The AI capex cycle is likely to end badly. Not catastrophically, but badly enough to reset expectations and take the air out of one of the most crowded trades in the market.
Here is why.
The five largest hyperscalers are spending more than $1 trillion on AI-related capex over 2025–26. At the same time, the S&P 500 is close to record highs. The market sees no contradiction between the two.
I do.
The important number is not total capex. It is capex relative to operating cash flow.
JPMorgan estimates that AI capex has risen from 33% of hyperscaler operating cash flow in 2023 to roughly 93% in 2026.
At 93%, there is very little room for disappointment.
That leaves less cash for share buybacks, dividends, acquisitions and debt reduction. More importantly, it means the infrastructure build is consuming almost all of the cash generated by the existing business.
The problem is not that AI fails
The bear case is not that AI does not work.
Microsoft, Amazon and Alphabet are not going to default. Their businesses generate substantial cash. The technology is real, and demand for AI services is real.
The problem is the amount of capital being committed relative to the returns that infrastructure may ultimately generate.
I have spent 25 years in credit markets and have watched several investment cycles develop along similar lines: fiber optics in the late 1990s, shale in the mid-2010s and cloud infrastructure in the early 2020s.
In each case, the underlying technology worked. Companies built real businesses and generated real revenues.
But capital was deployed faster than profitable demand could absorb it.
That distinction matters.
AI revenues could continue growing at 30–40% annually while capex grows at 50–70%. At some point, depreciation accelerates, electricity costs rise, computing becomes cheaper and data centers are not used as intensively as expected.
The result can be strong revenue growth alongside declining incremental returns on invested capital.
I have seen this pattern before.
The revenue line can look perfectly healthy while the returns deteriorate underneath it.
Why this matters for the broader market
The concentration of the S&P 500 makes this more than an AI-sector issue.
The Magnificent Seven account for roughly 34% of the index, while Nvidia alone represents around 8%.
That concentration creates a powerful transmission mechanism if the economics of AI infrastructure deteriorate.
The sequence is straightforward:
AI capex slows.
Semiconductor orders decline.
Credit spreads widen.
Data-center financing becomes more expensive.
Marginal projects are canceled.
Economic growth slows.
Earnings estimates come down.
Equity multiples contract.
I would assign a meaningful probability to four of those five steps occurring within 18 months of a capex peak.
I do not expect this to resemble 2008. The companies at the center of the investment cycle have real cash flows, substantial balance-sheet strength and strong credit ratings.
The better comparison is the telecom investment cycle of the late 1990s and early 2000s.
The infrastructure was real. The demand was real. But too much infrastructure was built ahead of the demand.
When the cycle turned, investors discovered that infrastructure could be economically unproductive even when the underlying technology was successful.
My base case
My base case is that pressure builds through 2027 as spending continues to increase while investors become increasingly focused on the returns generated by that spending.
By 2028, weaker AI projects begin to be canceled and the effects spread beyond the technology sector.
In that scenario, I would expect a 25–30% correction in the S&P 500, with credit markets repricing risk before equities fully reflect the deterioration.
The key question is simple:
Will the cash generated by AI infrastructure be sufficient to justify the capital being committed today?
My answer is no — not at current spending levels and not on the timeline currently embedded in valuations.
Watch the cash flow, not the headlines
I am not predicting a market crash.
I am arguing that investors should be more selective about what they are treating as an unqualified positive.
AI capex is a bet on future returns. Those returns have not yet been demonstrated at the scale required to justify today’s investment.
The longer the returns take to materialize, the larger the gap becomes between capital deployed and cash generated.
For me, the indicators to watch are straightforward:
Capex as a percentage of operating cash flow
Depreciation and amortization
Data-center utilization
Free cash flow after capex
Debt issuance and financing costs
Returns on incremental invested capital
Those numbers will tell us considerably more than the next round of AI announcements.
The question is not whether AI works.
It does.
The question is whether the industry is building enough infrastructure to satisfy future demand — or far more than the economics can support.
That is where the next phase of this cycle will be decided.
Alessandro Rovelli
Rovelli Research
Independent credit investment research
