Written By: Stephen Swensen
Date: June 30, 2026
Subject: Can AI Justify Higher Equity Market Valuations? Separating Cash Flow from Narrative in a Concentrated Market
In an earlier white paper, Beyond the Traditional Playbook: Adapting to the New Economic Game, we argued that the valuation anchors investors grew up with, in particular the reflex that the market is “expensive” above some fixed multiple of earnings, were becoming less reliable as the economy’s center of gravity shifted toward software, platforms, and intangible capital. That argument has since collided with the most expensive question in markets: whether artificial intelligence justifies paying more for stocks than history says we should.
The numbers make the question concrete. As of late April 2026, the S&P 500 traded at roughly 21x forward earnings (and, depending on the measure, in the mid-to-high 20s on trailing earnings), near double the practitioner shorthand of “15x” that anchored a generation of allocation decisions. The cyclically adjusted P/E sat around 41, within striking distance of its all-time record of 44 set in December 1999. The index’s top 10 names made up close to 40% of its weight, the most concentrated the US market has been in modern history, with Nvidia alone the largest single-stock weight on record. And the companies at the center of it are spending at a scale that would have looked reckless a decade ago: the four largest hyperscalers have guided to capital expenditure approaching $700 billion in 2026, up by roughly three-quarters in a single year. The market is being asked to believe that this spending earns back more than it costs—and to pay up today for the belief.
The argument in one breath: This paper takes that question seriously without surrendering to it. The short institutional answer is yes, but only conditionally. A higher valuation regime than the old “15x” benchmark can be justified if three things are simultaneously true: first, that index composition has structurally migrated toward higher-margin, more scalable, intangible-heavy businesses; second, that AI is producing measurable improvements in revenue growth, margins, and incremental return on invested capital; and third, that those gains spread beyond a handful of mega-cap leaders into the broader earnings base. The evidence strongly supports the first condition, partially supports the second, and only begins to show for the third.
And the honest caveat: Calling higher valuations justified is not the same as calling them riskless. The pages that follow weigh the case on its merits: 1) where the productivity gains are real, 2) where the monetization is genuine, 3) where the capital spending may be outrunning the 4) cash flows, and where the market may be paying for a future it has not yet earned.
History is a sequence of regimes, not an average. The most useful reframing in this whole debate is that valuation “history” is not one monolithic average. It is a sequence of regimes, and the multiple a regime can sustain depends on its economic structure, discount rates, payout policy, sector composition, and the durability of growth. Our own Beyond the Traditional Playbook put the point directly: the practitioner benchmark of “approximately 15x” is increasingly incomplete in sectors like technology and AI, where scalability and expected growth justify a different lens. And in Growth vs Value we traced how style leadership persists for long stretches under low inflation, low rates, and accelerating innovation, and how speculative excesses such as the 2000–2002 dot-com bust fall hardest on the most richly valued growth names. Comparing today’s market to a mid-century average without adjustment is too blunt to be useful, because today’s index is not the same economic object. Even long-horizon tools require care: Robert Shiller’s data now include a total-return version of his cyclically adjusted P/E precisely because buybacks have increasingly substituted for dividends, quietly changing what payout-based measures capture.
The current index structure strengthens the case for an upward reset to a “fair” multiple. The S&P 500 covers roughly 80% of US market capitalization, and more than half of it now sits in sectors tied directly to software, semiconductors, digital platforms, cloud, data, and AI deployment. Information Technology alone is about 36% of the index, with Communication Services another 11%. These are businesses that can scale revenue without the physical-capital intensity of prior industrial regimes. A higher “normal” multiple is more plausible for that index than for a market dominated by lower-growth cyclicals.
But composition explains only part of the premium. Concentration is a double-edged sword. The same composition that helps justify a higher multiple also makes the market far more dependent on a narrow leadership cohort. The top 10 names, among them Nvidia, Apple, Microsoft, Amazon, Alphabet, Meta, Broadcom, Tesla, Berkshire Hathaway, and Eli Lilly, now approach 40% of index weight, with the “Magnificent 7” alone around a third. A market this concentrated has little room for innocent mistakes. If AI monetization disappoints, or if the leaders keep spending without converting capex into incremental free cash flow and returns, the same concentration that supported the multiple can compress it just as quickly.
A useful cross-check comes from Morningstar, whose analysts value more than 700 US-listed companies on a bottom-up basis. On that composite, the US market traded at roughly a 5% discount to fair value at the end of April 2026, hardly the picture of a screaming bubble. But the same analysis carries a warning that should sober any bull: strip out Nvidia and the discount nearly vanishes; strip out Nvidia, Alphabet, and Broadcom and the market sits exactly at fair value. In other words, what looks like a modestly cheap market is really a fairly valued market wearing a few very large, very specific bets. That is the headline-multiple problem in miniature: a reported forward P/E near 21x is itself skewed by a cluster of mega-caps trading near 30x, with the median large-cap meaningfully cheaper.
Exhibit 1: Valuation Regimes at a Glance (Illustrative)
Exhibit 2: Market Concentration Over Time (Illustrative)
The test is whether AI changes the cash-flow math. A premium multiple is justified not because a company says “AI” with enough conviction to move the room, but because AI changes one or more of the core drivers of equity value: sales growth, operating margins, capital efficiency, competitive durability, or the duration of excess returns. That is a stricter test than the market often applies, and the evidence for it is strongest where studies examine live workflows rather than survey optimism.
The micro-evidence of real gains exists, and it is sharpest for less-experienced workers. Brynjolfsson, Li, and Raymond, now published in the Quarterly Journal of Economics, found that a generative AI assistant raised customer-support productivity 15% on average, with the largest gains accruing to lessexperienced workers, by diffusing the tacit know-how of the best workers to everyone else. A larger set of randomized trials at Microsoft, Accenture, and a Fortune 100 firm, pooling nearly 4,900 developers, found AI tools increased completed tasks by roughly 26%, again with the biggest gains for the less experienced. Economically, this points to margin and labor-productivity gains that can spread across firms even without owning a frontier model, especially in service industries with repeatable workflows.
Company disclosures show the bull case more concretely, and the most recent quarter is striking. Microsoft reported its AI business had surpassed a $37 billion annual revenue run rate, up 123% year over year, with Azure growing about 40% and more than 20 million paid Copilot seats. Alphabet’s Google Cloud grew 63% to $20 billion with operating margins near 33% and a backlog above $460 billion, and the company has since moved to raise roughly $80 billion of fresh equity to fund the buildout. Nvidia’s data-center revenue reached about $75 billion in a single quarter, up more than 90%, at gross margins near 75%. Amazon’s AWS reaccelerated to 28% rowth at roughly 38% operating margins, and Meta’s revenue grew 33% even as it guided 2026 capital spending up toward $125–145 billion. When revenue acceleration, backlog, and margin expansion move together, the market is rational to pay a premium relative to businesses still describing an AI future rather than reporting one.
It is worth pausing on a counterintuitive institutional read. Despite the run, Morningstar’s analysts assign all five of these companies a wide economic moat, the highest durability rating they offer, and currently judge every one of them undervalued relative to their fair value estimates, with Microsoft and Meta trading at the steepest discounts. That is not a green light; Morningstar attaches a “very high” uncertainty rating to Nvidia precisely because those fair values embed aggressive growth assumptions that may not hold. But it reframes the question. The debate is not simply “are these stocks too expensive?” It is “are their moats and growth durable enough to make even today’s prices reasonable?” That is exactly the underwriting question a disciplined investor should be asking.
Breadth is the swing factor, and here, the latest earnings season runs counter to the simple “narrow market” story. In the first quarter of 2026, the Magnificent 7 grew earnings by 63%, their fastest since 2021, while the other 493 companies also grew earnings by 17%, their strongest showing since late 2021. That is real breadth, and it is encouraging for the thesis. Two caveats keep it from settling the question. First, a meaningful part of the mega-cap number was non-operating: Alphabet’s reported earnings included a $37.7 billion gain on equity holdings, Amazon’s $16.8 billion from its Anthropic stake, and Meta’s an $8 billion tax benefit, all items that flatter headline growth without reflecting dayto-day operations. Second, the index’s earnings engine remains concentrated: four of the Magnificent 7 (Nvidia, Alphabet, Amazon, and Meta) are among the top five contributors to S&P 500 growth. Breadth is improving; it is not yet decisive.
Exhibit 3: The AI Leaders: Monetization and Morningstar's Read (Illustrative; ratings are Morningstar opinions, not Atlas recommendations)
Capability is not the same as realized productivity. The strongest counterargument is that markets may be treating capability as if it were already operationalized productivity. That leap is too fast. METR’s randomized study of experienced open-source developers found that early-2025 AI tools made them 19% slower, even though the same developers expected to be roughly 20% faster, a 40-point gap between perceived and measured productivity. The result should make investors wary of sweeping “AI equals immediate margin expansion” assumptions in complex, accuracy-sensitive work, and wary of survey-based productivity claims, which consistently run well above what controlled trials measure.
Step back from individual firms and the aggregate picture is more sobering still. Despite near-universal experimentation, McKinsey finds that only about 6% of companies report enterprise-wide earnings impact from AI, and a widely cited (if methodologically contested) MIT study puts the share of enterprise generative-AI pilots with no measurable profit-and-loss impact at as high as 95%. The macro data agree: US total factor productivity, the part of growth that reflects doing more with the same inputs, grew under 1% in 2025, well short of what a genuine productivity boom would show, and even bullish sell-side estimates credit AI with adding only a fraction of a percent to S&P 500 earnings through productivity in 2026. The market story so far is overwhelmingly a capital-spending story, not a realized-efficiency story. The productivity payoff is forecast—not yet measured.
That capital-spending story has become enormous, and it is where the clearest risk lives. The four largest hyperscalers are on track to spend approaching $700 billion in 2026, and Goldman Sachs estimates cumulative hyperscaler capex of roughly $1.15 trillion across 2025–2027, more than double the prior three years. Bain & Company’s 2025 technology report sharpened the worry into a number: meeting projected AI demand could require about $2 trillion of new annual revenue by 2030, leaving an estimated $800 billion annual shortfall even after counting AI-driven savings, the institutional successor to Sequoia’s earlier “$600 billion question.” The concern is not that the spending is irrational, but that the revenue needed to justify it has not yet shown up at the necessary scale.
The financing of that spending is the genuinely newer risk relative to a year ago. Bank of America estimates that in 2026 the largest spenders will commit close to 90% of their operating cash flow to capital expenditure, against a long-run average nearer 40%, pushing several of these companies toward negative free cash flow and into the debt and equity markets. We are seeing large bond deals, offbalance-sheet financing vehicles for data centers, and circular arrangements in which chipmakers invest in the model labs that then buy their chips. Investors including Michael Burry have argued that depreciating GPUs over five to six years, when their cutting-edge economic life may be closer to two or three, flatters reported profits; Goldman, more neutrally, calls the useful life of silicon the single most influential variable in the whole build-out. None of this is proof of a bubble. It is evidence that the cushion of self-funding that distinguished this cycle from past manias is thinner than the bull case assumes.
Exhibit 4: AI Capex Versus Operating Cash Flow (Illustrative)
Concentration is also a liquidity risk. This is also where our own prior work connects. In The Importance of Liquidity and its sequel, The Illusion of Liquidity, we argued that liquidity is most abundant exactly when you least need it and evaporates when you most do. A market in which 40% of the index sits in ten crowded names is a liquidity argument as much as a valuation one: if sentiment turns, everyone reaches for the same exit at once, and the apparent depth of these mega-caps can prove shallower than it looks. Concentration and crowding convert a valuation question into a drawdown question.
It is, therefore, no surprise that official-sector voices have grown cautious. In late 2025, both the Bank of England and the IMF warned that equity valuations, on some measures, were comparable to the dot-com peak and that the risk of a sharp correction had risen. The honest comparison cuts both ways. On one hand, concentration today exceeds 2000 and CAPE is near its record. On the other hand, the leaders are profitable cash generators, not the cash-burning concept stocks of 1999; today’s forward multiple of around 21x sits below the roughly 25x the index carried at the 2000 peak, and the leaders trade near 30x earnings rather than the triple-digit multiples investors paid for the dot-com darlings. The most useful framing is not “bubble or not” but “overbuild-and-timing risk”: even a correct long-term thesis can suffer an air pocket if capacity and capital arrive years before revenue does, as they did for telecom fiber in 2000–2001.
Finally, the labor and measurement picture warrants humility. Stanford researchers found that workers aged 22–25 in the most AI-exposed jobs have seen roughly a 16% relative decline in employment since late 2022, even as older workers in the same fields kept growing, a real and concentrated displacement that has not yet shown up in aggregate statistics. The IMF estimates AI could affect 40% of jobs globally and 60% in advanced economies, with only part of that exposure enhancing productivity. A world where AI complements capital more than labor can be bullish for some equities and destabilizing for the broader economy at the same time, and markets have historically climbed through exactly that tension.
Exhibit 5: Weighing the Evidence (Illustrative Synthesis)
Exhibit 6: The Capex Cycle: Scale Versus Cash Flow (Illustrative)
The point is to sharpen the tools, not abandon them. If AI is going to change how we value markets, it should change how we apply valuation discipline, not retire it. Four recalibrations follow from the evidence above.
First, move from static P/E to evidence-adjusted growth. A price-to-earnings-to-growth lens is more informative than a static multiple in technology-heavy markets, but the growth input should be discounted by evidence quality. Growth supported by current monetization, backlog, and conversion deserves more credit than growth supported only by management aspiration. A 25x multiple on a business compounding earnings 25% with visible AI monetization is a different animal from 25x on a business growing 25% through capex-fueled hope and adjusted-EPS engineering.
Second, focus on free cash flow yield after recurring AI investment. With capex now consuming most of the leaders’ operating cash flow, the valuation conversation has to separate reported earnings growth from owner earnings after normalized AI capex, stock-based compensation, and maintenance cloud costs. Rich multiples can be justified by powerful businesses; they become harder to justify when cash conversion deteriorates because the monetization treadmill runs behind the spending treadmill. This is the single discipline the current cycle most demands.
Third, emphasize incremental return on invested capital, not just consolidated margins. AI justifies a premium only if the next dollar spent on models, chips, and distribution continues to earn an attractive spread over the cost of capital. This is where it pays to distinguish AI enablers and disciplined adopters, both of which can deserve higher multiples, from AI storytellers, which warrant skepticism. It is also why the wide-moat framing matters: a durable moat is precisely the structural feature that lets incremental returns persist long enough to justify paying up today.
Fourth, judge the index by breadth, not just its weighted-average multiple. A fair-value multiple for the S&P 500 should be assessed against whether earnings-revision breadth, margin durability, and return persistence are spreading into the other 493 names. If the multiple stays elevated while the earnings base grows less dependent on a narrow cluster, the case for a structurally higher regime strengthens. If the market keeps paying up for the same handful of names while the rest of the index drifts, that multiple is fragile—the difference between a healthier market structure and an expensive popularity contest.
Exhibit 7: Recalibrating the Valuation Toolkit (Illustrative)
We think in layers, not in a single bet. Our posture is to underwrite the AI thesis the way we would underwrite any source of premium valuation, by asking what would have to be true and then watching whether it is. It is also how we approach individual names: estimating intrinsic value from the bottom up and comparing it to the market price before paying for a story, the same discipline that has led us, at times, to judge a popular AI stock worth a fraction of where it traded. That leads us to think in layers rather than in a single bet. The first layer is AI infrastructure and platform leaders with visible monetization, wide moats, and proven scale. The second is higher-return adopters outside the obvious mega-cap cohort: firms in software, industrial automation, health care, and financial services that can use AI to raise throughput or pricing without taking on frontier-model risk. The third is broad-market breadth exposure, which becomes more attractive only if earnings diffusion continues into the rest of the index.
It is worth noting that this is close to where independent institutional research is also landing. Morningstar’s 2026 outlook explicitly recommends a barbell, high-quality value on one end and select quality growth on the other, as the way to stay invested while respecting both concentration risk and the chance that leadership broadens. We arrived at a similar structure from our own analysis, and the convergence is reassuring: a measured stance, neither all-in on AI nor waiting on the sidelines for a reckoning, but a balanced implementation that owns the genuine winners while refusing to let a macro thesis collapse into a single crowded trade.
Concentration risk of this magnitude is also why we have spent so much prior work on downside protection. In our research on convexity and downside-protection overlays, we made the case that the goal is to stay invested through volatility rather than to time an exit, shaping the ride so that a sharp drawdown in a crowded market does not force selling at the worst moment. We will not detail mechanics here, but the principle is directly relevant: when so much of the index depends on so few names, managing the shape of potential losses matters as much as forecasting the direction of the next move.
The discipline lives in the monitoring. Where we think a white paper earns its keep is in monitoring, not forecasting. Rather than predict an outcome, we track the indicators that would confirm or challenge the thesis as it unfolds: concentration, breadth, margin durability, the ratio of capex to cash flow, and the underlying productivity evidence. Tracking them lets the position adjust as the facts do, instead of waiting for a single headline to settle the debate.
Exhibit 8: An AI Valuation Monitoring Dashboard (Illustrative)
The bottom line is not that AI has made valuation irrelevant. The opposite is true: AI makes valuation work harder. It widens the gap between firms with visible AI economics and firms with AI theater, between genuine multiple support and narrative multiple inflation, and between a healthy index-level re-rating and a concentration-driven melt-up. The evidence we have gathered supports a market with a somewhat higher justified baseline multiple than the old “15x” shorthand implied, anchored by real monetization, wide moats, and a genuinely changed index composition. But the non-negotiable conditions are that AI keeps converting from story to cash flow, that the capital spending earns its return rather than outrunning it, and that the gains broaden from a handful of leaders to the wider market.
For allocators, that calls for conviction and humility in the same portfolio: own the infrastructure and the disciplined adopters, respect the concentration by managing the shape of potential losses, watch the breadth and the cash flows that would validate a higher regime, and refuse to let a single crowded trade stand in for a diversified one. Higher valuations can be earned. Whether this market has earned them is a question the cash flows, not the narrative, will ultimately answer—and the time to decide how you will respond is before the answer arrives.
References and Sources for Further Reading
A considered piece deserves a transparent evidence base. Sources are primary where possible, with the as-of date and link; market and company figures are as of the spring 2026 reporting period and remain subject to revision. Estimates, forecasts, and third-party opinions are identified as such.
S&P Dow Jones Indices, S&P 500 Factsheet (as of April 30, 2026). Index-level P/E, top-10 concentration, and sector composition. spglobal.com/spdji.
FactSet Earnings Insight, John Butters (May 8 & May 21, 2026). Forward P/E (~21x) and the realized Q1 2026 split: Magnificent 7 +63.2%, other 493 +17.4%, plus the GAAP one-off items. insight.factset.com.
Morningstar equity research: US market price/fair value and the AI leaders’ moat, fair-value, and star ratings (June 2026). Market at ~5% discount to fair value, concentrated in a few names; wide moats and undervalued ratings across the five leaders. Morningstar opinions, not Atlas recommendations.
Robert Shiller/multpl.com (as of June 11, 2026). CAPE ~41.2 versus the 44.19 record of December 1999. multpl.com/shiller-pe; shillerdata.com.
Microsoft FY26 Q3, Alphabet, Amazon, and Meta Q1 2026 results (SEC EDGAR/IR, April 29, 2026); Nvidia Q1 FY27 (SEC EDGAR, May 20, 2026). Primary disclosures of AI revenue, cloud growth, margins, backlog, and capex guidance. sec.gov.
Alphabet, $80 billion equity capital raise (IR, June 1, 2026). Subsequent to Q1 earnings; includes a $10B Berkshire Hathaway placement, to fund AI infrastructure. abc.xyz/investor.
Brynjolfsson, Li, and Raymond, “Generative AI at Work,” Quarterly Journal of Economics 140(2):889 (2025). Peer-reviewed; +15% average customer-support productivity, largest for lessexperienced workers. academic.oup.com/qje.
Cui, Demirer, Jaffe, et al., randomized trials of AI developer tools (Microsoft/MIT, 2024-25). Pooled RCT (~4,867 developers): +26% completed tasks. SSRN 4945566.
METR, developer productivity RCT (July 2025). Experienced open-source developers 19% slower with early-2025 AI tools despite expecting gains. metr.org; arXiv 2507.09089.
McKinsey, “The State of AI” (2025); MIT Project NANDA, “The GenAI Divide” (2025). Adoption-to-impact gap: ~6% of firms see enterprise-wide EBIT impact; MIT’s ~95%-of-pilots figure is widely cited but methodologically contested (survey, not peer-reviewed).
US Bureau of Labor Statistics, total factor productivity (2025 release). TFP grew under 1% in 2025: no clear AI signal in aggregate productivity yet. bls.gov.
Bain & Company, 6th Global Technology Report (Sept. 23, 2025); Sequoia Capital, “AI’s $600B Question,” David Cahn (June 2024). Forecast of ~$2T new annual revenue needed by 2030 with an ~$800B shortfall; the earlier revenue-gap framing. bain.com; sequoiacap.com.
Goldman Sachs Research (2025); Bank of America, via CNBC (Feb. 6, 2026). Cumulative hyperscaler capex ~$1.15T for 2025-27; 2026 capex near 90% of the largest spenders’ operating cash flow (Big Five basis). Forecasts.
Bank of England Financial Policy Committee (Oct. 2, 2025); IMF Global Financial Stability Report (Oct. 2025). Official-sector warnings that valuations looked, on some measures, comparable to the dot-com peak and correction risk had risen. bankofengland.co.uk; imf.org.
Brynjolfsson, Chandar, and Chen, “Canaries in the Coal Mine,” Stanford Digital Economy Lab (Nov. 2025); IMF Staff Discussion Note SDN/2024/001 (Jan. 2024). ~16% relative employment decline for the youngest workers in exposed jobs; ~40%/60% global/advancedeconomy job exposure. digitaleconomy.stanford.edu; imf.org.
Atlas Investment Management, Beyond the Traditional Playbook (2024) and Growth vs Value. Prior firm research on regime change, market composition, and style leadership that this paper extends.
Atlas Investment Management, The Importance of Liquidity (2024), The Illusion of Liquidity (2026), and our convexity/downside-protection research. The firm’s work connecting concentration and crowding to liquidity risk, and the case for managing the shape of drawdowns.
Disclaimer: This report’s contents reflect the author’s opinions as of the indicated date and are subject to change without notice. Atlas Investment Management has no obligation to update the information provided here. It should not be assumed that past investment performance guarantees future results. Profit potential also entails the risk of loss.
The information presented is believed to be current and is not personalized investment advice. All opinions expressed as of the presentation date may change over time. All investment strategies carry the potential for profit or loss. Asset allocation and diversification cannot guarantee improved returns
or eliminate the risk of investment losses. References to specific securities are for illustration only and are not a recommendation to buy or sell any security. Third-party ratings and estimates (including Morningstar) are the opinions of those providers, not of Atlas Investment Management.
Securities are offered through Purshe Kaplan Sterling Investments, Member FINRA/SIPC.Investment Advisory services are offered through Ignite Planners, LLC. Purshe Kaplan Sterling Investments and Ignite Planners, LLC are not affiliated companies
