The bubble within the bubble: valuations, AI and inflated earnings

The debate over a possible stock-market bubble, fuelled by artificial intelligence, plays out on two registers that answer each other poorly. The first is qualitative: judging whether an industry’s real potential justifies its prices. The second is quantitative: comparing valuations with their own history. A market note relayed by the Financial Times in early July raised eyebrows precisely because it pushes the second register to its logical conclusion, with a deliberately provocative number. Let us take both registers, methodically, before drawing a reading.

Two ways to sense a bubble

This dual reading grid, framed by the entrepreneur and digital thinker Gilles Babinet in commenting on the episode, sets two ways of judging a bubble against each other. The first method assumes fine knowledge of the sector: estimating that its speed of development justifies very high prices. Applied to AI, this reading rests on an imaginary of rupture, summed up in a series of acronyms that have become totems in the industry: AGI for artificial general intelligence, ASI for superintelligence, RSI for an AI able to improve itself. These horizons are championed by the field’s leading players. In his essay “Machines of Loving Grace”, Anthropic’s CEO Dario Amodei places the arrival of a “powerful AI”, with capabilities matching or exceeding the best human specialists across most disciplines, as early as late 2026 or early 2027. Elon Musk announces his Optimus humanoid robot going on sale for late 2027, between $20,000 and $30,000, promised for domestic uses. The conviction that these milestones will be met, and on those dates, is the bedrock of the bullish scenario.

The second method dispenses with forecasting the future: it takes the big valuation ratios. The P/E, the ratio of a stock’s price to its earnings, and above all the CAPE, popularised by the economist Robert Shiller, which relates price to the ten-year average of inflation-adjusted earnings, so as to smooth the cycle. It is on this terrain that the number circulating this week sits.

The number that gave the FT pause

In FT Alphaville, on 3 July 2026, Bryce Elder relays a monthly note from strategists Joachim Klement and Francisca Reis, of Panmure Liberum. Their reasoning comes in two steps. First, the classic observation: per Shiller’s data, the S&P 500 CAPE was 32.6 in 1929, 1.8 standard deviations above its trend, and 44.2 in 2000, 3.3 standard deviations, a clear bubble signal. Today it is at 41.0, 2.9 standard deviations above trend. Bubble territory, then, but nothing unprecedented at this stage.

The second step is the more original. In 1929 as in 2000, earnings were within their normal range, less than one standard deviation from their trend. Today they are themselves 1.8 standard deviations above. In other words, the valuation is high at a moment when the denominator, earnings, is already abnormally inflated. Corrected for this anomaly, the CAPE would come out not at 41 but at 67.6, 4.6 standard deviations above trend, a level that exceeds anything US history has known. Klement draws a striking image from it: under the assumption, false and acknowledged as such, of a normal distribution of valuations, such a level would occur in 0.00019% of months, or once every 43,432 years.

A point of method is in order, and the FT author stresses it first: this “once every 43,432 years” must not be taken literally, because valuations do not follow a normal law. It is a way of expressing the size of the gap, not a real probability. Elder adds the usual caveat: supra-normal profits always end up normalising, but trying to time the turn is a mug’s game. Rigour commands keeping the order of magnitude without lending it false precision.

The bubble within the bubble: earnings themselves

The heart of the thesis is therefore not the CAPE, but the denominator. The idea that current earnings are abnormally high holds up beyond the Panmure note alone. US corporate margins run around 14%, a record high in the available data, supported by factors that are not all durable: sector concentration, favourable taxation, massive share buybacks that flatter earnings per share, and the direct effect of AI spending on a few champions.

Yet the profit margin is, in the phrase of investor Jeremy Grantham, “probably the most mean-reverting series in finance, and if profit margins do not mean-revert, then something has gone badly wrong with capitalism”. The asset manager GMO devoted a study with a telling title to this anomaly, “The Curious Incident of the Elevated Profit Margins”. If this logic holds, using inflated earnings in the denominator of a valuation ratio makes the market look cheaper than it is. That is exactly the mechanism of the “bubble within the bubble”: an overvaluation laid on profits that are themselves in excess.

The rising doubt over AI profitability

What makes the question burning in 2026 is that the bad news is piling up on the real profitability of AI, both for those who sell it and for those who buy it. Infrastructure spending is exploding far faster than revenue. The five largest hyperscalers plan between $700 and $900 billion of investment in 2026, up about 36% year on year, while the AI ecosystem would generate far lower revenue, leaving a shortfall estimated at around $600 billion a year. According to Allianz Research, the divergence between AI investment and revenue growth reaches about 46%, beyond the 32% seen during the 2001 telecom excess, which preceded a brutal and durable correction.

OpenAI’s case illustrates the tension: about $25 billion of annualised revenue in early 2026, but past losses of $540 million in 2022, $1.5 billion in 2023 then $5 billion in 2024, with no profitability expected before 2029. On top of that sits a partly circular financing architecture, where a fraction of declared revenue is capital recycled among interconnected players rather than independent organic demand, a risk we described in AI circular financing and the fragility flagged by the BIS. Finally, the value actually created stays uncertain: as our review of the evidence on AI productivity shows, the gains are real at the task level but struggle to diffuse to the economy, which limits the pool of revenue capable of justifying the investment.

The antithesis, taken seriously

Rigour forbids stopping at the bearish thesis, however well supported. Several solid objections deserve to be raised.

First, the CAPE has its limits, widely documented. It is criticised for ignoring changes in accounting standards, the rise of buybacks, the forty-year secular decline in interest rates, and the shift in the market’s sector composition, lighter in capital than it once was. Some argue that part of the rise in margins is structural, driven by dominant, capital-light firms, and not a mere cyclical excess bound to correct. If this reading is right, the denominator is not so inflated, and the “bubble within the bubble” deflates on its own.

Next, even accepting the diagnosis, the timing stays unpredictable. An extreme valuation says an asset is expensive, not that it will fall tomorrow. Bubbles can inflate for years, and betting on their bursting has ruined plenty of sceptics: over the past decade, recurrent warnings of a “tech bubble” caused many to miss a considerable rise in the indices. A high CAPE is an indicator of mediocre ten-year forward returns, not a timing signal.

Finally, the bullish scenario has its internal coherence. If the industry’s totems materialise, even in part, future earnings could grow fast enough to catch up with prices, and today’s valuation would then be nothing aberrant. A truly transformative AI would justify high and durable margins. That is the bet, perfectly defensible in theory, of those who buy the promise rather than the proof.

Reality’s verdict

The weak point of the bullish scenario is not its logic, it is its dependence on dated predictions. And dated predictions have a merciless judge: the calendar. Financial history is a graveyard of expert convictions belied at the worst moment. In October 1929, days before the crash, the economist Irving Fisher, a luminary of his discipline, assured that prices had reached “what looks like a permanently high plateau”, an episode John Kenneth Galbraith recounts in “The Great Crash 1929”. The lesson is not that optimists are always wrong, but that a player’s closeness to an industry does not protect it from excess enthusiasm, it exposes it to it.

What to conclude, without yielding to either catastrophism or denial? Three things, that the evidence permits. US valuations are, on historical measures, at rarely reached peaks, and the margin of safety is thin. These valuations rest on earnings whose durability is contested, which doubly weakens the structure. And the scenario that would justify them depends on technological promises still unkept, whose deadline is approaching. None of this says when, or even whether, a correction will come: the timing stays out of reach, as the classics of the genre recall, from Robert Shiller’s “Irrational Exuberance” to Charles Kindleberger’s “Manias, Panics, and Crashes”. But the honest analyst must distinguish a promise from a proof, and acknowledge that today, an important share of stock prices rests on the former.

Sources

1. Bryce Elder, FT Alphaville, 3 July 2026, relaying the Panmure Liberum note (Joachim Klement, Francisca Reis): CAPE of 32.6 in 1929, 44.2 in 2000, 41.0 in 2026; earnings 1.8 standard deviations above trend; corrected CAPE at 67.6 (4.6 standard deviations); 0.00019% of months, or once every 43,432 years, under a normality assumption explicitly presented as false: https://www.ft.com/content/8e9337f8-9191-48e9-9289-a8defda89431
2. Robert Shiller, historical CAPE data and “Irrational Exuberance” (Princeton University Press, 2000): construction and reading of the cyclically adjusted valuation ratio: http://www.econ.yale.edu/~shiller/data.htm
3. Dario Amodei, “Machines of Loving Grace”, personal essay: arrival of a “powerful AI” possible as early as late 2026 or early 2027: https://www.darioamodei.com/essay/machines-of-loving-grace
4. Entrepreneur / statements by Elon Musk: Optimus going on sale targeted for late 2027, between $20,000 and $30,000, domestic uses: https://www.entrepreneur.com/business-news/elon-musk-tesla-sell-optimus-humanoid-robots
5. Jeremy Grantham (GMO), quote on the mean reversion of margins, and GMO, “The Curious Incident of the Elevated Profit Margins”: https://www.gmo.com/americas/research-library/the-curious-incident-of-the-elevated-profit-margins-part-1_whitepaper/
6. Fortune, 19 May 2026, Jeremy Grantham on the “AI war” and coming pressure on margins: https://fortune.com/2026/05/19/blood-in-the-streets-jeremy-grantham-ai-monopoly-brutal-competitive-world-recession/
7. Allianz Research, 2026: AI capex-revenue divergence of about 46%, beyond the 32% of the 2001 telecom cycle: https://www.allianz.com/content/dam/onemarketing/azcom/Allianz_com/economic-research/publications/specials/en/2026/march/2026_03_25_AI.pdf
8. Forbes, 2 June 2026, widening gap between AI spending and revenue, hyperscalers between $700 and $900 billion of capex in 2026: https://www.forbes.com/sites/jasonkirsch/2026/06/02/the-ai-capex-to-revenue-gap-is-widening---and-markets-are-starting-to-notice/
9. FutureSearch, OpenAI financials: about $25 billion of annualised revenue in early 2026, losses of $540 million (2022), $1.5 billion (2023) and $5 billion (2024), profitability expected around 2029: https://futuresearch.ai/openai-revenue-forecast/
10. John Kenneth Galbraith, “The Great Crash 1929” (1955): Irving Fisher’s statement on a “permanently high plateau” on the eve of the 1929 crash.
11. Charles P. Kindleberger, “Manias, Panics, and Crashes: A History of Financial Crises” (1978): the recurring anatomy of bubbles and their turn.
12. Gilles Babinet, comment on X reacting to the FT article (secondary source, analysis rather than primary source): proposes the two-method reading grid taken up here, compiles the AI industry’s dated predictions and recalls the precedent of experts belied in 2000: https://x.com/babgi/status/2073731501045248405