Everything About AI
NextFin News - Wall Street is treating the AI slowdown as a pause, not a reversal. BMO Capital Markets' Jennifer Lee said on September 13 that while calls to slow AI development may create near-term market uncertainty, the technology's infrastructure, investment and demand are unlikely to reverse course. The view captures the prevailing mood on the Street: the growth rate of AI capital spending is decelerating, but the spending itself is still climbing toward records.
The tension is real. Company guidance puts 2026 capital expenditures for the four largest hyperscalers — Amazon, Microsoft, Alphabet and Meta — in a range of roughly $700 billion to $760 billion, up from about $413 billion in 2025. Yet the pace of that climb is slowing: hyperscalers spent $106 billion in the third quarter, a 75% year-over-year increase that analysts expect to moderate to 49% in the fourth quarter and to just 25% by the end of 2026. That second-derivative slowdown has already produced volatility: the Philadelphia Semiconductor Index fell 7.9% in a single session in late June, its worst week since early April, and chip stocks have traded well below their late-June highs since.
But the dominant view among major institutions is that this is a cyclical deceleration inside a structural expansion, not the start of a bust. The question investors are really asking is not whether AI spending will stop, but whether the companies funding it can convert it into revenue before patience runs out.
The Slowdown Is in the Acceleration, Not the Direction
The first thing to get straight is what "slowdown" means here. It does not mean cuts. It means the rate of increase is coming down from an unsustainable sprint.
Goldman Sachs Research tracks the consensus estimate for hyperscaler capital spending in 2026 at $527 billion, up from $465 billion at the start of the third-quarter earnings season — a pattern of upward revisions that has now run for two years. At the start of both 2024 and 2025, consensus estimates implied capex growth of roughly 20% for the year. In reality, it exceeded 50% in both years. Consensus has been wrong in the same direction, twice.
That history matters because it frames the current debate. The bears point to the deceleration — 75% to 49% to 25% — as evidence the boom is topping out. The bulls point to the absolute level: even at 25% growth, spending keeps rising, and the estimates keep getting raised after they are published. A deceleration from an extreme growth rate is not the same thing as a contraction, and confusing the two is the most common error in the current debate.
There is also a valuation gap that explains why investors have turned nervous. The average stock in Goldman Sachs Research's basket of AI infrastructure companies returned 44% year-to-date, compared with a 9% increase in the consensus two-year forward earnings-per-share estimate for the group. Price has run far ahead of the earnings revisions that are supposed to justify it. When a basket trades on a 44% rally against a 9% earnings upgrade, any slowdown in the narrative that drove the re-rating becomes a trigger for profit-taking.
"The timing of an eventual slowdown in capex growth poses a risk to these companies' valuations," Goldman Sachs Research analyst Ryan Hammond wrote in the firm's report.
The market has already started to act on that risk. Since June, the average stock price correlation across the large public AI hyperscalers has declined from 80% to just 20%. Investors are no longer buying "AI" as a single trade. They are rotating away from infrastructure companies where operating earnings growth is under pressure and where capex is being funded via debt, and toward companies showing a clear link between capex and revenues — such as the largest cloud platform operators.
"The combination of continued corporate AI adoption and growing concerns about the AI infrastructure complex has increased recent investor focus on the next beneficiaries of the ever-expanding AI trade," Hammond wrote.
The mechanism here is straightforward: when the growth rate of the addressable market slows, investors stop rewarding capacity builders and start rewarding revenue converters. That is a rotation within the trade, not an exit from it. The dispersion itself is evidence that the market is doing exactly what a maturing cycle should do — discriminating between winners and laggards instead of lifting everything together.
Why This Is Cyclical, Not Structural
The central judgment of the current market is that the AI buildout is a cyclical fluctuation superimposed on a structural shift — and the two are easily confused because they point in opposite directions at the same time. Getting this distinction wrong flips the conclusion, so it is worth separating them explicitly.
On the cyclical side, the evidence is the deceleration itself. The 75% to 49% to 25% path is a classic second-derivative pattern: after a period of explosive growth, the rate of growth normalizes. That normalization produces volatility, multiple compression, and sector rotation — all of which investors are seeing now. It is mean-reverting in the sense that growth rates cannot compound at 75% forever; they revert toward the underlying demand growth rate. Three historical cycles support the pattern: the late-1990s telecom buildout, the mid-2000s semiconductor supercycle, and the 2020 cloud migration wave all showed the same sequence — a spike in capital intensity, a growth-rate normalization, and then a long plateau of elevated but slower absolute spending.
On the structural side, the evidence is that the absolute level of spending remains far below historical technology-investment peaks. AI capex has recently equated to 0.8% of GDP, compared with peak levels reaching 1.5% of GDP or greater during other technology booms of the past 150 years. To match the peak of spending during the late-1990s telecom investment cycle, AI hyperscaler capex would need to reach $700 billion in 2026 — a level that company guidance is now approaching, and one that several analysts expect to be exceeded.
Goldman Sachs Research's broader estimate, which augments the hyperscaler figure to capture global AI investment, puts 2026 AI capex at 1.8% of US GDP (0.9% of global GDP), rising to 2.5% of US GDP in 2027 and 2.8% in 2028. "These levels are consistent with the 2%-5% of GDP peak investment impulses observed in prior general-purpose technology buildouts," Goldman Sachs economist Jim Briggs wrote. "And while our US portfolio strategy team has flagged that consensus capex projections for 2027 are likely too conservative, even significant upward revisions would leave the level of AI investment as a share of GDP comfortably within the historical range observed in prior technology cycles."
That comparison to prior general-purpose technologies — railroads, electrification, the internet — is the core of the structural argument. These buildouts take a decade or more, and the steepest spending comes in the middle years, not the first. A slowdown in the growth rate in year three of the cycle is what the pattern predicts, not evidence that the pattern has broken. The dot-com and telecom booms each lasted years beyond their first growth-rate inflection.
There is also a leading-indicator check. Goldman Sachs economists compiled a dashboard of semiconductor manufacturing equipment imports in Taiwan and South Korea, relevant Purchasing Managers' Index indicators, import prices, and memory purchase and GPU rental prices. "The good news for the capex outlook is that all leading indicators rank near the top end of their range since 2022," Briggs wrote. "This pattern suggests a robust near-term growth outlook." GPU rental prices are particularly telling: if demand were actually rolling over, rental rates would be the first thing to crack, and they are not.
Finally, the demand-side driver that BMO's Lee emphasized remains intact: labor shortages and the push for automation keep AI's longer-term economic expansion on track even if the pace of infrastructure investment moderates. Concerns about jobs, data centers and profitability may temper the speed of the buildout, but they do not remove the economic incentive to automate. A cyclical slowdown in spending growth does not erase a structural shortage of labor.
The Counter-Thesis: What If This Time the Boom Breaks?
The strongest case against the "temporary step" view comes from the bubble camp, and it should not be waved away. Jeremy Grantham, the GMO co-founder, has argued that AI "meets every condition of the railroads and the internet. It's a powerful idea that's attracted everybody's money. No one has any doubts." His warning is rooted in a simple arithmetic: "the higher the market, the lower the returns will be," because doubling the price of an asset halves the forward return.
The bear case has three pillars. First, the spending is enormous: Amazon, Alphabet, Meta and Microsoft have collectively earmarked roughly $725 billion in capital expenditures this year, about 2% of US GDP, according to analysis of company statements. Second, much of that spending is funded through debt, which raises the hurdle rate for returns — and several hyperscalers reported free-cash-flow compression in their most recent quarters as capital spending outpaced operating cash flow. Third, the payoff is uncertain: some investors worry that massive spending to build AI data centers may take too long to pay off.
GMO's own February 2026 paper, "Sink or Swim," examined whether capex booms "foretell wise investments or warn of over-optimism" and notably stopped short of a clean answer. That ambiguity is itself a warning: history does not guarantee that this boom ends well, and every previous general-purpose-technology buildout eventually produced a bust in the equities that financed it, even when the technology itself changed the world.
The bull response is that the comparison to the dot-com bubble fails on the funding side. The 1990s telecom boom was financed by speculative, unprofitable companies burning cash with no revenue. Today's hyperscalers are among the most profitable companies in history, with combined quarterly revenue of approximately $430.6 billion in the first quarter of 2026 and cloud segments growing roughly 35% year over year. In the second quarter of 2026, Alphabet's Google Cloud revenue jumped 82% to $24.8 billion, Amazon Web Services grew 37% — its fastest pace in 18 quarters — and Microsoft's Intelligent Cloud segment grew 32%. The capacity is being built by companies that already have paying customers and deep balance sheets.
The real vulnerability is not demand for compute but the timing of monetization. The software and services companies that are supposed to be the "AI Productivity Beneficiaries" have struggled because they have yet to realize AI-enabled revenue growth. Meanwhile, a group of "AI Platform" stocks — providers of database and development tools — has outperformed as corporate AI adoption increases, suggesting the market is already rotating toward the layer where revenue is showing up first. If AI-enabled revenue at the productivity layer fails to materialize while capex stays above $700 billion annually through 2027, the bulls' structural argument loses its anchor.
So the falsifying signal is specific: if hyperscaler capital expenditure turns negative in absolute terms for two consecutive quarters, or if GPU rental prices fall below the low end of their 2022–2026 range — signaling idle capacity rather than tight supply — the "temporary step" thesis breaks. A third trigger would be a sustained failure of AI-enabled revenue to accelerate at software and services companies while spending remains elevated. Any one of those would shift the debate from "pause" to "peak."
The Second-Order Question the Market Is Not Asking
The first-order effect of an AI capex slowdown is obvious: chipmakers and data-center equipment suppliers see order growth decelerate. The second-order effect is where the real risk — and opportunity — lies, and it runs through three channels.
First, the funding channel. Hyperscalers finance capex through operating cash flow, debt issuance and equity. If capex growth slows because cash flow cannot keep up, the same companies that have been net buyers of credit may become heavier issuers, putting upward pressure on investment-grade spreads and, indirectly, on Treasury yields if the supply overhang grows. A capex slowdown that is driven by funding stress is more dangerous than one driven by deliberate pacing, because it transmits into the bond market.
Second, the earnings-revision channel. The 44% rally in infrastructure stocks against a 9% earnings upgrade is sustainable only if revenue conversion accelerates. If it does not, the multiple compresses even if the absolute capex number keeps rising. This is why the correlation breakdown from 80% to 20% matters more than the headline spending figure: it shows investors are already discounting the earnings-revision risk stock by stock.
Third, the cross-industry channel. AI infrastructure spending has pulled forward demand for power, utilities, copper and data-center real estate. A slower growth rate in capex does not erase that demand, but it changes the timing — and timing is what leveraged projects care about. Projects financed on the assumption of perpetual 75% growth face refinancing risk if the growth rate settles at 25%.
The market has not fully priced the funding channel. That is the gap between the conventional read — "capex is slowing" — and the second-order read: "the composition of capex funding is about to change, and that changes the risk for credit markets." It is also why the pause could feel sharper than the headline numbers suggest.
What It Means for Investors
If the base case holds — a cyclical deceleration inside a structural expansion — the market impact is uneven rather than uniform, and it differs sharply across time horizons.
In the short term, sentiment and liquidity dominate. Volatility will stay elevated as investors digest each quarterly capex print and each data point on GPU pricing. The dispersion that has already opened up — the drop in hyperscaler stock correlation from 80% to 20% — is likely to widen, not close. Stock picking replaces index buying, and the easiest money in the infrastructure layer has already been made.
In the medium term, fundamentals take over. The beneficiaries are companies that can demonstrate a direct link between AI spending and revenue: the largest cloud platform operators, the AI Platform layer of database and development tools, and eventually the software and services firms that monetize AI productivity. The exposed are infrastructure companies where operating earnings growth is under pressure and capex is debt-funded, plus any company whose valuation was bid up on capex growth alone rather than on earnings. Credit investors should watch issuance volumes from the hyperscalers more closely than the capex headlines.
In the long term, the structural view implies that AI investment momentum should persist. Goldman Sachs Asset Management's third-quarter 2026 outlook states that "AI investment momentum should persist, with scope for upward capex revisions that benefit US and EM equities in particular." JPMorgan's 2026 market outlook is more cautious but points in the same direction: "we believe that fundamental, CapEx, sales growth, earnings, and buybacks will support the AI sector, we can't rule out episodes of volatility or temporary retracement."
Three scenarios frame the path ahead. The base case is a deceleration that stabilizes: capex growth slows to the mid-20% range, revenue conversion catches up, and the market broadens beyond the infrastructure layer into software and services. The upside case is another round of upward revisions, driven by stronger-than-expected cloud revenue and GPU rental prices that stay firm — in that scenario, the 2027 consensus estimates that Goldman Sachs already calls "too conservative" get revised higher again, and the correlation across hyperscalers stabilizes. The downside case is a funding shock: if debt markets tighten or cloud revenue growth disappoints, capex plans get cut in absolute terms, free cash flow does not recover, and the valuation gap between the 44% stock rally and the 9% earnings revision closes from the price side.
The market's current message is clear: it is pricing a pause, not an end. The risk is that a pause, if it lasts long enough, becomes something else. For now, the leading indicators, the GDP comparisons, and the profitability of the spenders all point the same way — this is a step, and the staircase is still going up. But the second-order funding channel is the watch item: if capex slows because the money runs short rather than because the builders are pacing themselves, the pause becomes a peak.
Data as of September 13, 2026. Capex figures are from company disclosures and consensus estimates through the third-quarter 2026 earnings season.
Explore more exclusive insights at nextfin.ai.
