
The AI Capital Loop
Institutional ownership and voting
Three distinct claims hide under ownership
FIELD NOTE Economic interest, voting authority, and index weight answer different questions.
Who bears the gains and losses?
The ultimate investor bears the asset’s economics through a direct holding, a retirement vehicle or a fund. In a pooled product, the investor typically owns shares of the fund, not separately registered shares of every constituent. An adviser manages portfolios and may administer voting for a fund, and a bank or nominee can appear as holder of record. The reporting word beneficial can refer to voting or disposition power under a defined legal rule, not merely ultimate economic entitlement. These layered interests can coexist without representing multiple economic stakes. [65] [86]
The attachment’s phrase “the asset managers own the tech firms” therefore combines influence and economic risk. Managers can be governance intermediaries for millions of investors without those positions belonging to the manager’s own balance sheet. The separation is not an argument that proxy power is unimportant.
Which question can a disclosure answer?
Form 13F primarily measures reported long securities under investment discretion for the reporting snapshot. Rule 13d-3 beneficial ownership can include power to vote or direct disposition, and specified rights to acquire shares. A proxy provides corporate-class and record-date voting rules. N-PX shows how covered funds or managers reported voting on particular items. Benchmark weight instead measures the constituent’s role in index performance. None answers all four questions. [64] [66] [86] [95]
A useful map attaches four attributes to each number: security class, legal entity, date and denominator. Separate who bears the dollars from who can vote. A substantial reported holding is evidence of exposure or influence potential, not proof that a holder directs AI strategy or coordinated with a competitor.
A four-filer reconciliation from EDGAR
FIELD NOTE Quarter-end share rows are reconstructed by CUSIP and not counted as votes.
Reproducible scope
The Q3 2026 exercise uses original SEC information tables for [*REDACTED*], [*REDACTED*], [*REDACTED*] (VCM), and [*REDACTED*] (VPM). Each table is joined by nine-character CUSIP. The reconstruction sums ordinary share rows for the identical class, drops put and call records, and retains counts for sole, shared and no voting authority rather than projecting votes from gross shares. VCM, VPM and [*REDACTED*] filed November 13; [*REDACTED*] filed November 12. Positions are dated September 30, not November. [67] [68] [69] [70]
This is a four-filer comparison, not an all-institution universe. The [*REDACTED*] parent filed a notice pointing to multiple separately reporting entities. The notice has no additive holdings table, and the two selected advisers do not exhaust the parent’s affiliates. [*REDACTED*]’s cover distinguishes discretion exercised by operating subsidiaries from the parent itself. [71] [67]
Avoid false precision
Match a denominator to the security class before calculating reported shares as a percent of shares outstanding. Where a nearby rather than exact date is used, flag the mismatch and use limited precision. Position lines split by discretion and manager code require inspection. Across independent filings, gross arithmetic sums may not fully eliminate all overlapping shared discretion arrangements; no deduplication beyond the identified filing set and ordinary-share filter is claimed. Do not describe the result as reconciled legal beneficial ownership. [64]
The method deliberately does not publish a sector-wide 50–70% figure. A legitimate rollup would need a full institutional manager universe, combination/notice relationships, amendments, confidential-treatment caveats, exact class balances and an explicit overlap approach. It would still not identify underlying retirement savers or investment strategy from each manager line.
[*REDACTED*] is a same-date test case
FIELD NOTE The September 30 denominator is known to the individual common share.
What the 13F rows show
Microsoft’s proxy identifies 7,426,426,719 common shares outstanding at the September 30, 2026 record date, one vote per share. At that same quarter-end, the selected EDGAR holdings totals are [*REDACTED*] 607.119 million, VCM 486.085 million, VPM 167.213 million, and [*REDACTED*] 320.716 million. Relative to the proxy denominator these approximate 8.18%, 6.55%, 2.25%, and 4.32%. The four gross percentages sum to 21.29% on unrounded values. The two [*REDACTED*] advisers together report about 8.80% of [*REDACTED*] shares in this limited filer set; the footnotes do not establish a common [*REDACTED*] vote. [74] [67] [68] [69] [70]
This validates the attachment’s approximate 8% [*REDACTED*] and 4.2–4.3% [*REDACTED*] intuition for [*REDACTED*], but revises [*REDACTED*]’s entity label: the cited mid-6% covers VCM only, not VCM and VPM together.
Why a contemporary proxy can contain older inputs
The October 21 proxy table shows 7.48% for [*REDACTED*] and 7.30% for [*REDACTED*], but footnotes use a Schedule 13G filed April 30, 2026 and Schedule 13G/A filed February 13, 2024, respectively. The proxy’s September heading does not independently update those filing inputs. Even two truly same-date 13F shares and Schedule 13G figures could differ because their scope and defined powers differ. Store each document’s effective inputs and filing date separately. [74] [75] [64] [65]
Use the 13F arithmetic for managed-position exposure and use the dated corporate statements for class rights. Neither substitution nor averaging across the two is appropriate. [*REDACTED*]’s December meeting was not yet held at the cutoff; outcomes cannot be inferred from the 2026 proxy.
[*REDACTED*] is not a five-class ownership puzzle
FIELD NOTE Use SH common lines and exclude options; do not treat the share denominator as exact September 30.
Reported shares and issuer basis
NVIDIA’s common 13F CUSIP shows [*REDACTED*] 1,940.886 million shares; VCM 1,540.601 million; VPM 512.235 million; and [*REDACTED*] 1,025.075 million. Dividing these quarter-end holdings by the approximately 24.1 billion shares [*REDACTED*] reported outstanding on August 21 gives roughly 8.05%, 6.39%, 2.13% and 4.25%. The implied gross four-line total is about 20.8%, not a precise same-day stake. A share-count exhibit avoids blending observation dates. [67] [68] [69] [70] [77]
The May 2026 [*REDACTED*] proxy states that the listed common stock had one vote per share and shows 7.43% [*REDACTED*] and 7.31% VCM beneficial ownership. Those numerators are sourced to different observation dates: the [*REDACTED*] footnote references December 31, 2023, while [*REDACTED*]’s references March 31, 2026. The recognized percentages should not be substituted for the newer selected four-filer 13F positions. [76]
Voting fields are narrower than exposure
In Q3 [*REDACTED*] reports about 1,755.0 million [*REDACTED*] shares with sole authority under the 13F field definition, [*REDACTED*] roughly 130.1 million sole plus shared, and VCM/VPM classify their reported position rows as no voting authority on the required form. These are not meeting tallies, and they do not imply all the index funds under a brand vote zero shares. The 13F authority instructions are specific to such reporting, including when only routine authority is exercised; the [*REDACTED*] fund filings retain a different voting role at the fund level. [64] [67] [70] [68] [69] [87]
Potential influence at a non-founder-controlled one-vote-per-share firm can be sizable, but market share, bargaining position, founder relationships, directors and management strategy must be analyzed separately.
[*REDACTED*], [*REDACTED*], and the limits of the peer estimates
FIELD NOTE Selected filing sums support overlap, not a sector-wide voting bloc.
Two single-common share examples
Rounding original 13F share counts against [*REDACTED*]’s 14.532259 billion shares on September 26 yields approximately 7.96% [*REDACTED*], 6.57% VCM, 2.27% VPM and 4.28% [*REDACTED*], approximately 21.09% combined on unrounded estimates. The holdings date is four days later than the issuer denominator. In [*REDACTED*]’s case a September 30 balance-sheet denominator of 10.811 billion outstanding shares gives about 6.92%, 5.91%, 1.09% and 3.70% in the same order, approximately 17.63% combined. The issuer’s quarter-end share figure is rounded to the million. [84] [85] [67] [68] [69] [70]
The company-by-company differences are evidence against treating a single industry band as a measurement. [*REDACTED*]'s 8.15% index weight on November 18 is a different number from the fraction of [*REDACTED*] stock reported by [*REDACTED*]. The near equality of two unrelated percentages would be coincidence, not corroboration. [93]
What varies even within one manager
13F does not label the underlying holding as passive, a specific retail ETF, institutional separate account or active strategy. An adviser may exercise different powers across these accounts; the filing's discretion scope helps detect that ambiguity but does not resolve it. The benchmark weight does not tell whether a manager used full replication, optimization, derivative exposure or delegated voting for one sleeve. [64] [96]
An analytical label such as “selected overlapping public managed positions” is supportable. “Three firms voted one fifth of these securities as a unit” is not. For investment committee tables, include issuer denominator date within the same exhibit and avoid implying that end investors all choose the same financial exposure or receive the same stewardship decisions.
[*REDACTED*] shows how minority economics can control votes
FIELD NOTE A dollar of common equity and a corporate vote do not map one-for-one across classes.
The controlling shareholder is explicit
Meta stated that as of October 1, 2026 [*REDACTED*] beneficially owned 1,220,703 Class A and 341,215,148 Class B shares, about 13.4% of total economic interests and 60.6% of voting power. Class B carried ten votes per share and Class A one. The definitive November information statement reports approved written consent from holders controlling approximately 56.1% of outstanding voting power on October 20, with those participating shares beneficially owned by [*REDACTED*]. Approval by that subset is not the same measure as his total voting position. [79] [80]
The documentation describes creation of a nonvoting Class C security and a recapitalization intended to preserve voting control against future dilution from Class A issuance. The approval and effectiveness/issuance are different events; the information statement establishes the former and describes the proposed implementation conditions. [80]
Index-manager leverage remains bounded
Selected September Class A 13F positions imply approximately 7.84% [*REDACTED*], 6.52% VCM, 1.74% VPM and 4.21% [*REDACTED*] against the near-date October 16 Class A denominator. The gross 20.31% Class A sum reflects roughly 449.8 million Class A shares and is not 20.31% of corporate votes. If every such share carried and exercised one vote and the class structure were unchanged, the shares would represent roughly 8% of all then outstanding voting units, before mandate restrictions. That is a mechanical upper illustration, not actual influence measured at a meeting. [81] [80] [67] [68] [69] [70]
Use separate charts for economic interests, common-class holdings and total voting power. Institutional voices can matter to disclosure and market reputation while remaining unable to outvote a controlling founder.
[*REDACTED*] splits benchmark exposure from ballots
FIELD NOTE Class A voting stock and Class C economic exposure should never be conflated.
Founder votes versus reported ownership
Alphabet’s April 6, 2026 proxy ownership table gives [*REDACTED*] 27.4% and [*REDACTED*] 25.3% of total voting power, about 52.7% in arithmetic aggregate. Their power comes largely from Class B shares with ten votes each. Class A has one vote per share; Class C is generally nonvoting and excluded from this voting table. The filing shows [*REDACTED*] at 6.1% of Class A and 2.5% of total voting power as represented in the table, with the filing’s source and voting/disposition footnotes important. A combined majority by two founders is not a claim about individual majority power or an assertion of a legally coordinated group. [82]
Class A and Class C footprints
The four 13F filers report approximately 1,183.1 million Class A shares combined and 945.8 million Class C shares combined. Using the later October 21 rounded class counts of 5.888 billion and 5.579 billion yields rough class ratios of 20.09% and 16.95%. The comparison illustrates the class distinction, not nine-digit estimates of September beneficial voting power. A cap-weighted fund may hold both public [*REDACTED*] lines; securities-level index reporting may separate them even though they are one business’s economic exposure. [83] [67] [68] [69] [70] [93]
For an index-concentration exhibit add the two weights when reporting [*REDACTED*] as an issuer. For governance, calculate votes by class and use recorded vote evidence. End investors may bear risk across both economic classes without acquiring a proportional say in [*REDACTED*]’s strategic direction.
[*REDACTED*] changed its reporting perimeter
FIELD NOTE 2026 was a structural reporting break, not necessarily a sale of underlying shares.
What changed
Vanguard says it completed a January 12, 2026 internal reorganization of U.S. portfolio management across [*REDACTED*] and [*REDACTED*]. The two subsidiaries have distinct investment and stewardship teams. [*REDACTED*] states that the funds’ existing investment strategies were unchanged by this operational change. The parent’s third-quarter 13F notice lists multiple other filing managers, including both new advisers. The correct interpretation is a change in filing entity and discretion perimeter, which can happen without a benchmark fund selling a share. [72] [73] [71]
Historical vendor charts that compare a consolidated brand before the change with a single newly reported adviser afterward can show a spurious loss of exposure. Retain CIKs, dates and notice relationships; present any reassembled family estimate as such.
Who controls the ballot?
The [*REDACTED*] Index Funds N-PX explanatory note says U.S. [*REDACTED*]-advised funds retain proxy authority; distinct teams administer voting policies. A filing’s financial-reporting control concept is therefore not a shortcut for predicting ballots. A percentage for VCM plus VPM can illustrate combined managed exposure but cannot be called a 13D group, a shared vote commitment, or all [*REDACTED*] activity. Other filing entities listed by the parent would need review for a fuller family picture. [87] [65] [71]
Keep adviser names fully qualified in research models, especially for separated teams and series. Check whether securities are listed as delegated, whether the relevant meeting was covered, and whether the actual voting measure is item-specific. Reorganization effects are a data-quality confound that matters precisely because well-known adviser brands appear to be stable labels.
The correct measure of votes is the proposal record
FIELD NOTE Actual votes can depart from management recommendations and from quarter-end 13F classifications.
[*REDACTED*] counterexample
At [*REDACTED*]’s June 24, 2026 meeting, the issuer recommended against a nonbinding shareholder proposal to replace specified supermajority provisions with a simple majority standard. [*REDACTED*] reported 14,589,671,908 votes for, 2,210,282,205 against, 56,735,273 abstaining and 2,829,718,733 broker nonvotes. For divided by for-plus-against is about 86.8%; this is not a fraction of all common shares outstanding. The vote approved an advisory proposal, not a finding of corporate control. [76] [78]
The [*REDACTED*] 500 Index Fund N-PX contains a record showing a 622.526 million-share segment in favor on that exact proposal. Other lines for the same series/meeting/proposal show segmented votes; the large segment is not the entire family’s vote. This directly demonstrates a major managed index exposure voting against a board recommendation on an actual governance item, without demonstrating how the same fund voted on AI spending. [87]
How to audit a ballot
Identify issuer, class, CUSIP, meeting date, series and proposal. Use the issuer proxy for the board recommendation and the 8-K or equivalent for the final tally. Match N-PX records across split segments and check whether shares on loan were available. Registered funds report a wider proxy record; institutional manager reporting on N-PX covers specified executive compensation votes rather than all governance issues. The quarter-end 13F and vote record may not reference the same position size even when both are correct. [86] [87] [88]
Record whether the item was binding, advisory, contested, withdrawn or rejected. One manager may support boards on standard matters and dissent on governance. Robust documentation avoids inferring private pressures or AI-specific intent from an ordinary governance vote.
Management alignment is a sample statistic
FIELD NOTE High aggregate agreement is not evidence of uniform voting or a conspiracy.
What the proxy-season data shows
Morningstar analyzed 50 large U.S. managers of equity and allocation funds across 21,208 voted resolutions in its [*REDACTED*] US Large-Mid Cap Index company sample over three proxy years ended June 30, 2025. It reports the three biggest index managers averaged 96.0% support for management proposals in 2023, 98.0% in 2024 and 98.7% in 2025. Their 2025 shareholder-proposal support averaged 7.5%. This validates the direction of the attachment’s management-alignment number, but neither number is a 2026 proxy season total or an all-company vote count. [*REDACTED*] reports unweighted manager peer averages, not shares controlled. [89] [90]
Director elections dominate management resolutions in the sample, so a high approval rate partly reflects proposal mix. A shareholder proposal can itself be supported by management. Percent approval of management-sponsored resolutions is not identical to percent agreeing with management across all contested issues.
The mechanism still needs evidence
Institutional influence can involve dialogue before a vote, published guidelines, choosing directors and pivotal proposal votes. Aggregate agreement may reflect directors’ conventional slates, robust reporting, similar fiduciary judgments or limited incentives to conduct governance research. Similar revealed preferences are insufficient evidence that funds coordinate investment decisions. When filings show segmented votes for one series, advisers’ published recommendations and end-investor choices can further diverge. [87] [99] [100]
[*REDACTED*] and [*REDACTED*]’s historical roughly 25% of cast votes for S&P 500 firms is a historical estimate with assumptions, including assumed participation by the large managers when modeling aggregate votes. It cannot establish what the managers cast at selected November 2026 tech companies. [91] [92]
A broad index can carry a concentrated risk slice
FIELD NOTE Use the official index panel; combine share classes only for issuer-level grouping.
The dated snapshot
State Street’s SPY page displays separate fund and index holdings panels. In the index panel as of November 18, 2026, weights are [*REDACTED*] 8.15%, [*REDACTED*] 5.98%, [*REDACTED*] 5.79%, [*REDACTED*] 4.16%, [*REDACTED*] 2.67%, [*REDACTED*] A 2.63%, [*REDACTED*] A 2.15%, [*REDACTED*] C 2.10%, [*REDACTED*] 1.75% and [*REDACTED*] 1.71%. Summing the top ten security lines gives 37.09%, based on displayed rounded weights. Top six security lines total 29.38%; top six issuer groups after combining [*REDACTED*]’s two lines total 31.48%. The six groups are [*REDACTED*], [*REDACTED*], [*REDACTED*], [*REDACTED*], [*REDACTED*] and [*REDACTED*], in descending grouped weights. [93]
These are component weights, not the fraction of shares any manager owns and not the fraction of firms’ revenues exposed to AI. The benchmark bundles multiple business models, some predominantly outside chip and cloud infrastructure.
What a weight can predict mechanically
For a simple point-in-time scenario holding other returns fixed, a 10% decline in the six issuer prices is approximately a 3.15 percentage point direct contribution to benchmark return before second-order interactions. This is weighted arithmetic, not a forecast, and is not an assumption that all other constituents fail to react. Fund performance may also differ from benchmark due to expenses, tracking error and implementation. [93] [95]
Index weights move with prices, share count, float and constituent composition. A different snapshot can present different rankings without evidence of sudden net selling. The September 16 [*REDACTED*] mirror, for example, showed [*REDACTED*]’s index line at 7.98% versus 5.79% November 18. That comparison is about benchmark weights; it does not identify what orders produced the move. Dynamic pages should be archived with dates for reproducibility. [94] [93]
Passive demand is a possible amplifier, not issuer financing
FIELD NOTE Distinguish price transmission in secondary trading from cash delivered in the primary market.
The arithmetic of an index mandate
In a capitalization-weighted benchmark new net investment generally seeks exposure proportional to constituents’ weights. If all else is equal, a larger weight means more dollars of new consolidated portfolio demand. A company’s price rising also lifts its weight in an existing full-replication portfolio automatically; no new share purchase is necessary solely to keep that weight matched. Index inclusion, corporate action, reinvestment, rebalancing and fund inflow are distinct operations. ETFs also trade between existing investors without necessarily creating fresh underlying demand. [95] [96]
The mechanical allocation claim can be true for marginal net inflows without the stronger claim that passive investing creates a one-way price ratchet. Net flows can reverse; other investors supply or absorb trades.
What issuer cash can and cannot do
Secondary-market buying normally pays the seller of the share, not the AI issuer. The issuer can benefit indirectly from a higher valuation by issuing securities later, using stock as acquisition currency, or improving perceived financing access. That bridge requires a primary issuance, explicit collateral channel or another financing action. Do not add the market-cap uplift to cash available for GPU purchases. [96] [98]
[*REDACTED*], [*REDACTED*] and [*REDACTED*] provide a model and supporting evidence consistent with disproportionately strong effects of passive flows on very large firms under their conditions. ECB analysis identifies possible comovement and liquidity channels and notes competing considerations. Neither measures the proportion of today’s AI revenue funded by passive flows or identifies a universal gain each time a firm joins an index. [98] [97]
The common-ownership hypothesis and how to test it
FIELD NOTE Overlap identifies a screening question, not a finding of market harm.
Specify the proposed channel before measuring it
If an investor simultaneously owns rivals, the account may benefit if total industry margins rise rather than if one competitor grows by cutting price. For a changed company decision to follow, the incentives need a transmission mechanism: engagement, director selection, compensation oversight, express coordination, or a management response to pivotal holders. Passive tracking alone does not identify how objectives were transmitted. Multiple funds under one adviser can have different benchmarks, clients and voting directions. [102] [103] [87]
In the AI sector, define the market carefully. A chip designer and a cloud customer are vertically related unless particular products satisfy rival roles; equity held in both is not automatically horizontal common ownership. Two companies could be partners for distribution and competitors in a downstream inference market.
Why measured estimates disagree
Azar, [*REDACTED*] and [*REDACTED*]’s airline study associates higher modeled common ownership with excess ticket prices. [*REDACTED*], [*REDACTED*] and [*REDACTED*] instead attribute the earlier relationship to endogenous market-share components and sensitivity to control assumptions. The original authors’ subsequent reanalysis challenges the critique’s method for holding market shares fixed. DOJ economists’ separate replication reports materially different results under alternative ownership sources. None estimates competition effects in AI training, cloud procurement or accelerator pricing. [102] [103] [104] [105]
A persuasive AI test would combine ownership and actual stakes held, a product-specific outcome, dated influence or contract evidence, and an identified counterfactual controlling for compute shortages, regulatory developments and demand. A correlation between several portfolio companies’ returns is insufficient.
Index governance research does not speak with one voice
FIELD NOTE Different comparisons can legitimately answer different questions about influence.
Two empirical results, different interpretations
Appel, [*REDACTED*] and [*REDACTED*]’s index assignment approach uses variation around [*REDACTED*] index thresholds and links passive ownership to governance changes such as more independent directors and fewer defenses. [*REDACTED*] and coauthors instead report that index funds are less effective monitors than active funds in specified settings, with less opposition to management on contentious issues and weaker governance outcomes in their estimates. A later working paper by [*REDACTED*], [*REDACTED*] and [*REDACTED*] challenges aspects of the [*REDACTED*] specification. The direction of effect and causal identification depend on which shareholders are displaced, how index assignment is used, and what behavior is measured. [99] [100] [101]
What a credible claim about AI direction requires
Private stewardship contacts can be hard to observe. Published policies and fund N-PX establish some observable decisions. They do not reconstruct pressure behind contracting, financing or AI safety choices. Request evidence specific to the deal, item or board decision before assigning governance causation. In founder-controlled firms, even demonstrated investor advocacy faces hard limits from class rights. In dispersed firms, institution-led proposals or directors’ votes can matter while management retains discretion over supplier contracts. [86] [80] [82] [78]
Treat lower aggregate dissent as one observation about visible behavior. It could imply limited intervention or reflect prior coordination with a board through normal stewardship; neither inference identifies why an AI infrastructure plan was chosen. Cross-check with changes in ownership, exact vote margin, binding status, risk disclosures and deal chronology.
Enforcement depends on conduct and legal posture
FIELD NOTE A filed allegation, settlement and merits determination are different evidence.
Texas litigation illustrates the distinction
The states alleged that major managers used positions in competing coal companies in ways that harmed competition. DOJ and FTC’s May 2025 statement of interest addresses how antitrust law might apply to the alleged use of common shareholdings. Its position in the case is not a trial finding, and the underlying market is coal, not AI. [*REDACTED*] and participating states reached a February 25, 2026 settlement; the district court’s October opinion records dismissal of those [*REDACTED*] claims with prejudice on February 26. [106] [107] [108]
On October 30, the district court granted in part and denied in part [*REDACTED*] and [*REDACTED*]’s motion for partial judgment on the pleadings. Judgment on particular pleadings questions is not a finding of the whole alleged scheme. Present dates and claims at the appropriate procedural stage, and do not repurpose this record as evidence of misconduct at technology companies. [108]
A separate governance-policy proposal
The SEC’s September 2026 proposal would rescind the federal Rule 14a-8 shareholder proposal framework and make related changes to discretionary proxy authority under Rule 14a-4. The SEC page listed November 20 as the public comment deadline. As of November 19 this was a proposal, not adopted law. Changes could alter how investors bring issues to a vote, but are not evidence that current managers collectively determine AI permit or safety outcomes. [109]
Private antitrust liability analysis depends on the law, acquisition, market definition, conduct, effects, and legal posture. A 13F table may identify potential parties for investigation; it does not prove coordination or an illegal competitive effect. This research does not offer a legal conclusion.
Connect the portfolio layer to the financing layer
FIELD NOTE Evidence of overlapping exposures is stronger than evidence that holdings caused the financing structure.
Different maps, linked by a stress path
Issuer financing is a map of corporate investments, provider commitments, leases and cash payments. The index map is a map of constituents’ economic sensitivity. Link the two with observed counterparties and end-demand stress, not with an assumption that all managed fund positions finance capex. A deterioration in external monetization could weaken a buyer’s demand and the market value of an investor’s holding at the same time. Several correlated equities might then reprice in retail and pension portfolios even if firms never coordinate their policies. [97] [93] [98]
A reverse mechanism is possible: higher discount rates and equity weakness can make subsequent primary issuance more expensive. That can feed back into projects that rely on continued financing. It still does not turn benchmark purchases into equity subscriptions or convert a publicly traded holding into a control vote.
Indicators to keep separate
Track weighted exposure to the relevant issuer groups, common broker/counterparty or lease guarantees, next financing dates, and actual external utilization. Monitor index weight and price contribution separately from flows and creations; two common windows do not imply one caused the other. Benchmark diversification is sector-weight contingent; class-specific obligations and manager mandates are governance contingent. [96] [93] [64]
The strongest evidenced conclusion is that multiple portfolios can be exposed to related AI and technology valuation factors. Against this run competing factors: broad portfolios outside AI, diverse customer sectors, founder voting blocks, rapid changes in weights, market demand not supplied by strategic investors, and contractual protections. Keep these offsets visible rather than constructing a fully closed ownership-financing circuit.
An auditable ownership and voting scorecard
FIELD NOTE Publish evidence grades; record what would falsify the stronger claim.
Quantitative fields
Record CIK, accession, source document type, issuer and CUSIP, reported shares and authority fields, class shares outstanding, numerator and denominator date, and security class rights. For a vote record unique by meeting and proposal, add source of proposal, positions for or against, abstentions, shares on loan, broker nonvotes, management recommendation, issuer result and whether it is binding. Reconcile fund series to advisers and distinguish active, passive and unidentified strategy according to actual fund documentation. [64] [65] [86]
For benchmark risk, record source panel, as-of date, issuer-level consolidation and what business exposures are excluded from the analysis. Keep the units aligned: weights and market-value exposures can be aggregated cautiously, whereas votes require classes, mandates and turnout.
Falsification conditions
A reported collective governance claim is weakened by fund votes that split on the matter, end-investor authority, a controlling shareholder who can independently set the outcome, or failure to find conduct linking holdings to the alleged response. A risk transmission claim is narrowed by independent end-user demand, high diversification outside the stressed subindustry, or financing commitments that are unenforceable or untriggered at the stress date. Use a current covenant or filing to change the assessment, not anonymous descriptions. [80] [87] [103]
The measurement framework leaves several figures deliberately unresolved: national institution-wide estimates for named issuers without a complete reconciliation; the exact meeting turnout of the largest managers in 2026 in every company; and the fraction of AI capex dependent on secondary-market index flows. The point is not that these are unknowable in principle, but that the inspected filings answer narrower questions.
Exhibit 11 · Investment exposure and voting are distinct
Beneficiaries, funds, investment advisers and equity issuers are not interchangeable. Investment exposure, reported holdings, authority to cast voting instructions and votes actually counted must be checked separately. Share classes and turnout can change the result.
48PEGASUS Research conceptual exhibit. Not empirical data.
Exhibit 12 · Founder votes can exceed economic ownership
[*REDACTED*] disclosed [*REDACTED*]’s approximate 60.6% total voting power versus 13.4% economic interest through beneficial ownership of outstanding capital stock. The points answer different questions. The example does not measure the votes of any asset manager.
[*REDACTED*], October 2026 SEC disclosure: primary source
Exhibit 13 · The largest S&P 500 issuer groups carry weight
Displayed S&P 500 index weights in [*REDACTED*]’s index panel. [*REDACTED*] combines Class A (2.63%) and Class C (2.10%). Six issuer groups total approximately 31.48% on the displayed rounded weights; the six largest individual security lines instead sum to 29.38%. These are not percentages of ownership or votes.
[*REDACTED*], SPY index panel, November 18, 2026: primary source
Exhibit 14 · Four reported [*REDACTED*] positions are not four votes
Selected Form 13F filer rows for [*REDACTED*] common stock, excluding options, divided by the 7,426,426,719 common shares outstanding on the same date. Managers and subsidiaries are identified separately. These figures are not votes cast, economic stakes of the managers themselves, or complete institutional ownership. Individual filings and extraction method are in the ownership chapter.
SEC Form 13F public filings; [*REDACTED*] share count from the proxy statement. See the ownership chapter source register.
Browse: White Papers · Reports · Downloads
