Loading reports … Fetching the current index of long-form research.
Reports
Loading reports … Fetching the current index of long-form research.
Passive funds hold $19 trillion and 53 percent of all US long-term fund assets.
The S&P 500 is sold as diversification and now carries a concentrated bet on one AI build-out. Earnings built most of that weight; passive flows amplify it at the margin and hold it in place. The harms that can be measured are small. The exposure is large, and the savers who carry it did not choose it. This report measures the weight, the flows that amplify it, the earnings beneath it, the reversal of 2026, and the households holding the position.
The top ten companies in the S&P 500 hold 37.9 percent of index weight, about 40 percent above the dot-com peak and a step below the record 40.7 percent they reached at the end of 2025. The level is half the finding. The other half is what those ten companies have in common.
In 1990 the S&P 500’s top ten companies held about 19 percent of index weight. In 2000 they touched about 27 percent at the dot-com peak and closed the year near 23. By 2015 they were back to about 19 percent, no higher than a quarter-century earlier. At the end of 2025 they held 40.7 percent, a record, and the Q1 2026 reading of 37.9 percent is the first step down from it.1 Those five readings are what the series publishes, and Exhibit 1 draws them without a line between them.
The step down is 2.8 points in one quarter on one series, a move of the size this series makes in ordinary conditions. Part IV reads it against the two comparable episodes on the record. The nearer of them, in 2022, reverted in full.
The dot-com top ten was General Electric, Cisco, Exxon Mobil, Microsoft, Intel, Oracle, IBM, Lucent, Walmart and Citigroup. Six of the ten were technology companies, but they sold routers, chips, software, mainframes and telecom equipment to different buyers on different cycles, and the other four were an industrial conglomerate, an oil major, a retailer and a bank. By free-float weight at March 30, 2026, the current leaders are Nvidia at 7.0 percent, Apple at 6.3, Alphabet (both share classes) at 5.8, Microsoft at 4.6, Amazon at 3.7, Broadcom at 2.4, Meta at 2.4, Tesla at 2.3 and Berkshire Hathaway at 1.8. The tenth slot, on a combined-Alphabet basis, is contested between JPMorgan Chase and Eli Lilly at about 1.4 percent, and is left unnamed.2 The nine named companies hold 36.3 percent between them; the concentration series puts the full top ten at 37.9.
Membership in the AI set is decided by one test: a company is in it if its reported earnings move directly on the build-out of AI compute. Nvidia and Broadcom sell the silicon. Microsoft, Amazon and Alphabet rent the compute. Meta has committed its own capital program to building it. Those six hold 25.9 percent of the index, and they are the lead figure here because the narrow set survives every objection. Apple and Tesla join on weaker ground, since their demand is AI-adjacent and the build-out does not drive their reported earnings; with them, eight names hold 34.5 percent. Both figures appear wherever the claim is about shared exposure.
In 2000 the technology names at the top of the index were exposed to the same bubble, and they still earned on different products. The largest firms today are components of one technology stack: the chip makers sell to the cloud renters, the cloud renters sell to the platforms, and the earnings of all of them rise and fall on the same assumption about when AI pays.
The concentration is held in place by the rules of index construction. The S&P 500 Index Committee updates share counts and float weights quarterly, in March, June, September and December, and adds or removes constituents as needed.3 The quarterly update does not buy winners. A holding that outruns the rest of the index gains value, and so weight, with no trade at all, and the rebalance then adjusts for issuance and buybacks, leaving price drift where it is.
The buying comes from new money. Every contributed dollar is invested at the weights of the day, so the largest constituents take the largest share of every payday. On its own, buying in proportion to weight lifts the whole index and leaves relative weights where they are. It tilts the index toward the largest firms only if their prices respond more to the same flow than the rest do, which is the Jiang, Vayanos and Zheng result read in Part II. That finding carries the loop, and the committee may cap a single constituent's weight only case by case.
Retirement plan design feeds the same loop. The Pension Protection Act of 2006 cleared the way for automatic enrollment and automatic escalation, which direct a rising share of each paycheck into target-date funds that hold the S&P 500 passively. In How America Saves 2024, Vanguard reported that 59 percent of the plans it administers enrolled workers automatically, and that about seven in ten of those plans also raised contributions automatically each year.4
At the level of one participant the arithmetic is small and relentless. A dollar paid into an S&P 500 index fund buys 7.0 cents of Nvidia, 6.3 cents of Apple and 5.8 cents of Alphabet. The same dollar in the equal-weighted version of the same 500 companies buys 0.2 cents of each. The first allocation expresses no view about any of the three. It records what they were worth at the last update, which is the accumulated result of every earlier contribution made on the same rule.
Passive funds crossed active funds in December 2023 and now hold $19 trillion. Peer-reviewed work finds that their flows raise the largest firms’ prices most, amplifying at the margin a weight that earnings built, and that the same flows make it riskier to trade against them.
The concentration built up alongside a change in how Americans invest. In December 2023 index mutual funds and ETFs passed actively managed funds in total assets for the first time, $13.293 trillion to $13.234 trillion on Morningstar's measure.5 The margin has widened since.
At the Investment Company Institute's March 2026 release, index funds held $19.09 trillion against $17.12 trillion for active funds, a 52.7 percent passive share of US long-term fund assets.6 In the decade to 2024, passive US equity funds took in $2.8 trillion of net inflows while active funds lost $3.0 trillion, close to a one-for-one substitution.7
The mechanism was set out in the Review of Financial Studies in December 2025. Hao Jiang of Michigan State, Dimitri Vayanos of the London School of Economics and Lu Zheng of UC Irvine show that inflows to passive funds raise the prices of the economy's largest firms disproportionately, and most of all the large firms the market already overvalues.8 Passive funds hold whatever the index holds, so as their share rises, a smaller share of the market is held by investors who trade on value.
The strength of the result is how conservative its calibration is. The authors' model puts the market's demand elasticity at 3.93, or 4.92 when demand varies by firm, which they place at the high end of the literature: Gabaix and Koijen estimate 0.2 for the aggregate market, and studies of index additions find 0.4 to 4 for single firms. A higher elasticity means prices move less for a given flow, so the model assumes a market about twenty times more elastic than Gabaix and Koijen's, and passive growth still lifts the largest firms most. When new investors enter through index funds in the calibration, the market rises 8.48 percent and the largest size group 11.78 percent, more than any other. In S&P 500 data on quarterly index-fund flows from 1996 to 2020, the largest stocks in the index outperform the index after inflows.
A second effect compounds the first. In the same paper, passive flows raise the idiosyncratic risk of the large firms they push up, which makes betting against them costlier, so the gap between price and fundamental value can widen and persist. Goldman Sachs put passive ownership of the median S&P 500 stock at 26 percent of shares outstanding in November 2024, up from 18 percent two decades earlier, and the average stock at 24 percent.9 That eight-point rise in the median is the input to the one quantified harm this report can attach to the mechanism.
The European Central Bank supplied the coefficient in its November 2024 Financial Stability Review. On euro-area equities from 2010 to 2024, each percentage point of passive ownership raised a stock's correlation with its index by about 0.005.10 Applied to the eight-point rise in the median S&P 500 stock, the implied cumulative effect is about 0.04. The coefficient is euro-area and is carried to US stocks by assumption, and 0.04 is a small number: stocks inside the index move together somewhat more closely than they did when active managers set most prices, and nothing more is claimed for it.
The declared figures understate the passive share. Alex Chinco of Baruch College and Marco Sammon of Harvard Business School, in the Journal of Financial Economics in 2024, use trading volume at index reconstitutions to estimate that passive investors held about 33.5 percent of the US stock market in 2021, roughly double the 16 percent held by declared index mutual funds and ETFs.11 The other 17 points sit in institutional index mandates and in active funds that hold close to index weights. A fund that tracks the S&P 500 inside an active wrapper behaves as a passive fund in price discovery.
Voting power has concentrated with the money. Counting index mutual funds, BlackRock, Vanguard and State Street together vote shares representing about 25 percent of S&P 500 market capitalization.12 Their stewardship teams set the terms of governance engagement for a large share of American public companies, an influence no 401(k) participant delegated by choice.
At the end of 2025 the top ten held 40.7 percent of index market value against about 32 percent of index earnings. In 2015 the two were level. The Q1 2026 earnings surge that looks like vindication shrinks once the one-time items are stripped out.
At the end of 2015 the top ten S&P 500 companies held about 19 percent of index weight and generated about 19 percent of index earnings. At the end of 2025 they held 40.7 percent of weight against about 32 percent of earnings, a gap of 8.7 points.13 The two legs are not on identical footing: the 2025 earnings share is what the ten were expected to generate, and the 2015 share is what they had generated.
One publisher carries both legs at both dates, and what the two vintages establish is the direction. Weight more than doubled, from about 19 percent to 40.7. The earnings share rose too, from about 19 percent to about 32, and the difference between the two rises is the whole of the gap. The width of the gap matters less than its existence, and the section that closes this part sizes how much of it growth explains.
The first quarter of 2026 looked like an answer. In FactSet's May 4 reading, taken while the season was still reporting, Alphabet, Amazon and Meta accounted for about 71 percent of the net dollar increase in S&P 500 earnings.14 The beats were large: Alphabet reported $5.11 a share against a $2.68 estimate, Amazon $2.78 against $1.63, and Meta $10.44 against $6.67. Most of each beat came from items the estimates had not included.
Alphabet's quarter carried $36.9 billion of gains on equity securities, mostly unrealized markups of private stakes including Anthropic, SpaceX and Databricks, worth about $2.35 a share; without them, earnings were about $2.76, a beat of about 3 percent. Amazon's carried a pre-tax gain of about $16.8 billion on its Anthropic stake, about $1.24 a share; without it, earnings were about $1.54, below the estimate. Meta's carried an $8.03 billion tax benefit under Treasury Notice 2026-7, the interim guidance on the corporate minimum tax and previously capitalized research spending; without it, earnings were about $7.31, a beat of about 10 percent.15
Company by company, little of the surprise survives. Of Alphabet's $2.43 beat, 8 cents is left once the gains are removed. Amazon's $1.15 beat becomes a 9-cent miss. Meta keeps 64 cents of a $3.77 beat, the one operating result of the three that cleared its estimate by a clear margin. The 71 percent contribution share stands as published, because a share of the index's dollar increase cannot be adjusted with per-share surprises.
The one-time items are $53.7 billion of pre-tax equity-security gains at Alphabet and Amazon and an $8.0 billion tax benefit at Meta. The Anthropic marks are a product of the same AI investment cycle. Alphabet and Amazon fund a private AI company, the funding round raises its valuation, the stake is marked up on the funders' balance sheets, and the mark-up lands in reported earnings that are then read as evidence that AI is paying. No customer appears anywhere in that sequence. An unrealized gain on a non-marketable security remains a valuation opinion until the position is sold. The $8.0 billion at Meta is a tax item, and the simplest of the three to strip.
Exhibit 4. Stripped of one-time items, the quarter's headline beats reduce to two modest beats and one miss
Q1 2026 · diluted earnings per share · three hyperscalers
| Estimate | Reported | Operating | The one-time item | |
|---|---|---|---|---|
| Alphabet | $2.68 | $5.11 | $2.76 | $36.9B gain on equity securities, mostly markups of private stakes including Anthropic and SpaceX |
| Amazon | $1.63 | $2.78 | $1.54 | $16.8B pre-tax gain on the same Anthropic position |
| Meta | $6.67 | $10.44 | $7.31 | $8.03B tax benefit under Treasury Notice 2026-7 |
Source: Company 8-K and 10-Q filings, Q1 2026; FactSet Insight, May 4, 2026, for the reported and estimated EPS; Treasury Notice 2026-7. Operating EPS excludes the stated one-time item. Against the estimate, operating EPS beats by 3.0 percent at Alphabet and 9.6 percent at Meta and misses by 5.5 percent at Amazon.
This is the bet inside most American retirement accounts. At the end of 2025, 401(k) plans held $10.1 trillion, 57 percent of it in mutual funds. On BrightScope and ICI's plan-level accounting, index funds held 44 percent of 401(k) assets in 2022, and the target-date funds that most plans use as their default hold US equity through broad index funds whose largest holdings are the same ten companies.16
The loop closes on the flow channel. Contributions enter passive funds, which buy in proportion to weight. The largest stocks receive the largest dollar allocation, and because their prices respond most to flow, that allocation lifts them, at the margin, by more than their earnings justify. Index weights tilt further toward them, and the next contribution repeats the cycle. Nothing in the architecture corrects it. Every step runs through the price-impact finding; without it, the loop would lift every stock alike.
Three readings compete with this one, and two of them are strong.
The first is the growth premium. A group of firms growing earnings faster than the index will trade above its current earnings share whoever owns it. The index's arithmetic bounds the point: a group's weight relative to the rest of the index equals its earnings relative to the rest, scaled by one plus its relative multiple premium. At a 32 percent earnings share, holding 40.7 percent of weight requires a relative premium of about 46 percent. A premium of that order is defensible for a group compounding faster than the index, so the gap alone cannot separate growth from flows. What this report claims is the residual, the part of the premium that survives once expected growth is priced.
The residual can be bounded on stated inputs. Value both the cohort and the index with a two-stage dividend discount model at a 10 percent discount rate, the same payout ratio, and index earnings growth of 4 to 6 percent a year in perpetuity, and let the cohort grow faster for ten years before it converges. Three points a year of excess growth supports a relative premium of about 26 percent, and four points about 37 percent. At a 32 percent earnings share those premiums put the top ten at 37.1 to 39.3 percent of index weight against the 40.7 percent they held, a residual of about 1.5 to 3.5 points. At five points of excess growth the model supports a premium of 46 to 49 percent, the whole of the observed one, and the residual disappears. The finding rests on that growth assumption, and a reader who expects the ten to grow five points a year faster than the index for a decade can conclude that the gap is fully priced.
The arithmetic also settles what built the concentration. Since 2015 the top ten's weight has risen by 21.7 points. Their earnings share rose by about 13 points, and the growth premium accounts for all but the residual of the rest. On these inputs, earnings and the premium on expected earnings built most of the climb, and passive flows can account for at most about a sixth of it, the 3.5-point upper bound of the residual. What passive investing adds is the margin, the lock-in through every payday, and the label.
The second reading reverses the causation. Strong mega-cap returns draw money into index vehicles, so the passive share is partly an effect of the concentration. Two findings help separate the channels. Jiang, Vayanos and Zheng identify the price effect from passive flows themselves, and Chinco and Sammon place most of the undeclared passive share in institutional mandates and index-hugging funds that do not chase performance. Neither settles it. Causation runs both ways, and this report measures one direction.
The third reading doubts the measurement. The gap pairs a realized weight share with an expected earnings share from one compiler. A realized 2025 earnings share, or an operating-earnings basis, would change the width in a direction the available data cannot sign. What two vintages of one series do establish is the direction: level in 2015, apart in 2025. Nothing in Part VI depends on the width to the point.
Equal weight has led cap weight through 2026, by 4.0 points at this report’s April pin and 3.6 at the end of August, the first crack in the concentration since 2022, when a wider one opened and closed. The rotation began in cyclical sectors in late 2025. In January the Federal Reserve paused and three hyperscalers guided capital spending sharply higher.
The Invesco S&P 500 Equal Weight ETF (RSP) holds the same 500 companies as the SPDR S&P 500 ETF (SPY) at 0.2 percent each. It led SPY by 4.0 percentage points year to date at this report's April 9, 2026 pin, by 1.3 points at June 9, and led the iShares Core S&P 500 ETF by 3.6 points at August 26. That reverses 2023, 2024 and 2025, when cap weight led by 12.5, 12.1 and 6.5 points.17 The 2022 episode is the reason to hold the reading lightly: equal weight led by 6.7 points that year, a wider gap than this one, and cap weight recovered all of it across the next three years. Table A4 prints both episodes on one measure, and the range of 2026 readings is the reading to carry.
The rotation started before the new year and started outside technology. By mid-February, Nationwide's chief market strategist recorded equal weight's strongest start to a year against cap weight since 1992, with energy, industrials and materials ahead of technology. He dated the turn in cyclical sectors to the autumn of 2025 and tied it to looser monetary policy, full expensing of capital spending, and margin forecasts improving outside the largest companies.18 The equal-weight index holds each sector in proportion to its number of constituents, so it captured the move.
Two later events bore directly on the largest stocks, and they fall inside nine days. From January 28 to February 5, 2026, three of the four largest hyperscalers put a 2026 capital-expenditure figure on the record. Meta guided $115 to $135 billion on January 28, against $72.2 billion spent in 2025. Alphabet guided $175 to $185 billion on February 4, against $91 billion. Amazon guided about $200 billion on February 5, against $131 billion. At midpoints that is $505 billion guided against $294 billion spent, a step up of about 70 percent, and none of the three publishes AI revenue as a separate line against which to check it. On January 28 the Federal Open Market Committee also held its target range at 3.50 to 3.75 percent, pausing after three consecutive quarter-point cuts, with inflation described as somewhat elevated and two members dissenting in favor of a further cut.19
The first raised the capital the AI thesis has to earn back. The second put further cuts on hold, which delayed any fall in the rate at which those earnings are discounted. Both bear hardest on the longest-duration earnings in the index. Neither can be separated from the cyclical rotation already under way, and the record supports a modest reading: the rotation began on broader earnings and policy support, and the January events gave holders of the largest stocks a reason to let it run. Microsoft, which reports on a fiscal year to June, put its calendar-2026 figure of about $190 billion on the record on April 29, and Alphabet and Meta raised their ranges in April. Those three April events take the four-company total to $710 billion against $412 billion, the figure Part VI and Exhibit 7 carry.20
One trigger often named does not fit the calendar. DeepSeek's R1 model was released on January 20, 2025, and the selloff it set off ran on January 27 of that year, taking Nvidia down 17 percent in a session and erasing about $600 billion of market value.21 That is a year before this rotation, and the drawdown was retraced within 2025.
The closest precedent is the unwind of the dot-com concentration from 2000 to 2002. As concentration in technology, media and telecom unwound over 30 months, the equal-weight index beat the cap-weighted index by about 27 percentage points of cumulative total return, front-loaded: about 19 points in 2000, 12 in 2001 and 4 in 2002. The 27 points is the difference between the two three-year cumulative returns, which is a different quantity from the sum of the annual differences.22 Table A4 prints the year-by-year gap.
The 1973 to 1974 episode followed a similar pattern around the Nifty Fifty, a small set of growth stocks whose valuations had run far ahead of earnings, in a market that fell 48 percent from January 1973 to October 1974.23 Both episodes came at passive-ownership levels far below today's, before the amplification Jiang and colleagues describe operated at scale.
The trigger differs this time. In 2000 earnings never arrived at the scale valuations implied, and the multiple compressed back to fundamental value. In 1974 an oil shock repriced discount rates worldwide and exposed the Nifty Fifty's duration premium. In 2026 the hyperscalers are profitable, and the open question is when AI earnings scale. That is a more ambiguous prompt than a revenue collapse, which is consistent with the modest compression so far and points to a drawn-out resolution.
The stakes run from the retirement saver to the price-discovery function of the market. A 60/40 portfolio indexed to the S&P 500 carries 22.7 percent of total assets in ten names, and 15.5 percent in the six that earn directly on the AI build-out.
The standard case for index investing assumes the index is diversified. With 37.9 percent of weight in ten companies and 25.9 percent in six that earn on one capital cycle, that assumption fails at the top of the distribution. A 401(k) participant fully invested in an S&P 500 fund holds a position that would draw scrutiny at many institutional risk committees without breaching a standard test. The index satisfies the diversification rules for regulated investment companies, and most institutional limits bind at the single name, where a 7.0 percent holding clears. The concentration sits at the level of the theme, where few frameworks look.
A portfolio of 60 percent equity and 40 percent bonds, with the equity indexed to the S&P 500, carries 22.7 percent of total assets in the top ten names, 20.7 points in the eight AI-linked names and 15.5 in the six on the strict test. On the narrowest of the three, one dollar in six of the whole portfolio, bonds included, rides on the same build-out. The risk can be sized directly: if the six fell by half, about what the S&P 500 as a whole lost from March 2000 to October 2002, the portfolio would lose 7.8 percent of its total value from those six names alone, before any loss elsewhere.
That arithmetic assumes the entire equity sleeve is the S&P 500, and Table A6 states the construction. A 2050 target-date fund at the major providers holds more equity and puts about 30 to 40 percent of it outside the United States, which brings the AI-linked share toward the high teens. The direction holds for a real target-date fund, and the headline shrinks: the exposure is large, undeclared and unowned, and below one dollar in five.
The ECB coefficient, applied to the eight-point rise in median passive ownership, implies a cumulative correlation effect of about 0.04, which leaves the index slightly less able to tell its components apart. The Jiang, Vayanos and Zheng finding that the largest stocks are priced above fundamentals points the same way: the market's price discovery weakens as the passive share grows.
The weakening is uneven, and ownership share does not explain it. In the same November 2024 note, Goldman put passive ownership of the Magnificent Seven at 22 percent, a little below the 24 percent of the average S&P 500 stock. What sets the largest firms apart in the Jiang, Vayanos and Zheng model is that the same flow moves their prices most and that their idiosyncratic risk makes it costliest to lean against them. The market works reasonably well for pricing small and mid-sized companies and least well for the largest, whose capital-allocation decisions carry the greatest macroeconomic weight.
Institutional allocators have few better options than retail savers. Equal-weight overlays add tracking error, and explicit underweights to the top names are active bets with active risk. Most allocators treat the concentration as an artifact of the benchmark and leave it unclassified as a risk. International holders meet it through the MSCI World Index, which puts about 70 percent of its assets in US stocks. Norway's Government Pension Fund Global, at about $2 trillion the world's largest sovereign fund, held an estimated $250 billion in the top five S&P 500 names at the end of 2025, near $62 billion of it in Nvidia, as a consequence of its mandate.24 That is about an eighth of the fund, held in five American companies because of how an index is built, and no investment committee approved it as a position.
A market in which passive flows distort prices should reward fundamental analysis, and the difficulty is timing. A manager who underweighted the top ten in 2022 would have trailed badly in 2023, 2024 and 2025 before the 2026 reversal began. Measured on cumulative returns, cap weight returned 85.6 percent across those three years and equal weight 42.6 percent, a cumulative gap of 43 points: a dollar left in cap weight ended about 30 percent above a dollar that moved. The three annual differences, 12.5, 12.1 and 6.5 points, add to 31.1, which is a smaller and different quantity because the underlying returns compounded. Equal weight's 3.6-point lead in 2026 to August 26 has recovered less than a point of the 43. Being right about the structure and wrong about the date costs a career, which is why most mandates file the concentration as a benchmark artifact and give it no owner.
The strongest argument against acting on this report is in Exhibit 6. Only 21 percent of active funds beat their passive peers over ten years on a survivorship-adjusted basis, and funds investing in US large caps were the weakest category.25 Set that against the harms this report can measure: a correlation effect of 0.04 and a valuation residual of up to about 3.5 points of index weight on a stated growth assumption. Add the fee gap between a large index fund and the median active fund, compounded over a forty-year contribution horizon, and the ledger for the default 401(k) participant favors staying in the index.
Switching to equal weight has its own price on this report's series: the 43-point cumulative gap an equal-weight allocator still trails by since the start of 2023, before the higher expense ratio an equal-weight fund charges for its rebalancing. This report compares long-run equal-weight and cap-weight returns only for 2000 to 2002 and for the years since 2022, and that record does not support a recommendation to switch.
The claim is therefore narrower than the diagnosis. The concentration is an undeclared and unowned risk inside a product sold as diversification, held by savers who did not choose it and cannot see it. The right response for most of them is to know the position, size the rest of the portfolio around it, and stop calling an index fund diversified when its top ten hold 38 percent of it.
At the May cutoff this report set one test: operating earnings, stripped of one-time items, growing fast enough to earn back the capital. The Q2 2026 season was its first reading, and it failed.
The test this report set at its May cutoff is whether hyperscaler operating earnings, stripped of equity-security gains and one-time tax items, grow fast enough to earn back a capital base that rose about 70 percent in a year. The test has two halves. The first is whether reported earnings can be read at face value: a quarter counts as clean only when no one-time item exceeds a tenth of reported earnings per share. The second is whether the operating earnings underneath grow fast enough. Q1 failed the first half at Alphabet and Amazon, and on the table in Part III little of the reported surprise survived the adjustment. Two clean quarters with strong operating growth would make the weight a fair price for real earnings.
The Q2 2026 season has reported, and it was not clean. Alphabet booked a $99.0 billion gain on equity securities, worth $6.26 of its $9.11 of diluted earnings per share, about 69 percent. Amazon booked a $53.4 billion pre-tax mark on its Anthropic holding. All four hyperscalers raised 2026 capital-expenditure guidance on their July calls, to a compiled total near $760 billion.26 The quarter failed the first half of the test again, for a different reason. Alphabet's 10-Q attributes its gain primarily to SpaceX, which merged with xAI early in 2026 and was reclassified as a marketable security after its June 12 listing, worth $94.0 billion at June 30, and to one unnamed private company. The SpaceX gain is a public share price, so the funding-round sequence in Part III, in which a funder's own investment lifts the valuation it books, describes Amazon's mark and at most part of Alphabet's. Two seasons in a row, reported earnings rested mainly on equity marks, and in Q2 the largest of them moved with a listed stock. The Q3 release has already shown the cost of reading such marks as earnings: Alphabet booked an $8.7 billion loss on equity securities, which its finance chief attributed primarily to the fall in SpaceX's share price. The second half of the test is still open, because the marks hide the operating trend as much as they flatter it.
Between earnings seasons, the spread between RSP and SPY is the clearest public signal of whether the AI premium is widening or compressing. A gap above five percentage points held for a full quarter would mark an unwind that feeds itself: slower passive flows to the top names weaken their price support, which lowers their weight, which sends a smaller share of each contribution their way. The 2026 gap has stayed below that threshold at every reading so far: 4.0 points against SPY at April 9 and 1.3 at June 9, and 3.6 against the iShares Core S&P 500 ETF at August 26. A series that halves and recovers inside a quarter is a range, and these two series have opened and given back a gap of this size inside a year before.
The $710 billion the four hyperscalers guided for 2026 at the May cutoff is about 72 percent above the $412 billion they spent in 2025, and the July raises took the compiled total near $760 billion. Every figure is total capital expenditure. Amazon's includes fulfillment and Kuiper, Alphabet's includes offices and non-AI infrastructure, and none of the four publishes AI capital spending separately, so the capital the thesis has to earn back is some amount below the total. Exhibit 7 shows the company breakdown. The signal to watch is the commentary on AI revenue conversion and unit economics on each call. If AI workloads displace commodity compute at high margins, the spending will produce earnings; at commodity margins, it produces an ordinary infrastructure build-out, short of the premium current multiples embed.
Three scenarios span the outcomes. In the first, the AI earnings case reasserts itself: the hyperscalers report clean operating beats without one-time marks, equal weight's lead reverses, passive flows continue, and concentration stabilizes near 38 percent and retests the 2025 record. The 2026 rotation proves to be a factor cycle.
In the second, enterprise adoption stalls: operating growth falls short of the rate the capital needs, the spending is judged premature, multiples compress, and the equal-weight gap passes the five-point threshold. With passive ownership at 53 percent of fund assets, the flow mechanism works in reverse with more force than in 2022, and concentration falls from 38 percent toward 30 to 32 percent over 18 to 24 months. This is closest to the 2000 to 2002 precedent, with profitable companies at its center.
In the third, concentration holds at 35 to 40 percent for several more years while passive flows grow and AI produces real operating leverage over a longer horizon than current multiples assume, until rate normalization or a broader earnings rotation forces the reversion. Active managers are right about the structure and early on the timing.
The first scenario needed two clean quarters, and Q2 was not one, so it has lost ground. The second needs an operating disappointment the record has not yet shown: the one-time items flattered earnings without concealing a collapse beneath them. The third asks only that nothing break, and it is where the evidence to September leaves the weight. Two clean quarters with strong operating growth would move weight back toward the first; an operating miss on the scale of Amazon's Q1 at two of the four would move it toward the second.
Six variables will show whether this reading holds, and each is observable in public data. Three of them would falsify the thesis if they moved far enough, and the trigger column states how far. One falsification sits outside the list: a sustained improvement in active funds' results against passive benchmarks would show that the flow regime had grown large enough to fund the active research that corrects it. The sixth variable has no publication calendar. The Chinco and Sammon estimate is an academic reconstruction whose latest reading is for 2021; it appears here as the level that would matter if it were measured again.
| Variable | What a move would mean | Reading | Falsifying trigger | Horizon |
|---|---|---|---|---|
| RSP minus cap weight, year to date | The AI premium widens or compresses, read in real time | +3.6 pts vs IVV (Aug 26), from +4.0 and +1.3 vs SPY | Above +5 pts, held a full quarter | Q4 |
| Hyperscaler operating EPS against consensus | AI monetization shows up stripped of one-time items | Q2 item at 69% of reported EPS, Alphabet | Two straight quarters with no one-time item above a tenth of EPS | Q3 and Q4 reports |
| Top-ten share of index weight | The concentration itself turns, or resumes | 37.9% (Mar 30), from a record 40.7% | Below 35% for two consecutive quarters | Q4 |
| ICI monthly active and passive flows | Price discovery begins to repair | 52.7% passive (Mar 2026) | Passive share falls three months running | Q4 |
| Combined 2026 hyperscaler capex guidance | The capital base the AI thesis has to earn back | Near $760bn after the July calls | A cut at any of the four on the Q3 calls | Oct and Nov calls |
| Passive ownership, Chinco-Sammon basis | The measured passive share reaches a policy threshold | 33.5% (2021, latest published) | Above 40% on the same method | No publication calendar |
Each falsifying trigger is set in the direction of the move that would change the view. The spread, the Q2 earnings and the capex guides run to September 2026; every other reading stands at the May 2026 cutoff.
Source: PortfoliosLab; ETFCentral; Fortune; company 8-K and 10-Q filings; Investment Company Institute; RBC Wealth Management / FactSet; Chinco and Sammon (2024). Three series updated through September 2026; analytical readings stand at the May 2026 cutoff.
The S&P 500 is sold as diversification and now carries a concentrated bet on one AI build-out. Earnings built most of that weight, and passive flows amplify it at the margin and hold it in place. The 2026 reversal is weak evidence of an unwind: the same two series opened a wider gap in 2022 and closed it within three years, though passive funds now hold more than half of fund assets and the four hyperscalers have guided near $760 billion of capital spending. The harms that can be measured are small, 0.04 of correlation and up to about 3.5 points of index weight. The exposure is large and undeclared, and the savers who carry it did not choose it. Knowing the position is the first step and sizing the rest of a portfolio around it the second. On this evidence the third step is to hold the index knowingly and review the position if the spread holds above five points for a quarter or a clean earnings season fails the capital test.
Every figure traces to a named, dated source, and a second reading is used as a cross-check wherever one exists. The analytical cutoff is May 2026. Three series are updated to September 2026 and say so at each use: the equal-weight spread, the Q2 2026 earnings and the July capital-expenditure guides. Index weights are free-float weights at March 30, 2026, with Alphabet's share classes combined; the tenth constituent is contested on that basis, between JPMorgan Chase and Eli Lilly, and is left unnamed. Every concentration reading is the RBC Wealth Management and FactSet series and every constituent weight is Slickcharts, and no calculation mixes the two; Morningstar's 37.3 percent for 2024 in Table A1 is corroboration only. Fund assets are US long-term mutual-fund and ETF assets on the Investment Company Institute's definition, except the December 2023 crossover, which is Morningstar's. Returns are total returns. Where a return difference spans more than one year, the annual differences and the difference in cumulative return are distinct quantities: the 27 points for 2000 to 2002 and the 43 points for 2023 to 2025 are both cumulative, and neither is the sum of the annual bars.
Some sources could be read only in part. The Jiang, Vayanos and Zheng figures are quoted from sections 5 and 6 of the authors' posted final version, since the journal's own PDF was not retrievable; the citation was confirmed against the journal record. The Q1 2026 estimates are FactSet's, as printed in its May 4 note. The Goldman passive-ownership readings come from the press reports of the note, which is behind a subscriber wall, and the Barometer's five-year success rate rests on Morningstar alone. Three figures earlier drafts carried are withdrawn for want of a source: intermediate points on the concentration series, a 2015 weight-and-earnings pair at 18 and 18, and the Big Three voting series before 2025. Four plates earlier editions carried are folded into the text and tables because every figure on them appears there: the declared and inferred passive-ownership pair (Part II), the 2015 and 2025 weight-and-earnings pairs (Part III), the 60/40 holdings (Part V and Table A6), and the 2000 to 2002 equal-weight spread (Part IV and Table A4). The comparison of a 71 percent contribution share with a per-share surprise ratio is withdrawn as invalid, since the two have different denominators.
Four limits bear on interpretation. The residual in Part III is a calculation on stated inputs, a two-stage dividend discount model at a 10 percent discount rate with three to four points of excess growth for a decade, and it moves with both; at five points it is zero. Every capital-expenditure figure is total company spending, since none of the four publishes an AI line. The 60/40 arithmetic assumes an equity sleeve made entirely of the S&P 500; a 2050 target-date fund holds more equity and 30 to 40 percent of it abroad, which moves the AI-linked share toward the high teens. And the weight-and-earnings pairs come from one publisher, with the 2025 earnings leg an expectation and the 2015 leg a realized share, so the direction is firmer than the width.
| Dataset | Source | Frequency | Coverage | Retrieved | Used in |
|---|---|---|---|---|---|
| Top-ten share of S&P 500 index weight | RBC Wealth Management / FactSet; Morningstar | Annual | 1990 to Q1 2026 | Mar 30, 2026 | Exhibit 1, Table A1 |
| S&P 500 constituent free-float weights | Slickcharts | Daily | March 30, 2026 | Mar 30, 2026 | Exhibit 2, Table A2 |
| Index and active US long-term fund assets | Investment Company Institute | Monthly | 2006 to March 2026 | Mar 2026 | Exhibit 3, Table A3 |
| Passive-active crossover, December 2023 | Morningstar | One-off | December 2023 | Jan 2024 | Part II, Table A3 |
| Passive ownership share, declared and inferred | Chinco and Sammon, JFE 157 (2024) | One-off | 2021 | 2024 | Part II, watchlist |
| Passive-flow price elasticity | Jiang, Vayanos and Zheng, RFS 38(12) | One-off | published Dec 2025 | Dec 2025 | Part II, At a Glance |
| Passive ownership and index correlation | European Central Bank, Financial Stability Review | One-off | 2010 to 2024 | Nov 2024 | Parts II and V |
| Passive ownership by stock | Goldman Sachs, via MarketWatch | One-off | 2004 to 2024 | Nov 2024 | Parts II and V |
| Top-ten share of index weight and of earnings | RBC Wealth Management / FactSet | Annual | year-end 2015 and 2025 | Dec 31, 2025 | Part III |
| Hyperscaler quarterly EPS and one-time items | Company 8-K and 10-Q filings; FactSet | Quarterly | Q1 and Q2 2026 | Jul 2026 | Exhibit 4, Parts III and VI |
| 401(k) assets and fund composition | ICI; BrightScope / ICI | Annual | 2022 and 2025 | Dec 2025 | Part III |
| RSP and SPY total returns | PortfoliosLab; ETFCentral; Fortune | Daily | 2022 to Aug 26, 2026 | Aug 26, 2026 | Exhibit 5, Table A4 |
| S&P 500 Equal Weight against S&P 500, total return | S&P Dow Jones Indices | Annual | 2000 to 2002 | 2026 | Part IV, Table A4 |
| Active-fund success rates by horizon | Morningstar US Active/Passive Barometer | Annual | to year-end 2025 | YE 2025 | Exhibit 6 |
| Hyperscaler capital expenditure and guidance | Q4 2025 releases and calls; April 2026 updates | Quarterly | 2025 actual, 2026 guided | Apr 29, 2026 | Exhibit 7, Table A5 |
| Norges Bank equity holdings | NBIM Annual Report 2025 | Annual | year-end 2025 | 2026 | Part V |
| FOMC decision | Federal Reserve Board | Meeting | January 27-28, 2026 | Jan 28, 2026 | Part IV |
Index weights move daily, and the March 30, 2026 file is a single-day reading; the top-ten total comes from the concentration series, since the constituent file ends one name short on a combined-Alphabet basis. Fund-asset figures are revised after first release, and the March 2026 vintage is used throughout. Company guidance is restated between calls: Meta's 2026 range moved twice before the May cutoff, and the April 29, 2026 raise is the figure carried.
Every row is a published reading, and no years between them are filled in. The 2024 row is Morningstar's, printed as corroboration; every other row is RBC Wealth Management / FactSet, and no calculation combines the two.
| Year | Share | What the reading is |
|---|---|---|
| 1990 | 19.0 | Published reading, RBC / FactSet |
| 2000 | 27.0 | Published reading, intra-year dot-com peak; about 23 at December 31 |
| 2015 | 19.0 | Published reading, post-1990 low |
| 2024 | 37.3 | Published reading, Morningstar, at December 31 |
| 2025 | 40.7 | Published reading, record, at December 31 |
| Q1 2026 | 37.9 | Published reading, March 30, first step down |
Source: RBC Wealth Management / FactSet, "The Great Narrowing: S&P 500 concentration," data as of December 31, 2025; Morningstar for 2024. Behind Exhibit 1.
The six names on the strict test are Nvidia, Microsoft, Alphabet, Amazon, Broadcom and Meta, at 25.9 percent; the eight AI-linked names add Apple and Tesla. The last three rows are aggregates of the rows above them, so the weight column does not add down its length. The top-ten total comes from the concentration series, because the retrievable ranking counts Alphabet's classes separately and the tenth constituent on a combined basis is contested between JPMorgan Chase and Eli Lilly.
| Company | Weight | AI-linked |
|---|---|---|
| Nvidia | 7.0 | Yes |
| Apple | 6.3 | Yes |
| Alphabet (Class A and C) | 5.8 | Yes |
| Microsoft | 4.6 | Yes |
| Amazon | 3.7 | Yes |
| Broadcom | 2.4 | Yes |
| Meta | 2.4 | Yes |
| Tesla | 2.3 | Yes |
| Berkshire Hathaway | 1.8 | No |
| Tenth slot, contested | 1.4 | No |
| Subtotal · eight AI-linked names | 34.5 | Yes |
| Subtotal · nine named companies | 36.3 | |
| Total · top ten, concentration series | 37.9 |
Source: Slickcharts, S&P 500 Companies by Weight, March 30, 2026; RBC Wealth Management / FactSet for the top-ten total. Behind Exhibit 2.
Index share is index assets divided by the sum of the two columns beside it. The December 2023 crossover is Morningstar's reading; every other row is the Investment Company Institute's, and the two houses count slightly different fund universes. The rows before 2021 are five-year samples from the Institute's Fact Book trend series. Fund assets are US long-term mutual-fund and ETF assets, excluding money-market funds, and the index column counts declared index vehicles only. Chinco and Sammon's 33.5 percent for 2021 measures passive ownership of the whole US stock market, a different universe, and is not comparable with this column.
| Date | Index | Active | Index share |
|---|---|---|---|
| 2006 | 1.00 | 7.50 | 11.8% |
| 2011 | 2.30 | 9.00 | 20.4% |
| 2016 | 5.30 | 11.60 | 31.4% |
| 2021 | 11.60 | 14.60 | 44.3% |
| Dec 2023 | 13.29 | 13.23 | 50.1% |
| Mar 2026 | 19.09 | 17.12 | 52.7% |
Source: Investment Company Institute, Active and Index Investing Release, March 2026, and ICI Fact Book trend series; Morningstar for December 2023. Behind Exhibit 3.
A positive number is equal-weight outperformance. The 2000 to 2002 rows are the S&P 500 Equal Weight Index against the S&P 500, backtested before its January 2003 launch; the 2022 to 2026 rows are RSP against SPY, except the 2026 row, which is RSP against the iShares Core S&P 500 ETF at August 26 (+4.0 at April 9 and +1.3 at June 9 against SPY). The two blocks are separate series, index against index and fund against fund. Annual rows do not add to the cumulative gaps the text cites: 27 points for 2000 to 2002 (the rows add to 34.5) and 43 points for 2023 to 2025, cap weight's 85.6 percent against equal weight's 42.6 (the rows add to 31.1).
| Year | Difference | Episode |
|---|---|---|
| 2000 | +18.8 | Dot-com unwind, year one |
| 2001 | +11.8 | Dot-com unwind, year two |
| 2002 | +3.9 | Dot-com unwind, year three |
| 2022 | +6.7 | Rate shock |
| 2023 | −12.5 | Cap weight leads |
| 2024 | −12.1 | Cap weight leads |
| 2025 | −6.5 | Cap weight leads |
| 2026 YTD | +3.6 | Reversal, to August 26 |
Source: S&P Dow Jones Indices; PortfoliosLab; ETFCentral; Fortune, August 28, 2026 for the 2026 year-to-date row. Behind Exhibit 5.
Ranges are taken at midpoint and Microsoft is on a calendar-year basis. The last row is an aggregate of the four above it and is labeled as such; it reproduces by addition from the columns printed here, and the increase is 72 percent of the 2025 base. Every figure in this table is total company capital expenditure. None of the four publishes an AI capital line, so the AI share of the 2026 guide cannot be separated from fulfillment centers, satellite constellations, offices and ordinary data-center replacement, and the capital the AI thesis has to earn back is some amount below the total printed here. Guidance is also not spending: each figure is what the company said it expected to spend at the date in the last column, and Alphabet's and Meta's figures had already been raised once when this report closed. A reader tracking the capex test in Part VI should watch the quarterly commentary on revenue conversion ahead of the headline, which moves on every call.
| Company | 2025 | 2026 | Increase | Guidance dated |
|---|---|---|---|---|
| Amazon | 131 | 200 | +69 | Feb 5, 2026 |
| Microsoft | 118 | 190 | +72 | Apr 29, 2026 |
| Alphabet | 91 | 185 | +94 | Apr 2026 raise |
| Meta | 72 | 135 | +63 | Apr 29, 2026 raise |
| Total · four companies | 412 | 710 | +298 |
Source: Q4 2025 earnings releases and calls, with the April 2026 guidance raises at Alphabet and Meta. Behind Exhibit 7.
Of total assets is index weight scaled by the 60 percent equity allocation. The equity rows are nested sets: Nvidia inside the six, the six inside the eight, the eight inside the top ten, so the column does not add down its length. The top-ten row is the concentration series' total; the rows beneath it come from the constituent weights in Table A2. The six are the names whose reported earnings move directly on the AI build-out, on the test stated in Part I; the eight add Apple and Tesla. The equity sleeve here is entirely the S&P 500. A 2050 target-date fund holds more equity and places 30 to 40 percent of it outside the United States, which pulls the AI-linked share toward the high teens. To carry the arithmetic to another portfolio, scale the index-weight column by its equity share, then by the share of that equity in US large caps.
| Holding | Index weight | Of total assets |
|---|---|---|
| Equity sleeve | 100.0 | 60.0 |
| Top ten names | 37.9 | 22.7 |
| Eight AI-linked names | 34.5 | 20.7 |
| Six on the strict test | 25.9 | 15.5 |
| Nvidia alone | 7.0 | 4.2 |
| Bond sleeve | 40.0 |
Source: Power Curve calculation on the index weights in Table A2.
Hart, Brendan. 2026. “Betting on the Index.” The Power Curve, October. https://brendanhart.com/reports/betting-on-the-index.
