Something unusual happened in May 2026. The cyclically adjusted price-earnings (PE) ratio for US equities—the Shiller CAPE, which smooths out the cycle to give a longer view of valuation—touched 40.8. The only other time in 145 years it had exceeded 40 was in the fevered months around the dotcom peak of early 2000.
Simultaneously, the University of Michigan’s long-running survey of American consumers recorded sentiment at its lowest level in over 70 years of measurement—some 10% below the nadir of June 2022, when inflation was at its most punishing. (Justin Lahart, Wall Street Journal, 23 May: bit.ly/3RAHYYQ)
These two facts are not in contradiction. They describe the same underlying reality from opposite ends of the income distribution.
Artificial intelligence (AI), the narrative sustaining the market, is good for capital. It compresses labour costs, widens margins and promises productivity gains that, at a sufficient discount rate, justify almost any present valuation.
For the worker contemplating what the next model release implies for her employment, the calculation runs rather differently. Both the market and consumer survey are right.
That is precisely the problem—and understanding it requires looking at what the AI investment cycle actually promises and what it is actually delivering. The promise, in financial terms, does not hold up.
Alan Greenspan used the phrase “irrational exuberance” in December 1996 to describe the technology boom of that era; the bubble burst in March 2000, three years and three months later.
Joachim Klement of Panmure Liberum has applied the same characterisation to the AI boom (bit.ly/3PSeOnp), and his arithmetic is difficult to dismiss. In 2025, businesses in leading advanced economies invested nearly $1.5 trillion in IT equipment and software—almost double the inflation-adjusted dotcom peak.
Over the past four quarters, 93% of US GDP growth has been explained by technology investment alone. Yet, when Klement examined consensus analyst estimates for the five principal hyperscalers—with capital expenditure growing at 20% annually to 2030, revenues at 15%—even the most charitable assumption, that every dollar of incremental revenue falls straight to profit, yields a negative return on investment for four of the five.
To achieve a 10% return, these companies, currently generating combined revenues of roughly $1.5 trillion, would need to find between $2 trillion and $5 trillion in additional annual revenue.
The operational evidence confirms the same from the ground up (bit.ly/4wWVy8Q). A company that positioned AI as central to its identity recently cancelled nearly all of its engineers’ AI coding licences: token-based pricing at scale had produced costs exceeding those of the human labour that the tools were meant to replace.
The chief technology officer of a major global platform described his annual AI budget as having been “blown away” by April. A senior executive at the world’s largest chip company, a direct beneficiary of every dollar of AI capex, observed matter-of-factly that computing costs for his team now exceed the cost of his employees. Goldman Sachs projects a 24-fold increase in token consumption by 2030; Gartner adds the uncomfortable corollary that even if individual token prices fall 90%, total enterprise AI costs will still rise, because agents consume exponentially more tokens per task than simpler tools.
The financial arithmetic is implausible from the above. The operational arithmetic is unworkable from below. Both point to the same destination.
Against this backdrop, the developing world’s alleged weak position on AI looks rather different from the conventional framing. The capital that major developing economies deployed over this period went into roads, power generation, metro systems, inland water transport and early semiconductor capacity.
The implicit comparison between this and data centres generating negative implied returns is not unfavourable to the former. If a correction comes, and the arithmetic suggests it will, the infrastructure will still be there. Roads carry freight regardless of whether a particular investment narrative has aged well.
There is also a systemic risk that the developing world is better off not inheriting. A computational hydrologist at the University of Calgary recently described how, in six days, he built a production-grade autonomous trading platform—linked to multiple exchanges, ingesting social media feeds, routing orders around the clock—using tools available to anyone with a laptop subscription (bit.ly/433Rwhh).
Regulators at the Financial Stability Board, Bank for International Settlements and Bank of England have all warned of “model monocultures”—the phenomenon in which thousands of AI agents sharing similar architecture reach the same conclusion simultaneously, at machine speed, about a sovereign debt position that human analysts might have read differently.
Unlike the crises of 1997 or 2010, there are no intervention windows when the herd moves in microseconds. The developing world’s relative distance from this architecture is not a disadvantage.
Three further channels reinforce the point. A US recession driven by AI capex retrenchment would be a sectoral correction, not a banking system crisis, and need not trigger the indiscriminate risk-off that strengthened the dollar in 2008.
Central bank rate cuts in affected economies would compress yield differentials and redirect capital towards economies still growing at 6% in real terms. And lower oil demand—data centres are significant energy consumers—would ease global crude prices, improving simultaneously the current account of every major oil-importing developing nation.
The standard geopolitical counter-argument—that any cooling of AI investment cedes ground to authoritarian competitors—conflates two things that are not the same. Maintaining AI capability is not the same as maintaining the current hyperscaler model. Early 2025 demonstrated, briefly and memorably, that frontier performance does not require $1.5 trillion in annual IT investment. Efficient IT architecture and open-source development can close gaps that concentrated capital was assumed to have permanently sealed.
Greenspan spoke of irrational exuberance in December 1996. The reckoning came three years and three months later. Timing, as ever, is the one thing the arithmetic cannot tell you. But the math itself is not in dispute. When the hype fades—and it will—roads will still carry freight, power plants will still generate electricity and the capital that built things would have outlasted capital that built narratives. That, in the end, may be the developing world’s quiet vindication.
These are the author’s personal views.
The author is chief economic advisor to the Government of India.