Mentioned in This Article
1. Torsten Slok's AI boom warning: the core argument
Torsten Slok, chief economist at Apollo Global Management, has spent recent days cautioning that the artificial intelligence boom may be built on figures that are starting to weaken. In an appearance on October 22, 2026, he described a wide gap between the heavy capital spending of the largest cloud and data-center operators and productivity that has barely moved across the wider economy.
His case rests on a simple tension. Hyperscalers, the companies that build and run vast computing capacity, are committing enormous sums to AI infrastructure, and share prices have rewarded that spending. For the investment to pay off, businesses elsewhere in the economy have to use the technology to earn more from each employee or to widen profit margins. Slok argues that the evidence for that second step is missing.
Slok is a well-known voice in macro research, and his role at Apollo, a large alternative asset manager with significant exposure to credit markets, gives his view on financing a particular weight. Credit investors tend to focus on whether borrowers can service their debts, which makes them an early reader of strain in any investment cycle. His concerns therefore combine two questions: whether AI will earn its keep, and whether the borrowing behind it can be sustained.
The warning matters beyond technology stocks. AI enthusiasm has been a leading driver of risk appetite across equity markets, and crypto assets have often traded in step with growth shares. A reassessment of AI valuations would therefore reach well past the companies directly involved.
2. What is the S&P 493 problem?
The S&P 493 is the S&P 500 with the Magnificent Seven removed. Those seven mega-cap technology companies have supplied a large share of the index's returns in recent years, so stripping them out leaves a picture of how the typical large American company is doing. It is the sharpest part of Slok's argument.
According to his analysis, those 493 companies show flat revenue per worker and flat profit margins despite the AI investment wave. If AI were lifting productivity broadly, revenue generated per employee should be rising, and margins should widen as work gets done faster or at lower cost. Neither shows up in the data he cites.
One consequence, in his reading, is that many firms treat AI spending as a cost of doing business rather than as a driver of growth. A company may buy AI tools because competitors do or because customers expect it, not because the tools raise earnings. Spending of that kind keeps revenue flowing to AI suppliers for a time, yet it does not create the new profits that would justify the valuations attached to the sector.
It helps to separate two ideas that are easy to blur. Revenue per worker measures how much a company sells for each person on its payroll, so it rises when staff become more productive or when the firm sells more without hiring. Profit margin measures how much of each dollar of sales remains after costs. A technology that genuinely saves labor or opens new markets should lift at least one of these over time. Slok's point is that, among the 493, neither has shifted.
Slok warns that AI valuations face a painful repricing if the expected returns do not arrive quickly. Repricing here means share prices adjusting downward to reflect lower growth than investors had assumed, and the word painful signals that he expects the adjustment to be sharp rather than gradual.
3. Why does hyperscaler spending near 3% of GDP stand out?
Slok projects that hyperscaler capital expenditure could reach roughly 3% of US gross domestic product annually from 2027 through 2029. Gross domestic product is the total value of goods and services the country produces in a year, so a single group of companies directing about three cents of every dollar of national output toward one category of investment is a very large commitment.
He compares the scale with earlier economic booms that ended in market corrections. Past waves of heavy investment in railways, telecommunications networks and internet infrastructure each brought genuine technological progress, yet each also saw prices for the related assets overshoot before falling back. The comparison does not predict a collapse of the technology itself, only that the financial returns on a wave of building can fall short of what prices implied at the peak.
The share of national output matters because it measures spending against the size of the whole economy rather than against any one company's budget. At that scale, the buildout begins to compete with other uses of money, labor and electricity, and its success or failure becomes a macroeconomic question as well as a corporate one. A disappointment would not stay confined to the firms that placed the orders.
Capital spending on this scale also has to be financed. Companies can pay from internal cash flow, but they can also borrow, and Slok points out that credit spreads on AI-related debt are widening. A credit spread is the extra yield investors demand to hold a risky bond over a safe government one, and when it grows, borrowing becomes more expensive and lenders signal greater caution.
4. How could debt and inflation slow the buildout?
Inflation is adding pressure on top of wider spreads. Higher prices tend to keep interest rates elevated, which raises the cost of the borrowing that funds construction. For projects that depend on borrowed money, the combination of costlier debt and rising input prices can change the arithmetic that justified them.
Slok warns that data-center projects relying on debt could therefore face delays, and that critical projects may be pushed back. A postponement would carry its own signal. Spending commitments are the foundation of current revenue forecasts for chipmakers, equipment suppliers and power providers, so any slippage in timelines would feed back into expectations for the whole supply chain.
The mechanism is worth spelling out. A data center takes years to plan, permit, build and fill with equipment, and each stage is paid for in advance of any revenue. That long gap between outlay and income is why the projects are so sensitive to the cost of money. When borrowing costs rise, the return a project must deliver to justify itself rises with them, and marginal projects stop clearing the bar.
It would also dampen the enthusiasm that has supported funding for AI projects in the first place. Investor willingness to finance new capacity depends partly on confidence that earlier capacity is being used profitably. If credit markets begin to doubt that, capital can become harder to raise precisely when the buildout requires the most of it.
5. What does this mean for investors and the AI trade?
The most direct effect falls on stocks whose valuations rely on continued growth in AI spending. If the revenue forecasts behind those prices assume steadily rising usage and investment, any evidence that the pace is levelling off makes the forecasts look optimistic, and prices can adjust quickly.
The S&P 493 data raises a separate issue. Even if hyperscalers keep spending, the long-term justification depends on other companies becoming more productive with the technology. Flat revenue per worker and flat margins suggest that the expected gains have not arrived on schedule. That does not rule out later benefits, since new tools often take years to change how businesses operate, but it leaves a gap between what is being paid today and what has been earned so far.
Investors weighing the argument face a timing problem. Technology cycles frequently take longer to deliver measurable results than early enthusiasm assumes, and markets can stay optimistic well beyond the point where skeptics first raise concerns. A warning of this kind identifies a vulnerability, not a date. What it asks investors to consider is how much of the future benefit is already reflected in today's prices, and how they would be affected if the evidence of payoff keeps lagging.
For readers who follow digital assets, the link is indirect but real. Bitcoin and other tokens have tended to move with the broader appetite for risk, and a sharp reassessment of one of the market's most popular themes can spill into other speculative holdings. Slok's remarks are about equities and credit rather than crypto, and they offer no forecast for digital assets.
6. What indicators should be watched next?
Several signals could show whether Slok's concern is borne out. The first is whether credit spreads on AI-related debt keep widening or settle. The second is whether data-center timelines begin to slip, since announced delays would confirm financing strain. The third is the pair of figures at the center of his argument: S&P 493 profit margins and revenue per worker.
Those two measures may matter most. If they begin to rise, the case that AI is lifting the broader economy gains support and the bull argument is strengthened. If they stay flat while hyperscaler capital spending climbs toward 3% of GDP, the risk of repricing becomes harder to dismiss.
No date has been set for any single test of the thesis. Quarterly earnings reports from large companies, which include revenue and margin figures, will provide the next regular data points, and further commentary from Slok and other strategists is likely as those results arrive. Readers should treat the warning as an analyst's assessment of risk rather than a prediction of a specific outcome, and the analysis does not constitute investment advice.