S17 A data-driven diagnosis of the AI market
Priced for everything
to go right.
Seven public datasets. One question. By the metrics Wall Street actually tracks, the U.S. market is the second-most expensive on record, eclipsed only by December 1999. The leaders make money this time. The followers tell stories. The page argues both at once.
Scroll through six interactive charts drawn from Robert Shiller, FactSet, FINRA, MSCI, Stanford HAI, METR, and the four hyperscalers' filings. Each chart asks one question. Together they answer a bigger one: is this a bubble, and if so, where? The verdict sits at the bottom.
§ I The question two camps fight over
The bull and the bear are looking at the same chart pack. They disagree about which lines matter.
The bull points at NVIDIA's 215.9 billion dollars in fiscal-2026 revenue, Microsoft's 281.7 billion, Alphabet's 402.8 billion, and Amazon's 716.9 billion. The leaders are profitable. Cloud backlogs sit at 460 billion at Google and 627 billion at Microsoft. METR's capability benchmark keeps doubling every four to seven months. None of this looks like 1999, when the largest internet names had weak revenue quality and negative earnings.
The bear points at the Shiller CAPE at 42, the Buffett indicator at an all-time high of 232 percent of GDP, the top ten stocks at 36 to 40 percent of the S&P 500, and margin debt at a record 3.83 percent of GDP. The bear points at Anthropic going from 61.5 billion to 380 billion in private valuation in under a year. The bear points at hyperscaler 2026 capex of roughly 725 billion against combined OpenAI and Anthropic ARR of 54 billion, and asks who pays.
The honest reading is that both can be right at the same time. The cleanest framing comes from Bain's chairman of global technology, David Crawford, in the September 23, 2025 report: by 2030, hyperscalers need roughly 2 trillion dollars in new annual revenue to profitably fund the data centers they are committing to today. That number is not a bubble. It is the gap that has to close.
S17.1 The CAPE chart
Robert Shiller's cyclically-adjusted price-to-earnings ratio. The S&P 500 price divided by the ten-year average of inflation-adjusted earnings. The long-run median is roughly 16. Above 30 has been rare. Above 40 has happened twice in 145 years.
What this means
The current reading is roughly 42, second only to the December 1999 peak of 44.2. The 1929 peak hit 32.6 before the crash. The CAPE is not a timing tool. It has stayed elevated for years at a stretch. But every prior reading above 30 has eventually mean-reverted, and the only mechanism for that mean-reversion is either falling prices or rising earnings. Today's CAPE says future long-horizon returns are more fragile than recent returns. It does not say when.
S17.2 The concentration
Share of the S&P 500's total market capitalization held by its largest stocks. The chart shows the top ten as one series, then breaks today's Magnificent Seven out individually. Cap-weighted index investors own this composition whether they know it or not.
What this means
The top ten stocks are roughly 38 to 40 percent of the S&P 500 by mid-2026, the highest concentration in at least fifty years. In 2000, the top ten were about 23 to 27 percent. The Magnificent Seven alone, NVIDIA, Apple, Microsoft, Alphabet, Amazon, Meta, and Tesla, are 33 to 35 percent of the index. NVIDIA by itself is roughly 7 percent. Critically, this concentration is not yet matched by earnings concentration. The Mag 7 generate roughly 24 percent of S&P 500 forward revenue and 13 percent of forward earnings. The valuation has moved faster than the earnings have.
S17.3 The capex gap
Hyperscaler capital spending against the revenue actually being earned by the frontier AI labs they are building for. Capex is what Microsoft, Alphabet, Meta, Amazon, and Oracle disclose in filings. Revenue is what OpenAI and Anthropic confirm as run-rate. This is the central bubble question.
What this means
Big 5 hyperscaler capex is going from 256 billion in 2024 to 410 billion in 2025 to roughly 725 to 805 billion in 2026, a near-tripling in two years. Capex intensity is the highest on record at Oracle (~86 percent of revenue), Meta (~54), Microsoft (~47), and Alphabet (~46). On the other side, OpenAI is at about 24 billion ARR and Anthropic crossed 30 billion in April 2026. Even granting a generous 514 billion in total global AI vendor revenue, 2026 hyperscaler capex still exceeds it by roughly 200 billion. Sequoia's David Cahn calls 2026 the moment of truth for capacity validation versus writedowns.
S17.4 The margin debt
FINRA-reported margin debt in customer securities accounts, expressed as a percentage of U.S. GDP. The level is the single most alarming reading in the entire data set. The composition is what keeps it from being a clean replay.
What this means
Margin debt as a percent of GDP hit roughly 3.83 percent in early 2026, edging close to the 3.97 percent 2021 peak and well above the 2.6 percent reading at the 2000 top. The January 2026 dollar level reached 1.28 trillion, the highest on record. By March it was 1.22 trillion. Retail trading volume hit 5.4 trillion dollars in 2025, the most since the FINRA data began. Same-day-expiry SPX options reached 57 percent of total SPX volume in Q3 2025. The plumbing is hot. Survey sentiment is curiously not. That gap is unusual and worth watching.
S17.5 The capability curve
METR measures how long a task a frontier model can complete autonomously. The y-axis is time, log scale. The trend has doubled every four to seven months since 2019. This is what the bull case keeps pointing at, and it has not yet flattened.
What this means
In 2019 the frontier could handle a few seconds of autonomous task work before failing. In May 2026, Claude Mythos shows a 50 percent time horizon of at least sixteen hours and an 80 percent horizon of three hours and six minutes. METR notes that measurements above sixteen hours are at the edge of its task-suite validity. So the line should be read with caution. But it has not bent down. This is the single strongest argument against a clean dot-com analogy. In 1999, the underlying technology was already on the diffusion curve. In 2026, the curve is still climbing.
S17.6 The bubble lineup
Five historical peaks compared on the indicators that actually distinguish a bubble from an expensive market. Each row is one signal. Each column is one peak. The color is how extreme the reading is relative to history. 2026 is the rightmost column.
What this means
The pattern is consistent across every dimension. By valuation, the 2026 market is closer to 2000 than any other peak. By profitability, it is closer to 2021. By leverage, it has surpassed every prior peak. By sentiment, it is curiously restrained. That mix is not a clean diagnosis. It is exactly what a localized bubble inside a richly valued market looks like.
§ VIII Verdict and receipts
Mixed, leaning bubble in narrow segments.
Broad U.S. equities are richly valued. The AI complex contains both fundamentally justified leaders and clearly bubble-like pockets. Private frontier-model financing is the most euphoric corner. The likely failure mode is multiple compression at the periphery, not a 1929-style system cascade.
- 35% Broad U.S. market already in a full-system bubble
- 65% Localized bubble in second-tier AI software and private frontier-model rounds
- 45% 12 to 24 month AI-led valuation reset of 25%+ without a recession
- 40% Leaders grow into current valuations through earnings over 2 to 4 years
§ IX Methodology & Colophon
Every chart is hand-drawn in SVG by a small vanilla JavaScript module. No frameworks, no build step, no analytics trackers beyond Cloudflare's privacy-respecting page-view counter. Data is baked into the page as static JSON. Hovering and pill filters redraw locally without a network round-trip.
Shiller CAPE from multpl.com / Robert Shiller's data. Buffett indicator from Advisor Perspectives. Top-10 concentration from Slickcharts and MSCI factsheets. Hyperscaler capex from company filings, Statista, CreditSights, and CNBC. AI lab ARR from OpenAI and Anthropic public disclosures, plus Sacra and Stanford HAI 2026 AI Index. Margin debt from FINRA. METR time-horizon from metr.org. All sources are linked in the body and the reading list. Snapshot vintage: May 12, 2026.
multpl · Shiller CAPE history ↗
Current Market Valuation · Buffett indicator ↗
Stanford HAI · 2026 AI Index ↗
METR · Time-horizon paper ↗
FINRA · Margin debt statistics ↗
Bain · Global Technology Report 2025 ↗
None of this is investment advice. Several anchor figures are mid-May 2026 estimates and will move. The Buffett indicator has a structural upward drift from globalization of revenue, so the level-comparison to 1929 is imperfect. METR explicitly notes that measurements above sixteen hours are at the edge of measurement validity. Private valuations cited for OpenAI and Anthropic are last-disclosed; private markets reprice between rounds. The 2 trillion dollar gap from Bain is a 2030 projection, not realized data. Past projections of compute demand have been wrong in both directions. The data is the strongest argument here, but the verdict is editorial.