AI vs Dot-Com Bubble: Five Measures of Market Risk

Five measures—valuation, earnings, concentration, market breadth and capital expenditure—give a more useful comparison of the AI rally with the dot-com era than a similar-looking price chart. Amundi’s comparison finds no explosive valuation dynamics in its studied AI basket, while identifying concentration and the durability of earnings and spending as risks. The evidence supports neither a simple replay of the late-stage dot-com bubble nor a verdict that AI-related shares are safe.

The risk is consequential even if the technology succeeds. The ECB’s analysis places US cyclically adjusted stock valuations near their historical peak and argues that a correction is likely whether enthusiasm for AI is rational or excessive. Its warning concerns the broad market and the effects of a possible correction; the AI basket study tests a narrower set of stocks and their valuation dynamics.

The five-metric scorecard

The historical figures below come from Amundi’s portfolio tables. Its dot-com basket contains technology, media and telecom stocks, while its recent basket selects S&P 500 companies using the researchers’ AI-exposure method. The groups are constructed differently, and their observation periods differ. Their figures reveal patterns within each cycle rather than a precise ranking of which market was more speculative.

Valuation: Compare how quickly prices rise relative to earnings, as well as the multiple reached. The dot-com basket’s average trailing price-to-earnings ratio rose from 21.4 in 1997 to 43.7 in 1999. The AI basket’s annual average fell from 40.4 in 2023 to 34.2 in 2025, even as its share prices rose over that period. Earnings growth can make a multiple fall during a rally; a falling multiple does not make the shares cheap. Persistent price gains that again outpace earnings would bring this measure closer to the late dot-com pattern.
Earnings: Reported profit is a stronger foundation than a claim about distant demand, but the source of that profit matters. J.P. Morgan Asset Management’s earnings review found that S&P 500 earnings grew 29% in the second quarter of 2026 after excluding one-off items, while only 2% of companies quantified AI-driven productivity gains on earnings calls. Those are broad-index and company-disclosure measures, not a profit figure for Amundi’s AI basket. They distinguish profits already earned from the wider productivity gains still expected to justify investment.
Concentration: Measure both the theme’s index weight and the number of stocks carrying it. The constructed dot-com basket reached 44.8% of the S&P 500 across 96 stocks in February 2000. Amundi’s AI basket accounted for 31.4% across 48 stocks at the end of March 2026. The latter is a later observation, not the AI basket’s peak, so comparing the percentages alone understates the methodological difference. Fewer constituents mean that weakness in a handful of large companies can have an outsized effect on an index fund, including for investors who never deliberately bought an AI stock.
Market breadth: Look at returns outside the leading theme as well as its contribution to the index. In 1999, the dot-com basket supplied 82.1% of the S&P 500’s gain while the basket excluding it returned 4.5%. In 2025, the AI basket supplied 57.3% of the gain and the basket excluding it returned 12.8%. Both years had narrow leadership, but the stronger return outside AI gave the more recent index another source of gains. These are different calendar years and differently selected groups, so the gap is evidence about breadth within each period, not a bubble threshold.
Capital expenditure: Spending matters most in relation to cash generation, financing and the eventual return on new capacity. S&P Global’s capital-spending research estimates that five major technology companies spent $1.1 trillion in total capital expenditure over the preceding five years; it identifies a shift from cash-funded to debt-funded expansion before returns are validated as a central risk. That total covers their capital expenditure, not an audited AI-only amount. Large spending can be productive, but its scale raises the cost of a mistaken demand forecast.

Why the institutional conclusions differ

The valuation measures have different subjects and time horizons. A cyclically adjusted price-to-earnings ratio compares a broad market price with inflation-adjusted earnings averaged over a long period. The trailing multiple for a selected AI basket compares that group’s price with more recent profits. Strong current earnings can compress the basket multiple while the broad US market remains expensive against its longer earnings history. Both findings can therefore be true at once.

The studies also test different propositions. The absence of explosive multiple growth in the AI basket over the studied period is evidence against one hallmark of a late-stage speculative bubble; it cannot establish what valuations will do later. The correction argument does not depend on AI failing. As a technology spreads through the economy, risks once confined to a few companies may become harder for investors to diversify, increasing the return they demand. That change can lower share prices even if the technology produces real profits. Neither approach identifies a date or level at which a correction must begin.

Where the analogy becomes more persuasive

The strongest dot-com resemblance would be a combination: multiples accelerating ahead of earnings, index gains relying on an ever narrower group, weaker returns outside that group, and capital commitments rising without matching cash flow or returns. Any one measure can mislead. A high multiple can reflect expected growth, and heavy investment can precede valuable capacity; together with deteriorating earnings and breadth, they present a different risk than spending or valuation alone.

For the periods measured here, the scorecard is mixed. The AI basket’s falling earnings multiple and the market’s reported profits weaken a direct late-bubble analogy. Concentration, expensive broad-market valuations and the size of the capital build-out leave substantial room for a painful correction. That is the distinction the five measures expose: a market need not repeat the dot-com valuation path for its losses to matter.