NextFin

AI Boom Has Years to Run Before Bubble Risks Bite, CIO Says

Summarized by NextFin AI
  • Federated Hermes' Steve Chiavarone argues the AI boom is not yet a bubble, as infrastructure buildout is funded by cash-generating incumbents rather than indiscriminate speculative capital.
  • Amazon, Microsoft, Alphabet and Meta plan roughly $725 billion in 2026 capex, up 77% year-over-year, with about three-quarters tied directly to AI infrastructure like GPU clusters and data centers.
  • Valuations remain stretched but below 1999 extremes: the Nasdaq trades near 33x earnings versus over 200x at the dot-com peak, though S&P 500 concentration hit a record 43% for the top 10 stocks.
  • The key risk is a potential capex strike if hyperscaler returns keep compressing from 40% to 25%, which would hit chip suppliers like Nvidia hardest before affecting cloud platforms.

NextFin News - Investors worried that the artificial-intelligence boom is turning into a bubble may be right — just several years too early, according to Steve Chiavarone, chief investment officer for global equities at Federated Hermes. The infrastructure buildout reshaping the technology sector still has years to run, he argues, before the risk of a hangover becomes more pressing than the opportunity in front of it.

The comment lands as the debate over AI valuations has hardened into two camps. On one side stand investors who see echoes of the dot-com mania in the concentration and multiples of the AI leaders. On the other stand strategists who argue that this cycle is financed by cash-generating incumbents committing capital at differentiated rates — the opposite of the indiscriminate funding that defined 1999. Chiavarone sits firmly in the second camp, but with a caveat that matters: the concerns are legitimate, just premature.

The Situation: A Buildout Backed by Cash, Not Hype

Chiavarone's distinction rests on how capital is being allocated. In a bubble, he noted, "everybody gets financed at the cheapest rate." What investors are seeing now is closer to the reverse: markets are being discerning about which companies receive capital and at what cost. That selectivity, in his view, is what separates a rational investment cycle from an irrational one.

The scale of the spending helps explain why the debate has intensified. Amazon, Microsoft, Alphabet and Meta collectively plan roughly $725 billion of capital expenditure in 2026, up 77% from the prior year's record of about $410 billion, according to company guidance compiled from quarterly earnings. Amazon leads at about $200 billion, Microsoft is tracking near $190 billion, Alphabet is targeting $175 billion to $185 billion, and Meta guided to $115 billion to $135 billion in its first-quarter outlook before raising the range toward $125 billion to $145 billion as component pricing climbed. Broadening the group to include Oracle, Goldman Sachs Research estimates the largest U.S. hyperscalers are on track to spend about $800 billion in 2026, with consensus estimates implying $1.1 trillion in 2027. Roughly three-quarters of that spend is tied directly to AI infrastructure — GPU clusters, custom silicon, and data-center construction.

The equity market has rewarded that commitment, but not without volatility. Nvidia, the primary supplier of the chips powering the buildout, trades near $216 a share with a market capitalization of about $5.2 trillion, down from a 52-week high of $243.37. Microsoft, Alphabet and Meta each fell between 1% and 2% in recent sessions, while the S&P 500 holds near record territory. The concentration that has accompanied the rally is the statistic that keeps bubble-watchers awake: the 10 largest stocks now account for roughly 43% of the index's market capitalization, a record, compared with a historical average of 18% to 23% across the 1990–2015 period. Nvidia alone accounts for about 7.4% of the index.

"Unlike the bubble days of the late '90s, we think that the exuberance, this far, is rational," Chiavarone said.

The question is how long that rationality can hold as the spending bill grows faster than the revenue it produces.

Why This Cycle Is Not 1999 — Yet

The most common bubble comparison is also the easiest to overstate. At the peak of the dot-com mania in late 1999, the Nasdaq Composite reached a price-to-earnings ratio of more than 200, according to historical market data. The AI-era average for the same index has run closer to 33. The leading technology stocks of the dot-com era commanded forward earnings multiples near 70 times; today's hyperscalers trade closer to 26 times forward earnings.

The difference is not an accident of accounting. It reflects who is spending. In 2000, capital flowed to companies with .com suffixes and little or no revenue, funded by venture capital and convertible debt that would vanish in the crash. Today it flows to four companies that each generate tens of billions in annual free cash flow and that are funding the buildout largely from internal resources. Cisco Systems, the infrastructure beneficiary of the internet buildout, briefly became the world's most valuable company on a trailing earnings multiple that reached as high as 472 times by some measures and roughly 30 times sales. Nvidia's valuation, by contrast, sits near 56 times trailing earnings and about 23 times forward earnings, with the company expected to deliver roughly $410 billion in revenue in the coming fiscal year.

That does not make the current market cheap. The Shiller cyclically adjusted price-to-earnings ratio for the S&P 500 has hovered around 41 in 2026 — the second-highest reading in 125 years of U.S. market data, behind only the dot-com peak near 44.2 in December 1999. The historical median is 17. Valuation, in other words, is stretched by any long-run measure; it is simply not stretched to 1999 extremes.

The deeper distinction is behavioral. Bubbles are defined less by high prices than by the collapse of skepticism. Chiavarone's point about capital being rationed at differentiated rates is a claim that skepticism is still functioning. Investors are still punishing companies that spend without a credible path to returns — a discipline that was largely absent in the final year of the dot-com boom. When capital is scarce and priced according to risk, bad projects die before they become systemic.

The Transmission Mechanism: How a Capex Strike Would Travel

Understanding where the risk actually sits requires following the money through the chain. The AI buildout has three layers: the chip and equipment suppliers at the front end, the hyperscalers and cloud platforms in the middle, and the application layer at the back end. So far, the financial returns have been concentrated at the front end. Nvidia's market capitalization of $5.2 trillion reflects the market's conviction that chip demand will keep rising.

The middle layer is where the arithmetic gets difficult. Hyperscaler capital spending now exceeds cash flow from operations, which means large further growth will increasingly need funding via debt and equity issuance, according to Goldman Sachs Research. That is the first structural change worth watching: for the past decade, the mega-cap technology complex has been a net buyer of its own shares. A sustained shift to net issuance to fund AI infrastructure would mark a regime change in how these companies allocate capital — and it would put the cost of capital, rather than internal cash generation, at the center of the investment case.

The second-order risk is not that AI fails. It is that the hyperscalers, faced with compressing returns, eventually slow the pace of spending — a capex strike that would hit the semiconductor and equipment complex long before it hits the cloud platforms themselves. Realized return on invested capital for the hyperscalers has declined from about 40% to 25%, a sign that each incremental dollar of capex is producing less than the last. Capital intensity has climbed to levels that would have been unthinkable a decade ago — as high as 57% of revenue for Oracle and 45% for Microsoft in recent quarters.

Nvidia's forward price-to-earnings multiple of about 23 times and a consensus analyst price target of roughly $339 imply that the market is still pricing robust growth in chip demand. If hyperscaler budgets flatten, that multiple has further to fall than the platforms' own earnings. The platforms can lean on their existing cash flows and diversified revenue; the pure-play suppliers cannot. That asymmetry is the real fragility embedded in the trade, and it is not visible in the headline index level.

The Math That Could Break the Thesis

The strongest argument against the "years to run" view is not valuation; it is arithmetic. At current spending rates, the hyperscalers need to generate roughly $300 billion in annual AI revenue within the next few years simply to break even on their capital investments, according to research published by Goldman Sachs in September. Current AI revenues remain below that threshold.

There is, however, momentum on the revenue side. Cloud revenue growth for Amazon, Alphabet, Microsoft and Oracle accelerated from 25% in 2024 to 48% in the second quarter of 2026. In dollar terms, annualized cloud revenues running above the pre-AI growth trend reached nearly $70 billion in the second quarter and have been roughly doubling over the past several quarters. The three major U.S. public-cloud companies have also announced a combined revenue backlog of about $1.7 trillion, which provides visibility into their ability to absorb the capacity they are building.

That backlog is the bull case's strongest piece of evidence — and its weakest. Backlogs are not revenue until they are recognized, and in a fast-moving technology cycle, committed capacity can become stranded capacity if a newer generation of chips renders today's GPU clusters uneconomic before they are fully depreciated. The useful life of an AI server is measured in years, not decades. If the depreciation schedule outruns the monetization, the write-downs will arrive well before the bubble narrative does.

The Counter-Thesis: Grantham's Bubble Warning

No analysis of the AI trade is complete without the bear case, and few bears are as experienced as Jeremy Grantham of GMO. In a January 2026 paper, Grantham 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." He has warned that declines of as much as 70% in some high-flying stocks cannot be ruled out if sentiment weakens and valuations retreat toward historical norms.

Grantham's central point is about starting valuations, not technology. "The higher the market, the lower the returns will be," he has said, because "if you double the price of an asset you halve the return from holding it from there on." On that logic, the AI revolution can be genuine and the stocks can still be terrible investments from current levels.

The concentration data gives the bears their strongest footing. The top 10 stocks' 43% share of the S&P 500 exceeds the roughly 27% peak weight recorded in March 2000, even though today's leaders trade at lower multiples. That combination — record concentration with still-elevated valuations — is precisely the setup that has preceded the worst drawdowns in market history.

The rebuttal rests on earnings. In 2000, the leading technology companies could not grow into their multiples; many never generated meaningful profits. Today's AI leaders are printing record earnings and free cash flow while funding the buildout. A bubble financed by profits is a different animal from one financed by speculation — but it is not immune to a repricing if growth merely slows rather than accelerates.

What to Watch: The Signal That Would Prove the Bull Case Wrong

The bull case for "years to run" depends on one assumption: that capital continues to flow into AI infrastructure at or near current rates. The single most informative signal would be a sustained deceleration in hyperscaler capital expenditure — specifically, two consecutive quarters of year-over-year capex declines among the four largest platforms. That would indicate management teams are seeing returns deteriorate faster than revenue is growing.

A secondary signal sits on the revenue side: cloud growth falling back below 30% year over year while capex remains above $600 billion annually. That divergence — spending staying high while the revenue engine slows — is the classic setup for a capex strike and the margin compression that follows it.

For investors, the practical implication is a split by time horizon. In the short term, sentiment and positioning dominate, and the record concentration leaves the index vulnerable to a sharp but shallow pullback if any of the mega-cap leaders disappoint. Over the medium term, earnings growth and cloud-revenue acceleration are the variables that matter; so far they are holding. Over the long term, the structural question is whether AI produces productivity gains large enough to justify a permanently higher level of capital intensity in the technology sector. On that question, the evidence is still being written.

The base case is that the buildout continues through 2027 and beyond, with spending supported by backlogs and accelerating cloud revenue, but with equity returns for the builders moderating as capital intensity rises. The upside case is that AI monetization accelerates faster than capex, lifting margins across the software and cloud complex. The downside case is a capex strike triggered by compressing returns, which would hit chipmakers and equipment suppliers hardest.

The AI boom may indeed have years to run. But the investors who make the most money from a structural shift are rarely the ones who pay the most for it — and at 43% concentration and a 41 CAPE, the price of admission to this one is already well above historical norms.

Explore more exclusive insights at nextfin.ai.

Insights

Why isn't the AI boom a bubble yet?

How does AI capex compare to dot-com?

What defines rational investment cycle?

What hyperscaler spend is planned 2026?

Who leads AI infrastructure spending?

What is Nvidia's current market cap?

Why does stock concentration pose risk?

How does Nasdaq PE compare to 1999?

What signals a capex strike risk?

What AI revenue breaks even on capex?

What is Grantham AI bubble warning?

How long can AI rationality hold?

What risks chip suppliers first?

Why is capital rationing key now?

What is the Shiller CAPE ratio now?

How does cloud revenue growth trend?

What is the AI server useful life?

Who funds the AI buildout today?

What breaks the AI bull case thesis?

Is AI monetization faster than capex?

Search, ask or analyze anything...