By any historical measure, the generative AI trade has become the most dominant investment theme the U.S. equity market has ever seen. And that dominance is not subtle.
As Michael Cembalest writes at the outset of J.P. Morgan Asset & Wealth Management’s 2026 Eye on the Market Outlook1, “65%–75% of S&P 500 returns, profits and capital spending since the launch of ChatGPT in 2022 have been derived from 42 companies linked to generative AI.” Without them, “the S&P 500 would have underperformed Europe, Japan and China.”
This is not just market leadership. It is market smothering.
Cembalest frames the moment precisely: after corrections, investors ask what could go right, but when markets are “highly concentrated and near all-time highs,” the correct question becomes “what could go wrong?”
This year’s Outlook is structured around that question.
The AI Moat: Enormous, Narrow, and Historically Unprecedented
The foundation of the AI moat is not abstract. It is industrial, physical, and highly specific:
- NVIDIA chip designs
- TSMC manufacturing
- ASML extreme ultraviolet lithography
Cembalest describes this stack as the literal wall depicted on the cover. The market capitalization of just four hyperscalers and four semiconductor ecosystem firms has surged from $3 trillion to $18 trillion in seven years. Together, they now represent roughly 20% of developed-market equities and 16% of global equities.
The U.S. has so far imported this revolution. Tariff exemptions have insulated the semiconductor ecosystem, and while a Section 232 investigation looms, Cembalest notes that exemptions may persist for firms committing to U.S. production.
The result is an AI complex that has grown faster, and with greater macroeconomic impact, than any prior infrastructure cycle. Tech capital spending in 2025 rivals, as a share of GDP, the Manhattan Project, the Apollo Program, and the Interstate Highway System combined.
Capital Spending Without Precedent, and (So Far) Without Leverage
One of the report’s most underappreciated insights is that this capital spending boom has not yet relied heavily on debt.
Meta’s capex and R&D now consume 70% of revenue, compared with 10% for the median S&P 500 company. Yet most hyperscalers entered this cycle with excess cash, not leverage. As Cembalest notes, “unlike capital spending booms of the past… the latest one has been mostly financed via internally generated cash flow… until very recently.”
That qualifier, until very recently, matters.
The First Cracks: Oracle, Meta, and Financial Engineering
Two firms mark an inflection point.
Oracle has committed to provide OpenAI with $60 billion per year in cloud capacity it has not yet built, requiring 4.5 gigawatts of power, equivalent to “2.25 Hoover Dams or four nuclear plants.” With weaker free cash flow than hyperscalers, Oracle has turned aggressively to debt markets.
Meta’s Hyperion data-center project goes further. Through a joint-venture SPV with Blue Owl, Meta effectively finances data centers via off-balance-sheet project debt. Cembalest cuts through the accounting treatment bluntly:
“Let’s focus on real economics rather than financial engineering and consolidate the SPV debt.”
On that basis, Meta’s net debt profile looks very different than it did a year ago, still manageable, but no longer trivial.
Credit markets are already responding. Since Cembalest highlighted Oracle’s financing challenges in September, “its stock is down ~35% and its credit spreads have risen by 90 bps.”
Valuations: High, But Internally Consistent
Cembalest pushes back on the idea that valuations alone doom the AI trade.
The market is paying more for tech because tech earns more. Higher profitability commands higher multiples, across sectors and within individual stocks. “Just looking at higher valuations in isolation misses the extraordinarily high profit margins of the tech sector vs the rest of the market.”
He reinforces the point with several anchors:
- Global tech PEG ratios are not meaningfully higher than the broader market
- The forward P/E of the S&P 500 rose only ~3% in 2025
- The remainder of returns came from earnings growth, not multiple expansion
The risk, then, is not valuation. It is whether profits materialize at scale.
WCGW #1: A “Metaverse Moment” for Hyperscalers
This is the most immediate threat.
Since Q4 2022, hyperscalers have spent $1.3 trillion on AI-related capex and R&D. Cembalest asks the core question plainly: “Will all this investment eventually result in commensurate profits?” If not, “moat companies could face a version of the Metaverse reckoning that took place in 2022 when the Mag7 stocks fell in half.”
He structures the analysis into six lenses, benchmarks, adoption, profit impact, hyperscaler disclosures, depreciation risk, and OpenAI-driven demand, to separate hype from economics.
AI Capability vs. AI Economics
Benchmark performance has improved dramatically. Models now exceed human performance in coding competitions, image recognition, and medical literature synthesis. Yet hallucination rates remain “shockingly high” in many models, and real-world deployment remains uneven.
Adoption surveys show rapid usage growth, but Cembalest warns explicitly: “These kind of surveys tell us nothing about how much companies or individual users are willing to pay.”
That warning becomes central in the MIT-Wharton divergence.
MIT vs. Wharton: A Chasm in AI ROI
MIT’s 2025 assessment is blunt:
- “95% of organizations are getting zero return” on genAI investments
- CEO confidence in AI strategy collapsed from 82% to 49% in one year
- Successful deployments remain rare and brittle
Wharton, surveying a different cohort, finds much higher reported ROI and aggressive future budgets.
Cembalest’s conclusion is skeptical and disciplined: “That’s the challenge with surveys; they’re not substitutes for tracking actual profit and spending data.”
Hyperscaler Profits: Visible, but Narrow
Only Microsoft discloses AI-specific revenue. Others offer fragments. Free cash flow margins are declining, cash balances are falling, and profitability visibility remains limited.
Cembalest notes that “a clearer path to profitability on AI investments may be needed in 2026 for current valuations to be sustained.”
And that path is clouded further by depreciation risk.
The Quiet Risk: GPU Depreciation Assumptions
Hyperscalers have steadily extended GPU depreciation lives, arguing that older chips retain utility. But rental rates for GPUs have declined 20%–25% in a year, and CoreWeave’s stock has fallen more than 50% from its peak.
Cembalest models a return to three-year depreciation lives. The result:
- 6%–8% EPS hits for most hyperscalers
- 17% EPS decline for Oracle
Not catastrophic, but meaningful.
Why the Capex Juggernaut May Continue
Despite all of this, Cembalest does not dismiss the bull case.
Token consumption is exploding. Google’s monthly token usage rose 100× in fifteen months. A thought experiment suggests that if AI becomes embedded in daily life for billions of users, 23–92 GW of inference capacity would be required, far above today’s AI-capable infrastructure.
If that future materializes, today’s capex may still prove insufficient.
Key Takeaways for Advisors and Investors
- AI concentration is historically extreme, but not yet irrational
- Valuations are supported by profitability, not speculation alone
- The real risk is earnings realization, not adoption headlines
- Debt and financial engineering are creeping in, watch credit markets
- GPU depreciation assumptions matter more than most investors realize
- Infrastructure winners dominate today; downstream beneficiaries lag
- 2026 will test whether AI capex converts into durable free cash flow
Cembalest does not predict collapse. He does something more useful: he defines the fault lines.
As he reminds investors, when a moat grows this large, the correct posture is not awe, but vigilance.
1 Cembalest, Michael. “Smothering Heights. Eye on the Market: Outlook 2026”, J.P. Morgan Asset & Wealth Management, 1 Jan. 2026.
