When Will the AI Bubble Burst? 5 Signals That Move Before the Market

The question everyone asks has a better version#

Ask when the AI bubble will burst and you are really asking two things. The first, what date, has no answer. Anyone who offers one is guessing. The second question is more useful: what would we see first?

Bubbles rarely end on a single dramatic morning. They end the way a bridge fails, with hairline cracks in unglamorous places long before the span comes down. In markets those places are credit spreads, financing structures, cash flow statements and adoption surveys. They move ahead of share prices, and unlike share prices they are published on a schedule anyone can follow.

There is a reason to look now. The Bank of England's July 2026 Financial Stability Report records that AI companies account for around half of the S&P 500's market value, up from roughly a quarter at the end of 2022. On 28 August 2026, after Fed Chair Kevin Warsh's Jackson Hole speech, markets put a 57% chance on a September rate rise, with the 10-year Treasury yield at 4.72%. A boom financed increasingly by borrowing has met a rate path that is no longer falling.

What a bubble is, and what has actually been built#

A bubble is not simply something expensive. It is a price that depends on future earnings the underlying business is unlikely to deliver. The test is arithmetic, not mood.

The IMF ran that arithmetic in its April 2026 Global Financial Stability Report. To justify prices at the time, the S&P 500 would have needed roughly 30% annual earnings growth through 2027, and the Nasdaq about 35%. Both figures sit well above what analysts themselves were forecasting. That gap is the bubble question in one line.

Four terms carry most of the weight in what follows.

Capital expenditure, or capex, is money spent on long-lived physical assets: here, data centres, chips and power connections. Free cash flow is what a company has left after that spending; when capex exceeds it, the shortfall must come from borrowing or issuing shares. A credit default swap, or CDS, is insurance on a company's debt, and its price, quoted in basis points, is the market's live estimate of default risk. Depreciation spreads an asset's cost across its useful life. Assume a chip lasts six years rather than three and today's reported profits look better.

None of the technology is imaginary. Nvidia's quarter ending 26 July 2026 produced $96.2 billion of revenue, up 106% on the year, with $89.0 billion of that from data centres, according to its results filed with the SEC. Real chips are being sold and real buildings are going up. The argument is about whether the revenue arriving at the other end will ever justify the cost.

The five signals that move before the market does#

1. The financing mix: cash flow gives way to debt#

For most of the boom, the big cloud companies paid for data centres out of operating cash flow. That has changed. FactSet's analysis of Alphabet, Amazon, Meta, Microsoft and Oracle found incremental debt rose from 9% of capex in fiscal 2024 to 32% over the twelve months to mid-2026. Aggregate capex is expected to exceed $690 billion in FY26, while free cash flow heads towards zero or turns negative for all except Alphabet and Microsoft. In June 2026, Alphabet raised $84.75 billion in equity.

This signal has the longest lead time, because it appears in quarterly filings and is hard to disguise. A boom paid for out of profits can slow gracefully. A boom paid for by capital markets stops when capital markets stop.

2. Credit spreads: the bond market votes first#

Equity holders own the upside. Lenders own only the downside, so they tend to notice trouble sooner. The Bank for International Settlements, in its March 2026 Quarterly Review, found that hyperscaler bond issuance topped $100 billion in 2025 and that credit default swap spreads had risen, particularly for lower-rated borrowers, reflecting both the volume of supply and "uncertainties around the projects' payoffs".

Watch for divergence rather than direction. When lenders stop treating AI as one trade and start pricing individual borrowers differently, that discrimination usually arrives before the equity market catches up. S&P Global Market Intelligence flagged exactly this pattern in a note dated 2 March 2026.

3. Debt that never reaches the balance sheet#

The BIS describes a second channel. Data centres are increasingly financed through special purpose vehicles and joint ventures, where the operator holds a minority stake and signs a long-term lease or capacity agreement while the debt sits outside its own accounts. The BIS calls this shadow borrowing: obligations economically akin to debt but largely residing off corporate balance sheets.

Who holds that debt matters. The Bank of England, citing OECD figures, reports that the share of AI investment financed by private credit rose from 9% in 2024 to 34% in 2025. It also cites Morgan Stanley estimates that more than half of global data centres' external financing need between 2026 and 2028 could be met with debt. Separately, a BIS bulletin dated 14 July 2026 found that business development companies had lent roughly $115 billion to software firms, more than 80% of their technology portfolios, while spreads on those loans had actually fallen and converged. Concentration compounds the problem: the five largest such lenders account for around 37% of software lending.

4. The adoption gap: what businesses are actually doing#

Spending is a forecast of demand. Adoption is demand. The US Census Bureau's Business Trends and Outlook Survey, reported in May 2026, found overall AI use among American businesses hovering between 17% and 20%. Among firms with 250 or more employees it reached 37%. Among firms with four or fewer employees it stayed below 20%, and adoption among firms with fewer than 20 employees did not change significantly between December 2025 and May 2026.

Set that against the projections the IMF cites, under which hyperscaler AI investment could top $3.4 trillion by 2030. The gap does not prove a bubble. Electricity and the railways both had long lags between building and using. But it is measurable, fortnightly and free, which makes it the most accessible signal here.

5. Concentration, and the assumptions under the earnings#

Concentration does not cause a crash. It determines how much damage one does. The IMF found that two of six major equity markets exceeded the 95th historical percentile on the Herfindahl-Hirschman Index, a standard measure of how much of an index sits in a handful of names. The Federal Reserve's May 2026 Financial Stability Report added that the equity risk premium remained well below its historical average and that hedge fund leverage was near all-time highs.

The quieter half of this signal is accounting. Depreciation schedules for AI hardware are set by management judgement and disclosed in annual filings. Shortening an assumed useful life would cut reported profits without a single customer being lost, which is why those footnotes deserve more attention than they usually get.

What the signals say in September 2026#

Three of the five are amber, one is arguably red, and one is not flashing at all.

The financing mix has changed decisively. That is fact rather than interpretation. Off-balance-sheet structures are growing and are, by construction, hard to size. The adoption gap is real and persistent, especially among small firms. Concentration sits at or near historical extremes.

The signal that is not flashing is fundamental performance. Nvidia's most recent quarter beat expectations and it guided to $108 billion for the next one. Anyone claiming the numbers have already cracked is not reading the filings.

One detail sits awkwardly between the two camps. Nvidia's accounts receivable stood at $63.1 billion at the end of July 2026, against $38.5 billion at its fiscal year end. Receivables growing faster than revenue can mean strong demand booked late in a quarter. It can also mean customers taking longer to pay. That is a question rather than a verdict, and worth carrying into the next set of accounts.

Where the dot-com comparison helps, and where it misleads#

The 2000 parallel is popular because the shape rhymes: a genuine technology, an investment surge, valuations that need heroic growth. The Bank of England notes that the S&P 500's excess cyclically adjusted earnings yield has edged towards levels last seen in the dot-com era.

Three differences matter. Today's biggest spenders are profitable businesses rather than pre-revenue start-ups. The assets are physical and have alternative uses, unlike fibre laid for traffic that never arrived, though depreciation still bites. And more of the financing sits with private credit funds and off-balance-sheet vehicles rather than public equity holders, which changes who absorbs losses and how visibly.

The BoE also notes something that cuts against the simple "it's just AI" story. Excluding the top 30 AI-related stocks, US valuations are around their lowest since 2007 on the same measure. That suggests a broad stretch in valuations rather than one sector's enthusiasm.

Using signals without pretending to predict#

None of this tells you what to buy or sell, and this article does not try to. What the signals offer is a way to swap a date for a dashboard.

Three habits make that practical. Track the quarterly financing mix, capex against operating cash flow, in filings rather than headlines. Read central bank stability reports, which are free and written for non-specialists. The BoE's scenario of a 45% US equity fall over six quarters, alongside a 350 basis point widening in credit spreads, is a stress test rather than a forecast, and knowing that difference is most of financial literacy. And check whether your own exposure is more concentrated than you assume. An index fund holding "the whole market" now holds one where AI-linked companies are about half of US market capitalisation.

The honest answer to when it will burst is that nobody knows, and that the question may be the wrong one. Booms can also deflate slowly, through years of flat prices while earnings catch up. What is knowable is which numbers move first, and those are published, free, on a schedule.

Key takeaways#

1. The financing mix has already shifted. Incremental debt rose from 9% of hyperscaler capex in FY24 to 32% by mid-2026, with free cash flow heading towards zero or negative for most of the group.
2. Credit markets lead equity markets. The BIS recorded rising CDS spreads on lower-rated AI borrowers alongside more than $100 billion of hyperscaler bond issuance in 2025.
3. Much of the debt is hard to see. Private credit's share of AI financing rose from 9% to 34% in a year, and off-balance-sheet vehicles keep obligations outside published accounts.
4. Adoption lags spending. Only 17% to 20% of US businesses report using AI, and small-firm adoption has not moved since December 2025.
5. Concentration sets the damage, not the timing. With AI names at around half of the S&P 500, index exposure is less diversified than it looks.

Frequently asked questions#

Is there definitely an AI bubble?

No official body has said so. The Bank of England, the IMF and the Federal Reserve describe valuations as stretched and the risk of a sharp correction as material. That is a statement about vulnerability, not a prediction.

Can anyone actually predict the date?

No. Forecasts of bubble timing have a poor historical record. The signals described here indicate stress building, not when it releases.

Would a burst hit my pension or index fund?

It would affect any portfolio holding US large-cap equities. The BoE's stress scenario models a 45% US equity fall over six quarters reducing UK GDP by 2.2 percentage points. That is a deliberately severe test, not an expectation.

Why does depreciation matter so much?

Because the assumed useful life of AI hardware directly determines reported profits. Shorter assumed lives mean higher annual depreciation charges and lower earnings, with no change in the underlying business.

How is this different from 2000?

The main spenders today are profitable and the assets are physical. But more of the financing sits with private credit and off-balance-sheet vehicles, which makes losses harder to trace.

What is the cheapest signal to follow?

The Census Bureau's Business Trends and Outlook Survey. It is published fortnightly, it is free, and it measures actual business use of AI rather than expectations of it.

Does high concentration always end badly?

No. Concentration raises an index's sensitivity to a small number of companies. It amplifies whatever happens next, in either direction.

Should I sell my AI-related holdings?

That depends on personal circumstances this article cannot know. Nothing here is investment advice, and a regulated adviser is the appropriate source for it.

Glossary#

Capital expenditure (capex) Money spent acquiring or upgrading long-lived physical assets such as data centres, servers and power infrastructure.

Free cash flow Cash generated by operations after capital expenditure. When it is negative, the shortfall must be met by borrowing or issuing shares.

Credit default swap (CDS) A contract insuring against a borrower's default. Its price, quoted in basis points, reflects the market's estimate of default risk.

Basis point One hundredth of a percentage point. A 350 basis point move equals 3.5 percentage points.

Special purpose vehicle (SPV) A separate legal entity created to hold specific assets and debts, keeping them off the parent company's balance sheet.

Private credit Lending by non-bank institutions such as investment funds, typically outside public bond markets and with less public disclosure.

Herfindahl-Hirschman Index (HHI) A measure of concentration. Applied to a stock index, a high reading means a small number of companies dominate its value.

Cyclically adjusted price-to-earnings (CAPE) ratio A valuation measure comparing prices with average inflation-adjusted earnings over ten years, which smooths out the business cycle.

Depreciation Spreading an asset's cost over its assumed useful life. Longer assumed lives produce lower annual charges and higher reported profits.

Equity risk premium The extra return investors expect from shares over government bonds. A low premium means investors are accepting little compensation for risk.

References#

1. Bank of England, Financial Stability Report, July 2026.
2. International Monetary Fund, Global Financial Stability Report, April 2026.
3. Board of Governors of the Federal Reserve System, Financial Stability Report, May 2026 (Overview). Data as of 23 April 2026.
4. Bank for International Settlements, "Financing the AI infrastructure boom: on- and off-balance sheet", BIS Quarterly Review, March 2026.
5. Bank for International Settlements, BIS Bulletin No 128, "AI disruption in private credit: exposure to software firms", 14 July 2026.
6. NVIDIA Corporation, financial results for the second quarter of fiscal 2027, filed with the US Securities and Exchange Commission, 26 August 2026.
7. US Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users", May 2026.
8. FactSet, "Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow".
9. S&P Global Market Intelligence, "What Do Credit Default Swaps in the Technology Sector Signal Ahead of Upcoming Earnings?", 2 March 2026.
10. Yahoo Finance, "Stock market today: Dow, S&P 500, Nasdaq end week on down note", 28 August 2026.