares, but closer to a normal correction than a systemic bust |
2. AI earnings and capital-expenditure bust | Monetisation disappoints, orders are delayed or cancelled, earnings estimates fall, depreciation and write-downs rise, and financing becomes more expensive | A deeper sector downturn with real effects on chips, construction, power and suppliers |
3. Systemic liquidity or credit accident | Falling asset values meet hedge-fund deleveraging, private-credit stress, funding problems and forced sales across markets | The least predictable and most dangerous case; not confirmed by current bank and credit indicators |
The sequence matters more than any single red light. The warning chain I would watch is: weaker narrative, weaker earnings revisions, weaker capital expenditure, wider credit spreads, then forced selling. If the chain stops at the first link, investors may get a valuation reset. If it reaches the last links, the problem becomes much bigger.
A hedging policy for people who cannot see the future
The purpose of hedging is not to look clever on the day of a crash. It is to avoid being forced into bad decisions and to preserve the ability to act when prices become attractive.
1. Map the exposure you actually have
Look through the labels. A portfolio can hold an S&P 500 fund, a semiconductor fund, a utility building data-centre capacity, a private-credit vehicle financing infrastructure and a few technology shares, while still being one large AI trade.
Map direct holdings, index exposure, suppliers, customers, debt-financed infrastructure and employment income. A founder whose business also depends on AI spending has a different risk budget from a retired investor with no other exposure.
2. Hold liquidity before you need it
A fall can create an excellent entry point, but only investors with available liquidity can exploit it. Keep spending reserves and strategic “dry powder” separate. The first prevents forced selling. The second allows deliberate buying.
Cash and high-quality short-duration government instruments are the simplest candidates, subject to currency, tax, inflation and counterparty considerations. Do not assume long-duration bonds will hedge every shock. They can fall with equities when inflation or sovereign yields are the problem.
3. Rebalance before sentiment makes the decision for you
Set exposure limits and rebalancing bands while markets are calm. If AI-linked holdings have grown beyond the approved allocation, trim them according to a rule rather than a mood. Rebalancing will not maximise returns in a straight-line boom. That is precisely why it is useful when the boom ends.
4. Diversify across return drivers, not ticker symbols
Ten AI companies are not ten independent bets. Neither is a capitalisation-weighted index with a large top-ten weight. Genuine diversification may include different regions, sectors, valuation exposures and asset classes, but each addition should be tested for hidden dependence on the same capital-expenditure cycle.
5. Pre-commit to staged buying
If a decline is the entry opportunity you have been waiting for, decide in advance how capital could be deployed in tranches. Use time, valuation or market levels as possible rules, while leaving room to reassess if the earnings story has genuinely broken. A mechanical buying plan should never become an obligation to catch a falling knife.
6. Use options only with a budget and a defined job
Protective puts can insure a specified exposure for a specified period. Put spreads can reduce the premium while limiting protection. Collars can fund some protection by giving up upside. None is free once premiums, rolling, volatility pricing, tax and imperfect index matching are included. Treat options as insurance contracts, not as a source of magical safety.
7. Treat gold as a diversifier, not a promise
Gold may help in some currency, inflation or confidence shocks. It can also fall during a liquidity scramble, produces no cash flow and can be expensive after strong demand. A controlled allocation may diversify risk; it does not provide a guaranteed crash hedge.
8. Avoid leverage and be cautious with shorting
Leverage turns a temporary loss into a potential forced sale. Shorting adds timing risk, borrowing costs, crowding and theoretically unlimited losses. It can be appropriate for specialist mandates with strict controls, but it is a poor default hedge for most long-term investors. Being right about an overvaluation in 2028 does not help if the position becomes unmanageable in 2026.
The bottom line
AI can be as transformative as electricity or the internet and still produce a terrible period for investors. The technology’s usefulness does not establish a floor under today’s valuations.
At the same time, a high CAPE ratio is not an air-raid siren. Current earnings are strong, forward valuation measures are less extreme, credit spreads remain contained and banks appear resilient. The strongest reasons for caution are concentration, dependence on expected earnings, the sheer scale of infrastructure spending and the growing role of debt and private credit.
So, I would not build a policy around a six-month crash prediction. I would build one around resilience: know the exposure, avoid forced selling, keep liquidity, diversify the actual drivers of return and write the buying rules before the opportunity arrives.
That approach will not identify the top. It does something more useful. It allows you to survive being wrong about the timing.
Data and source dates run through 30 October 2026 unless stated otherwise. This article is general market analysis, not personalised investment advice. Options, short selling and leveraged products can involve substantial losses. Tax and implementation rules vary by investor and jurisdiction.
[1] Fidelity, “Is AI a bubble? 5 signs to watch for,” including historical peak-to-trough comparisons. https://www.fidelity.com/learning-center/trading-investing/ai-bubble
[2] Federal Reserve Bank of St. Louis, FRED, NASDAQ Composite, series NASDAQCOM. Calculations use daily closing values. Peak: 5,048.62 on 10 March 2000; trough: 1,114.11 on 9 October 2002; first subsequent close above the old peak: 5,056.06 on 23 April 2015. The 2021 to 2022 decline was 36.4%, and the 2 June to 29 July 2026 decline was 9.8%. https://fred.stlouisfed.org/series/NASDAQCOM
[3] Multpl, “Shiller PE Ratio by Month,” observations through 30 October 2026. https://www.multpl.com/shiller-pe/table/by-month
[4] FactSet, “S&P 500 Earnings Season Update: October 23, 2026.” Earnings figures were preliminary because only 24% of companies had reported. https://insight.factset.com/sp-500-earnings-season-update-october-23-2026
[5] Multpl, “Shiller PE Ratio by Month,” observations through 30 October 2026. https://www.multpl.com/shiller-pe/table/by-month
[6] FactSet, “S&P 500 Earnings Season Update: October 23, 2026.” Earnings figures were preliminary because only 24% of companies had reported. https://insight.factset.com/sp-500-earnings-season-update-october-23-2026
[7] FactSet, “S&P 500 Earnings Season Update: October 23, 2026.” Earnings figures were preliminary because only 24% of companies had reported. https://insight.factset.com/sp-500-earnings-season-update-october-23-2026
[8] Visual Capitalist, “The Entire S&P 500 in 2026 in One Chart,” using Slickcharts data as of 30 March 2026. https://www.visualcapitalist.com/the-entire-sp-500-in-2026-in-one-chart/
[9] Citadel Securities, “October 2026 Global Market Intelligence.” These figures describe the firm’s analysis and classification of AI-infrastructure companies. https://www.citadelsecurities.com/news-and-insights/global-market-intelligence/october-update/
[10] Bank of England, Financial Policy Committee Record, September 2026. The debt and capital-expenditure figures are estimates by Morgan Stanley and JPMorgan cited by the Bank, not audited totals. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026
[11] Federal Reserve Bank of St. Louis, FRED, ICE BofA US High Yield Index Option-Adjusted Spread, series BAMLH0A0HYM2. Latest value used: 3.26 percentage points on 29 October 2026. https://fred.stlouisfed.org/series/BAMLH0A0HYM2
[12] Board of Governors of the Federal Reserve System, Financial Stability Report, May 2026, overview and leverage sections. Data conditions were generally as of 23 April 2026. https://www.federalreserve.gov/publications/2026-may-financial-stability-report-overview.htm
[13] Bank of England, Financial Policy Committee Record, September 2026. The debt and capital-expenditure figures are estimates by Morgan Stanley and JPMorgan cited by the Bank, not audited totals. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026
[14] Federal Reserve Bank of St. Louis, FRED, 10-Year Treasury Constant Maturity Minus 2-Year Treasury Constant Maturity, series T10Y2Y. https://fred.stlouisfed.org/series/T10Y2Y
[15] Citadel Securities, “October 2026 Global Market Intelligence.” These figures describe the firm’s analysis and classification of AI-infrastructure companies. https://www.citadelsecurities.com/news-and-insights/global-market-intelligence/october-update/
[16] Citadel Securities, “October 2026 Global Market Intelligence.” These figures describe the firm’s analysis and classification of AI-infrastructure companies. https://www.citadelsecurities.com/news-and-insights/global-market-intelligence/october-update/
[17] Meta Platforms, “Meta Reports Fourth Quarter and Full Year 2025 Results,” 28 January 2026. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Fourth-Quarter-and-Full-Year-2025-Results/default.aspx
[18] Bank of England, Financial Policy Committee Record, September 2026. The debt and capital-expenditure figures are estimates by Morgan Stanley and JPMorgan cited by the Bank, not audited totals. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026
[19] Meta Platforms, “Meta Reports Fourth Quarter and Full Year 2025 Results,” 28 January 2026. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Fourth-Quarter-and-Full-Year-2025-Results/default.aspx
[20] Bank of England, Financial Policy Committee Record, September 2026. The debt and capital-expenditure figures are estimates by Morgan Stanley and JPMorgan cited by the Bank, not audited totals. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026
[21] International Energy Agency, “Energy and AI,” section on energy demand from AI. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
[22] Reuters, republished by CNBC Africa, “IMF flags risks to global economy from AI market correction, war and debt,” 13 October 2026. https://www.cnbcafrica.com/2026/imf-flags-risks-to-global-economy-from-ai-market-correction-war-and-debt
[23] Bank of England, Financial Policy Committee Record, September 2026. The debt and capital-expenditure figures are estimates by Morgan Stanley and JPMorgan cited by the Bank, not audited totals. https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026
[24] Federal Reserve Bank of St. Louis, FRED, NASDAQ Composite, series NASDAQCOM. Calculations use daily closing values. Peak: 5,048.62 on 10 March 2000; trough: 1,114.11 on 9 October 2002; first subsequent close above the old peak: 5,056.06 on 23 April 2015. The 2021 to 2022 decline was 36.4%, and the 2 June to 29 July 2026 decline was 9.8%. https://fred.stlouisfed.org/series/NASDAQCOM
[25] International Monetary Fund, Global Financial Stability Report, October 2026. https://www.imf.org/en/publications/gfsr/issues/2026/10/13/global-financial-stability-report-october-2026
[26] International Monetary Fund, World Economic Outlook, October 2026. https://www.imf.org/en/publications/weo/issues/2026/10/13/world-economic-outlook-october-2026
Comments