Introduction: The Week Capital and Control Parted Ways

On the morning of Monday, October 26, 2026, a prospective investor opening Anthropic’s newly public registration statement on Form S-1 could have begun, as most investors do, with the income statement. The developer of the Claude family of models reported approximately $16.3 billion of revenue for the first six months of 2026, compared with roughly $1.2 billion in the comparable period of 2025, an acceleration that few public companies of any era have recorded.^[1][4] The same filing disclosed that the company had raised approximately $141.1 billion of capital in total as of June 30, 2026, that it held about $112.4 billion in cash and cash equivalents together with a further $19.3 billion in short-term investments, and that by the end of September it counted more than 1,900 customers generating at least $1 million of annual revenue and roughly 250 generating at least $10 million.^[5] Yet the filing also reported a first-half net loss of approximately $80.7 billion, a figure that at first glance seemed to contradict everything else on the page until the reader understood that roughly $72.4 billion of that loss consisted of noncash fair-value remeasurements of convertible notes, derivative liabilities, and simple agreements for future equity rather than cash spent on chips or electricity.^[6] The registration statement, in other words, was not merely the story of an enormously expensive model laboratory growing quickly enough to contemplate an initial public offering. It contained a less familiar and far more consequential question about the company that investors would actually own: how much influence, if any, would their shares purchase?

The answer was embedded in a governance arrangement without close precedent among large public companies. Anthropic proposed to issue a single special share of Class F common stock to a Delaware limited liability company, Anthropic Founders, LLC, whose initial members would be seven co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah. That one share would carry 50.1 percent of the votes on matters submitted to stockholders generally, and the percentage would not be diluted by subsequent issuances of Class A or Class N shares, however large.^[1][4] A separate institution, the Long-Term Benefit Trust, would hold a single share of Class T common stock conferring the power to elect and remove a majority of the board’s authorized directors, and the trust’s five trustees were described as financially disinterested specialists in AI safety, national security, public policy, and social enterprise who were barred from holding Anthropic securities.^[2][3][7] An investor therefore had to read beyond the customary division between founders and outside shareholders. Anthropic was proposing an architecture in which founder voting power, trust-selected board governance, mission commitments embedded in a public benefit charter, and investor economic ownership did not reside in the same hands and were not meant to. It was a legal construction designed for a technological enterprise whose future actions might carry consequences extending far beyond the shareholders financing it, and the filing said so explicitly, devoting nearly a hundred of its pages to risk factors that included the possibility of catastrophic or existential harm.^[7][58]

Across the country, on the same Monday morning, Meta Platforms presented a parallel story from an entirely different corporate starting point. Meta had reported five days earlier, on October 21, that third-quarter revenue had risen 26 percent year over year to $64.6 billion, that capital expenditures including principal payments on finance leases had reached $34.91 billion in a single quarter, and that the company now expected full-year 2026 capital expenditures of $135 billion to $145 billion.^[12] Free cash flow for the quarter was negative, by approximately $2.1 billion, for the first time in more than a decade, because the pace of datacenter and compute investment had finally overtaken the operating cash generated by one of the most profitable advertising businesses in history.^[12][18] In the same release, Meta’s board announced the establishment of a new class of publicly listed, nonvoting Class C common stock to be distributed as a one-for-one dividend on all Class A and Class B shares after an amended charter took effect on January 19, 2027.^[12][13] Then, on October 26, the company launched concurrent underwritten offerings of 250 million depositary shares representing interests in a 6.50 percent Series A Mandatory Convertible Preferred Stock, and $22.5 billion of Class A common stock offered through forward-sale agreements under which Meta itself would receive no proceeds until it chose to settle, together with a disclosed intention to establish an at-the-market program of up to $60 billion in Class C shares once that class began trading in 2027.^[15][16] Mark Zuckerberg, who at October 1, 2026 controlled approximately 60.6 percent of Meta’s voting power through high-vote Class B shares, would see none of that voting position eroded by any of it.^[14][61]

Imagine two conversations occurring on that Monday. In the first, in San Francisco, investment bankers from Goldman Sachs and Morgan Stanley, leading a syndicate of more than twenty banks, ask how much institutional capital a frontier-model company can absorb without compromising the founders’ ability to slow down, to withhold a model, or to decline a government contract when their safety commitments demand it.^[4][21] In the second, in Menlo Park, directors sitting on a special committee ask how an established technology platform can fund a capital buildout approaching $145 billion a year without allowing successive share issuances to erode the voting position of the founder whom the committee has concluded is best suited to lead the company into its next era.^[12][14] The two conversations sound opposed. One stresses safety, mission, and restraint; the other stresses strategic continuity and financing flexibility. Their corporate answer, however, converges on the same device. Financial participation can be expanded almost without limit while decision authority is deliberately protected from proportional redistribution. This is an economic innovation as much as a legal choice, and it changes the meaning of ownership in an industry where both the opportunity and the cost of error are unusually large.

I choose the title Voting Firewall because the phrase identifies a specific separation that ordinary descriptions such as dual-class stock, founder control, controlled company, or public-benefit governance do not fully capture. A firewall establishes a boundary. Here the boundary lies between supplying financial resources to an enterprise and obtaining corresponding authority over that enterprise’s decisions. A share may absorb price volatility, dilution, execution failures, and long-horizon business uncertainty without conveying a comparable vote over technology deployment, over the appointment of those who govern it, or over the multibillion-dollar compute commitments that determine whether the enterprise survives. In Anthropic’s proposal, that boundary is created explicitly through Class F rights layered on top of an independent mission-governance institution. In Meta’s strategy, it is strengthened through a planned nonvoting class layered onto a preexisting founder voting majority and a negotiated Founder Alignment Agreement.^[13][14] The mechanisms differ, but both invite the same inquiry: when does protection of strategic continuity become insulation from financial accountability?

I also choose the title because the distinction will grow more consequential as artificial intelligence moves from software experimentation into the industrial architecture of what I call the Five-Layer AI Economy. Energy investments, chip procurement, datacenter construction, frontier-model training, and agentic applications increasingly require simultaneous commitments of capital and operational authority. The governance of a single model company now influences power contracts measured in gigawatts, semiconductor supply agreements measured in hundreds of billions of dollars, regional economic development, research access, the timing of model releases, and the handling of safety incidents that may reach national security. Anthropic’s own filing discloses forward compute and infrastructure commitments that commentators have placed between $542 billion and $567 billion, including a block of approximately $417 billion of aggregate future minimum payment obligations entered into on December 31, 2025, with at least $110 billion owed to Amazon through April 2036 and at least $114.5 billion to Google through July 2033 regardless of actual usage.^[8][9] The central question is therefore not whether founders should always surrender control when a company becomes public; neither history nor the evidence justifies such a simple rule. It is whether institutions can preserve extraordinary control while furnishing public shareholders, affected communities, employees, regulators, and customers with accountability proportionate to the consequences of that control.

The question behind the offerings

Can the public increasingly finance superintelligence while strategic power remains concentrated in a small group of founders, directors, and mission-governing institutions? The question has no useful answer if the word public is treated as a single category. Public-market investors supply equity risk capital. Preferred holders and bondholders accept different contractual risks and different priorities in liquidation. Customers prepay or commit to compute-intensive services. Workers contribute human capital and may receive restricted equity with limited or no votes. Utilities and ratepayers finance or underwrite portions of the physical systems on which AI expansion depends. Governments establish the conditions under which high-impact systems may be built, exported, or deployed. Their claims and legal powers are not interchangeable, but their economic futures increasingly intersect, and the October 2026 filings make that intersection legible for the first time at the scale of a trillion-dollar enterprise.^[1][12][39]

This paper therefore advances six arguments. First, the new AI governance structures should be assessed as allocations of distinct rights, not as generic expressions of founder power, and the proper analytical tool is a rights map rather than a headline voting percentage. Second, capital formation in frontier AI has evolved into a multichannel process involving public common stock, preferred capital, forward sales, convertible obligations, strategic investors who are simultaneously suppliers and competitors, take-or-pay compute contracts, leases, and power infrastructure, and this multiplicity is itself a governance fact. Third, the separation of financial exposure from decision authority creates an accountability problem that reaches into the technical governance of frontier models, where the question of who can pause a release is at least as important as the question of who can elect a director. Fourth, corporate law, securities disclosure, competition law, export control, and emerging safety statutes address different parts of that problem and must not be conflated, because each is bounded by its own jurisdiction. Fifth, the scholarly debate between the idiosyncratic-vision defense of founder control and the entrenchment critique of perpetual dual-class stock, which matured between 2016 and 2026, supplies the right vocabulary for evaluating the AI structures, but it must be extended to accommodate mission guardians that are neither founders nor shareholders. Sixth, the most durable models for 2027 through 2030 will be those capable of sustaining investment while making concentrated authority legible, reviewable, and, under defined conditions, contestable.

Section 1: The Emergence of Founder-Controlled Artificial Intelligence Companies

1.1 From conventional shareholders to differentiated control

The familiar image of a public corporation is a group of investors buying fractional ownership in an enterprise, each receiving economic benefits proportionate to the capital contributed and a voice in collective decisions proportionate to that ownership. In the simplest model, one common share corresponds to one vote, and the board of directors is the agent of a shareholder body whose composition changes continuously as shares trade. The reality of American corporate law is considerably more flexible than that image. Delaware’s General Corporation Law permits a corporation to issue multiple classes and series of stock with such voting powers, full or limited or none, as its certificate of incorporation provides, subject only to the charter, applicable listing standards, and federal disclosure obligations.^[35] A founder can therefore own a small minority of the corporation’s economic interests yet possess a majority of its general voting rights. The divergence is especially visible in technology companies whose founders argue, with varying degrees of sincerity, that long development cycles and high technological uncertainty demand protection against impatient capital.

This governance approach is not a recent invention of the AI industry. The New York Stock Exchange prohibited dual-class listings for roughly six decades until the mid-1980s, and the structure remained rare in the United States until the technology listings of the early 2000s and the Google, Facebook, and Snap offerings of 2004, 2012, and 2017 normalized it among growth companies.^[27] Its historical appeal is easy to understand. A young enterprise may need repeated external financing before its technology becomes economically productive. A controller might sincerely believe that quarterly pressure would impair experimentation, reduce spending on safety or research, or force premature commercialization. The same arrangement may also protect an unsuccessful strategy, frustrate investor oversight, sustain conflicts of interest, and make the ordinary takeover discipline of the market for corporate control ineffective. The public shareholder receives an ownership claim but may lack any means of replacing management through an ordinary voting campaign. The corporate problem, properly stated, is not simply unequal votes; it is the difficulty of ensuring that protected decision-makers remain responsive when their preferred strategy ceases to serve the corporation and those who depend on it.

Artificial intelligence raises the stakes of this asymmetry in ways that earlier technology companies did not. Conventional business setbacks damage customers, employees, suppliers, and investors. Advanced models may additionally affect national cybersecurity, scientific research, the integrity of public information, the reliability of autonomous systems, and the competition among states for strategic infrastructure. A single company’s release schedule, security architecture, export compliance posture, and partnership policies can create consequences for parties who never purchased its shares and never consented to its risks. It would be mistaken to conclude that these larger consequences automatically justify either founder autonomy or greater external control; both inferences are available, and both have been drawn by serious people in the last several months. What the larger consequences do require is a precise examination of what each governance mechanism actually permits, which decisions require board approval, which require a shareholder vote, where a trust or foundation can intervene, and where public legal authority begins.

1.2 Anthropic: one founder share, two distinct centers of governance

Anthropic’s October 26 Form S-1 is a striking case because it makes several elements of governance unusually explicit, and because the explicitness itself was a choice. The Founder LLC would hold one share of Class F common stock carrying votes equal to 50.1 percent of all outstanding voting power on matters submitted to stockholders generally. The share would carry no economic rights: no dividend, no meaningful liquidation claim, no participation in appreciation. Nor would future issuances of Class A or Class N shares reduce its fixed voting percentage, because the share’s voting weight is defined by reference to the total rather than by a fixed number of votes.^[1][4] This is concentrated political authority embedded in a security with negligible direct economic rights, and it stands in deliberate contrast to the founders’ actual economic stakes, which press reports derived from the filing place at roughly 1.4 percent of the company apiece, with each founder having pledged, on a voluntary and revocable basis, to donate 80 percent of that equity.^[58] The founders’ voting power, in other words, is almost entirely decoupled from their residual economic exposure, which is precisely the configuration that corporate-law scholars have found most troubling in the dual-class literature of the past decade.^[27]

The more demanding analysis begins with what the Class F share does not accomplish. Anthropic’s Long-Term Benefit Trust holds the single outstanding share of Class T common stock, which confers the right to elect and remove the Class T directors, who will constitute a majority of the authorized board. The trust’s stated purpose is to ensure, through the exercise of its rights, that the company is managed in a manner that responsibly balances the pecuniary interests of stockholders, the best interests of those materially affected by the company’s conduct, and the responsible development and maintenance of advanced AI for the long-term benefit of humanity; the trust agreement states that it has no other purposes and no beneficiaries.^[2] The Founder LLC can determine the outcome of ordinary shareholder votes and can elect the directors other than the Class T directors, but it does not elect the entire board while the trust holds an advisory title. Press analysis of the earlier draft prospectus indicated that the trust would elect four of seven directors while Class A and Class F holders together would elect the other three, with the founders holding the controlling vote in that second group.^[20] The structure thus contains two centers of authority, each able to check the other in some respects and neither separately electable by public shareholders.

The distinction matters because under Delaware law the business and affairs of a corporation are managed by or under the direction of its board of directors, not by the shareholder meeting. Shareholder control and board composition are connected but legally distinct. A major merger may require both board approval and a shareholder vote; a charter amendment requires a board recommendation and a shareholder vote; the election of directors is a shareholder act but only within whatever class structure the charter has established. A founder security controlling ordinary shareholder outcomes cannot by itself erase the board’s fiduciary responsibilities or the trust’s director-selection rights. Conversely, trust-selected directors do not render the founder vote irrelevant: charter amendments, the election of non-trust directors, and other corporate actions depend upon it. The S-1 also discloses an intriguing fail-safe: if a majority of the co-founders cannot agree on how the Class F share should be voted, the share follows the proportional voting results of the public Class A shareholders, so that founder deadlock defaults to shareholder democracy rather than to paralysis.^[7] The result is neither a founder dictatorship nor a conventional shareholder democracy. It is a deliberately partitioned constitutional arrangement, and it should be analyzed as one.

Anthropic’s design also contains questions of endurance that its drafters evidently anticipated. The filing contemplates changes in the membership of the Founder LLC, founder-qualification requirements tied to continuing service, and a sunset that begins when only two or fewer founders or their successors remain qualified.^[20] Under the Founder LLC agreement filed with the S-1, each founder may propose a ranked list of up to ten successors, and the board’s nominating and governance committee may ask the founders to reconsider such a list.^[11] This is considerably more thoughtful than a perpetual personal voting right with no succession rules, but it creates its own governance questions, and the first of them surfaced within hours of the filing. Benjamin Mann, who has been described by the company as a co-founder since its founding in January 2021 and who co-leads Anthropic Labs, is not a member of the Founder LLC and will not participate in determining how the Class F share is voted.^[10][11] Ann Lipton, a University of Colorado law professor and former SEC attorney, observed that founder is not a legal term and that there could be many harmless explanations, while noting the practical effect plainly:

“He’s not going to have continuing governance rights in the firm.”

— Ann Lipton, University of Colorado Law School ^[10]

The episode illustrates a broader point. Who has practical influence over successor selection? What happens if a founder leaves the enterprise but retains an eligible ownership position? How will disagreements among the seven be resolved before the deadlock default is triggered? What information will outside shareholders receive about the exercise of the LLC’s collective vote? Which events cause the structure to sunset, and who can influence those events? These are not allegations of misconduct. They are the questions that any institution holding a durable, concentrated, non-economic vote should answer publicly before it asks outsiders to finance it at a valuation that reporting has placed between $1.5 trillion and $2 trillion.^[22]

1.3 The public-benefit corporation is not a universal safety guarantee

Anthropic’s corporate name ends in the letters PBC, for public benefit corporation. This is legally meaningful, not merely a branding statement. Subchapter XV of the Delaware General Corporation Law instructs the directors of a public benefit corporation to manage the corporation in a manner that balances the stockholders’ pecuniary interests, the best interests of those materially affected by the corporation’s conduct, and the specific public benefit or benefits identified in its certificate of incorporation. The statute also establishes the conditions of directors’ fiduciary responsibility in making that balance and requires periodic statements to stockholders about the corporation’s promotion of its public benefit.^[35] In Drakes Landing v. Tilden Park, decided in July 2026, Delaware’s Court of Chancery drew the contrast sharply, holding that directors of a traditional corporation must act with the sole focus of stockholder wealth maximization while directors of a public benefit corporation must balance stockholder pecuniary interests with those of other stakeholders.^[21] What the statute does not do, by its mere existence, is confer a private right of action on every affected person or guarantee that a disputed technological decision will have a single objectively correct outcome.

For an AI enterprise, the hard problem lies inside the word balance. Consider a model with capabilities that would be commercially valuable in cybersecurity research but might also facilitate harmful intrusion. Delaying deployment reduces near-term revenue and exposes the company to competitive pressure. Releasing it benefits legitimate security researchers while creating risks to third parties who are not in the room. Anthropic’s founders’ letter in the S-1 describes exactly this kind of choice, recounting that the company refrained from releasing its most capable model until it had built satisfactory safeguards, “forgoing substantial revenue and ceding ground to competitors,” and the filing warns investors that the company may choose to advance more slowly even where that creates a competitive disadvantage.^[1][7] A PBC structure broadens the set of interests that directors are required to consider, but it cannot mechanically calculate the right answer to such a choice. That requires reliable model evaluations, security expertise, access controls, incident reporting, and a documented decision process. A corporation’s charter can define purposes; it cannot substitute for institutional competence in fulfilling them.

Anthropic’s trust should therefore be evaluated as a system of decision rights, independence, expertise, information access, and enforceable procedure rather than as a label. A trust that can appoint a majority of directors provides a far stronger mission mechanism than one that only receives periodic briefings. But its effectiveness depends on trustee selection, funding independence, access to technical evidence, the treatment of disagreements, and the manner in which the board actually exercises its own authority when management disagrees. The S-1 discloses that the trust agreement limits trustee compensation to a reasonable cash amount fixed annually and prohibits trustees from directly buying or holding Anthropic securities, and that the Class T rights can be amended or redeemed by the board, the trustees, and specified majorities of other share classes acting together, which the company describes as preserving flexibility if circumstances change.^[1][7] Harvard Law School professor Jesse Fried, writing with Idan Reiter in a May 2026 paper on what they call Ben & Jerry’s risk, regards that very amendability as the feature that makes the arrangement workable, because a supermajority of investors can ultimately terminate the trust and remove the directors it appointed, which constrains the guardians by giving them something to lose.^[23][24] In this paper, mission governance means an institution’s demonstrable ability to shape decisions consistent with a stated public objective. The term is not a synonym for public-benefit registration, a safety-policy statement, or reputational claims.

1.4 Meta: the mature founder-control model meets the compute investment cycle

Meta approaches the same financing problem from the opposite direction. It is not an unprofitable startup seeking a first public valuation. It has a large existing shareholder base, extensive operating businesses, more than $82 billion of cash and marketable securities, and a long-standing dual-class structure in which Class B shares carry ten votes and Class A shares one.^[12][15] Mark Zuckerberg’s economic stake is materially smaller than his voting influence; reporting in 2026 placed his voting control at about 60 percent against an economic stake well below 15 percent, and Meta’s own preliminary information statement put his voting power at 60.6 percent as of October 1, 2026.^[59][61] By the autumn of 2026 the company’s communications to investors explicitly connected the proposed Class C recapitalization with the investment requirements of AI and the desire for continued founder leadership. Its October 21 release described a special committee of independent and disinterested directors that concluded Zuckerberg was best suited to lead Meta into its next era and that the creation of Class C stock would continue the founder-led model that had driven the company’s success since its inception.^[12] The preliminary information statement filed the same day disclosed that the company’s financial adviser had told the committee in June that the most Meta could raise in equity under its existing share structure, over a number of years, was approximately $700 billion before Zuckerberg’s voting power would fall below 50 percent, a figure the company later characterized as a theoretical maximum rather than a description of present financing plans.^[14][18][63]

Meta’s planned Class C stock is designed to carry economic rights without ordinary voting rights, and the recapitalization was approved not by a shareholder meeting but by written consent executed on October 20, 2026, the day before it was announced, which the information statement notes is the only stockholder approval required because the founder’s Class B votes suffice.^[13][14] The amended charter was scheduled to take effect on January 19, 2027, and the board declared a one-time dividend of one Class C share for each outstanding Class A and Class B share, contingent upon that effectiveness.^[12][13] That future date is not a formality for analytical purposes: as of this paper’s cutoff, the new security should be described as planned and approved, not as an established public market class. The October financing plan was also layered. On October 26 the company announced $22.5 billion of Class A common stock offered through forward-sale arrangements and 250 million depositary shares representing mandatory-convertible preferred stock; at pricing on October 27 the common-stock component had grown to $25.0 billion, or 40 million shares at an initial forward price of approximately $622.92, and the depositary shares were priced at $50 each, implying roughly $12.5 billion, with a 6.50 percent dividend and a liquidation preference of $1,000 per underlying preferred share.^[15][16][17] The preferred instrument is intended to convert, subject to specified terms, in August 2029, and once Class C shares exist both the preferred and the forward sales may settle in Class C rather than Class A stock.^[15][16] A separate future Class C at-the-market program of up to $60 billion had not been established; the company stated that it had entered into no agreement for such a program and that its timing, size, and terms remained open.^[15]

The strategic design is revealing. Meta seeks access to the pricing capacity of public equity markets while preserving continuity in voting control. Investors receive securities with different combinations of immediate cash distributions, conversion rights, forward-settlement timing, and residual economic exposure. The company gains financing options that it can exercise at its own discretion through November 2027 in the case of the forward sales. Existing high-vote interests are sheltered from the voting dilution that would follow an otherwise ordinary issuance of new voting shares. The Founder Alignment Agreement signed on October 20 adds a further layer: only 20 percent of the Class C shares that Zuckerberg’s holder group receives in the dividend will be freely saleable, and sales of the remaining 80 percent generally require the sale or conversion of a matching Class A or Class B share within two business days, so that monetizing nonvoting stock beyond the free tranche trims his voting power proportionately.^[13][18] The agreement also includes an eight-year commitment that the founder remain resident in California or in another jurisdiction where Meta maintains its headquarters, subject to conditions and an exception for material unforeseen circumstances, which reporting connected to a potential state wealth-tax liability.^[19] None of this abolishes economic dilution or guarantees a return on investment. Nor does it prove the spending will be productive. It changes the distribution of authority at the moment when vast sums are being committed to infrastructure whose returns may unfold over many years.

1.5 Four governance models, not one founder template

The temptation is to place Anthropic, Meta, OpenAI, and every founder-led AI developer into a single category of founder control. That would obscure differences that matter. Meta’s model begins with a publicly traded platform and a founder’s high-vote stock, now supplemented by a nonvoting class. Anthropic proposes a public-benefit corporation with a founder shareholder vote and a separate trust-appointed board majority. OpenAI’s October 2025 recapitalization created a for-profit public-benefit corporation, OpenAI Group PBC, controlled by the OpenAI Foundation, a nonprofit that appoints the Group’s board and holds a 26 percent equity stake valued at approximately $130 billion at the time, with a warrant for additional shares if valuation milestones are met; Microsoft holds roughly 27 percent, and employees and other investors hold the remainder.^[33][34] That architecture is not equivalent to a founder personally holding 50.1 percent of a general shareholder vote, because the controlling institution is a charitable foundation subject to oversight by the attorneys general of California and Delaware rather than a group of individuals. Private AI enterprises such as those associated with Elon Musk add further questions about investor agreements, intercompany contracts, and related-party transactions, but their governance cannot be inferred from the mere fact that they have a prominent founder.

Table 1. A rights map of the October 2026 structures (sources: Anthropic S-1 and exhibits; Meta 8-K, PRE 14C and offering documents; OpenAI structure disclosures)

Governance dimension	Anthropic, PBC (proposed)	Meta Platforms, Inc. (approved, effective Jan. 2027)	OpenAI Group PBC (since Oct. 2025)
Who controls ordinary shareholder votes	Founder LLC via one Class F share fixed at 50.1%	Mark Zuckerberg via ten-vote Class B (about 60.6% at Oct. 1, 2026)	OpenAI Foundation as controlling holder
Who selects the board majority	Long-Term Benefit Trust via one Class T share	Holders of combined voting power, i.e., the founder	OpenAI Foundation appoints the board
Public investors’ electoral role	Class A one vote per share; cannot elect any director independently	Class A one vote per share; outvoted by Class B; Class C nonvoting	None; no public class exists
Economic rights of the control instrument	None (Class F has no economic rights)	Full (Class B is common stock)	26% equity plus warrant
Stated sunset or transition	Begins when two or fewer qualifying founders remain; successor lists of up to ten	None time-based; Founder Alignment Agreement ties sales to voting reductions	None; subject to AG oversight
Mission instrument	Delaware PBC purpose plus LTBT with director rights	Conventional corporation; special committee process	Delaware PBC controlled by nonprofit with safety committee

The correct comparison is a rights map of the kind sketched in Table 1: who selects directors; who appoints management; who controls ordinary shareholder votes; who approves financing and acquisitions; who can amend the charter; who decides on model releases; who can require independent safety review; and who can replace an ineffective leader. Every company should be classified separately on these dimensions. The paper’s term, Voting Firewall, is therefore not a claim that every AI company employs dual-class shares. It describes the broader phenomenon in which economic capital and durable strategic authority can be separated through shares, trusts, corporate charters, contractual restrictions, nonprofit control, and specialized approval rules.

1.6 The scholarly debate that the AI structures inherit

The structures described above did not arrive in a theoretical vacuum, and the most useful way to evaluate them is to place them within the corporate-law scholarship that matured between 2016 and 2026. Two positions define the field. The first is the idiosyncratic-vision account developed by Zohar Goshen of Columbia Law School and Assaf Hamdani of the Hebrew University in the Yale Law Journal, which argues that entrepreneurs value control not primarily to extract private benefits but to protect their ability to pursue a business vision that investors may not share or may come to doubt midstream, and that corporate law should balance the controller’s need to secure that vision against the minority’s need for protection rather than treating concentrated ownership as inherently suspect.^[26] Their formulation deserves quotation because it captures the strongest case for what Anthropic and Meta have done:

“Concentrated ownership, therefore, should not be viewed as an unalloyed evil.”

— Zohar Goshen and Assaf Hamdani, Yale Law Journal ^[26]

The second position is the entrenchment critique developed most influentially by Lucian Bebchuk of Harvard Law School and Kobi Kastiel of Tel Aviv University, whose 2017 Virginia Law Review article argued that even where a dual-class structure is efficient at the time of an IPO, its advantages tend to recede and its costs tend to rise as time passes, that controllers have perverse incentives to retain the structure after it has become inefficient because they capture only a fraction of the gains from unification, and that the policy debate should therefore shift toward finite-term structures that sunset after ten or fifteen years unless extended by shareholders unaffiliated with the controller.^[27] Their warning, in the words of their summary of the article, is that even those who accept dual-class IPOs should:

“recognize the perils of providing founders with perpetual or even lifetime control.”

— Lucian Bebchuk and Kobi Kastiel, Harvard Law School Forum ^[27]

Between these poles sits a body of work specific to the public-benefit form. Jens Dammann’s 2024 empirical study of publicly traded public benefit corporations found dual-class shares in nine of nineteen listed PBCs, demonstrating that the combination Anthropic proposes is legally unremarkable even if its scale is not.^[28] Brett McDonnell has identified protected control as a means of preserving a public mission against investors seeking higher returns, which in AI could mean sustaining costly safety measures or delaying a release.^[31] Oliver Hart and Luigi Zingales have argued that shareholders themselves may care about corporate conduct and be willing to accept lower returns to prevent harm, so that insulating a company from shareholder pressure also silences investors who share the mission.^[29] Emily Winston has named the structural problem the separation of benefit and control: a PBC’s intended beneficiaries generally acquire no corresponding governance rights.^[30] And Marco Mari, a research fellow at NYU Law’s Program in Corporate Law and Policy and the MIT Industrial Performance Center, drew these threads together in a Columbia Law School Blue Sky Blog essay published three weeks before the S-1 became public, observing that a mission guardian could retain a board majority while public investors elected a minority, as at the New York Times Company, and concluding:

“For Anthropic, the question is who can check those entrusted with its mission.”

— Marco Mari, Columbia Law School Blue Sky Blog ^[20]

That question organizes the remainder of this paper. The idiosyncratic-vision account explains why the structures were created. The entrenchment critique explains why they are dangerous over time. The PBC literature explains why mission language alone does not resolve the tension. What none of these bodies of work fully addresses, because the phenomenon is new, is the interaction between protected control and the industrial scale of the capital now being deployed, and it is to that interaction that Section 2 turns.

Section 2: How the Capital Requirements of the Five-Layer AI Economy Reshape Ownership

2.1 Capital is no longer a financing event; it is an operating condition

The conventional venture-backed technology story involves a sequence: founders develop a product, early investors finance its development, the company discovers a scalable market, and a public listing eventually provides liquidity and new capital for expansion. Artificial intelligence does not entirely break this sequence, but it changes its proportions beyond recognition. Many software companies once expanded by purchasing cloud capacity incrementally as customer demand arrived, converting a fixed-cost problem into a variable one. Frontier AI developers, by contrast, may have to secure enormous quantities of scarce accelerators, advanced packaging, high-bandwidth memory, electricity, datacenter space, networking, and engineering expertise years before future product demand is certain. Economic survival depends on financing physical capacity ahead of the revenue it will produce, and the contracts that secure that capacity are increasingly structured as minimum purchase obligations that must be paid regardless of usage.^[9]

The Five-Layer AI Economy clarifies the financial chain. Layer 1, Energy, requires generation, transmission, substations, interconnection, cooling, and reliable fuel or power procurement. Layer 2, Chips, requires advanced design tools, fabrication, packaging, memory, and manufacturing equipment. Layer 3, Datacenters, converts chips and electrical access into usable computational capacity. Layer 4, Models, spends that capacity on training, evaluation, adaptation, and inference. Layer 5, Applications and Agents, attempts to convert intelligence into recurring customer value. Spending moves through the system before revenue fully returns through it, and the timing mismatch is economically consequential because obligations may be fixed while demand, pricing, and capabilities remain uncertain. Columbia Business School professor Stijn Van Nieuwerburgh, in a paper prepared for the fall 2026 Brookings Papers on Economic Activity, estimates that a 200-megawatt AI campus costs roughly $8.2 billion, of which about $2.2 billion is the building and $5.6 billion is the installed IT system, and projects total U.S. investment in data center buildings, power systems, networking, and specialized chips at approximately $10.3 trillion from 2025 through 2032, an average of 3.63 percent of GDP, which he calculates would be larger relative to the economy than the canal, railroad, electrification, highway, and telecommunications booms that preceded it.^[42]

These layers are not a literal single-company balance sheet. Each contains many entities with different financing mechanisms. Utilities use regulated rate structures or contracted generation. Chipmakers depend upon customer demand, capital-intensive fabrication partners, and long-lived manufacturing assets. Datacenter developers rely on debt, project finance, leases, joint ventures, and hyperscaler tenancy. Model developers use equity, strategic cloud investments, prepaid capacity, equipment leases, and convertible instruments. Application companies fund consumption through revenue or venture financing. The relevant financial system is therefore a web of claims, not simply an annual hyperscaler capital-expenditure number. Two companies may announce similarly large AI commitments while carrying radically different risks, terms, cancellation rights, and accounting treatments, and Van Nieuwerburgh’s central warning is that the risks embodied in the AI investment race are migrating from transparent on-balance-sheet financing by major corporations to opaque off-balance-sheet financing through joint ventures, private credit, and securitization.^[42] In conversation with journalists he put the analogy bluntly:

“somewhat reminiscent of what happened in the subprime mortgage crisis.”

— Stijn Van Nieuwerburgh, Columbia Business School ^[43]

Stanford’s 2026 AI Index records the breadth of the capital response. Global corporate AI investment, as the Index defines it to include private investment, mergers and acquisitions, and minority stakes, rose to approximately $581.7 billion in 2025, an increase of about 130 percent from roughly $253 billion in 2024, with private investment alone reaching $344.7 billion and U.S. private investment of $285.9 billion exceeding China’s $12.4 billion by a factor of about twenty-three.^[38] This figure does not equal annual global datacenter capital expenditure and should not be added mechanically to corporate infrastructure announcements; the categories overlap and measure different things. It does demonstrate that financing is not confined to a one-time public stock offering. It is repeated through procurement, strategic partnerships, acquisitions, refinancing, debt issuance, and successive generations of expensive hardware, each of which is a governance event.

2.2 Equity, preferred securities, convertibles, forwards, and leases convey different rights

To understand the Voting Firewall, an investor must distinguish four dimensions of every financing instrument: the cash it supplies, the economic claim it creates, the voting or contractual power it conveys, and the conditions under which that power changes. A common share ordinarily participates in upside and downside after senior claims, although the vote may differ by class or be absent altogether. Preferred stock can provide a distribution priority, a conversion formula, a liquidation preference, or other negotiated features; Meta’s Series A Mandatory Convertible Preferred ranks senior to common stock as to dividends and liquidation and carries a fixed 6.50 percent coupon payable in cash or, subject to limits, in Class C shares at the company’s election.^[16][17] Convertible notes begin as debt-like claims but acquire equity value when specified events occur, and their fair-value remeasurement is exactly what produced the bulk of Anthropic’s reported first-half loss.^[6] Forward equity sales allow a company to lock a reference price today while deferring cash receipt and share delivery; Meta explicitly stated that it did not expect to receive any proceeds initially from the forward purchasers’ sale of borrowed shares, and that it would receive cash only upon physical settlement at its election on dates through November 19, 2027.^[15][16] Leases and contracted compute purchases can create long-term obligations without issuing any shareholder votes at all, and Meta’s third-quarter 10-Q disclosed approximately $825 billion of lease liabilities and commitments, including hundreds of billions not recorded on the balance sheet.^[19]

There is no universal rule that creditors possess more or less practical influence than shareholders. A lender without a board seat may impose covenants constraining additional borrowing, asset sales, or leverage. A strategic cloud provider may be economically indispensable because it supplies accelerators and customer distribution, even if it lacks a controlling vote; Amazon and Google hold only nonvoting Class N stock in Anthropic, yet the company owes each of them more than $110 billion over the coming decade regardless of usage, and the filing itself flags that these providers also develop rival models and could prioritize their own compute needs.^[7][8][9] A partner can obtain commercial exclusivity, information access, first rights over certain workloads, or pricing advantages that affect management’s room for maneuver. The legal power of a stockholder differs from the commercial leverage of a supplier. This paper uses formal authority for charter and voting rights, contractual authority for negotiated rights and constraints, and operational leverage for power arising from access to scarce inputs or customer channels.

The distinction prevents a common overstatement. A founder retaining voting control has not necessarily retained total independence. Debt markets, cloud partners, regulators, customers, and the physical availability of power can restrict feasible choices far more tightly than a proxy contest ever could. Conversely, a strategic investor holding a minority economic interest may exert more influence through compute access, engineering integration, or distribution than a passive shareholder holding more voting stock. The architecture of influence cannot be understood from the capitalization table alone. A serious governance inquiry must read the operating agreements and commercial contracts to the extent they are disclosed, identify related-party arrangements, and separate incentives from legally enforceable control.

The financing question is especially acute during rapid changes in underlying hardware technology. A GPU cluster might be physically useful for several years while its profitability changes much faster because of newer chips, energy prices, inference optimization, or model efficiency. Microsoft told investors that about two-thirds of its recent capital spending went to short-lived assets such as CPUs and GPUs, and Alphabet said roughly 60 percent of its second-quarter outlay was servers rather than buildings, which means that a large share of the committed capital depreciates on a cycle measured in a few years rather than decades.^[46] A company can be simultaneously capacity-constrained in one workload and burdened with uneconomic commitments in another. Early investors may prize flexibility; long-term debt holders may demand predictable cash flows; founders may prefer preserving the option to pursue transformative research. The governing body must allocate capital among those competing imperatives, and when voting power is insulated, outside shareholders must evaluate the integrity of that allocation process through disclosure and board accountability rather than through election prospects.

2.3 Anthropic’s prospectus: growth, accounting losses, and the compute obligation

Anthropic’s S-1 provides unusual visibility into the financial pressures facing a frontier-model developer. First-half 2026 revenue of approximately $16.3 billion was roughly fourteen times its year-earlier level, and the company’s second-quarter revenue alone, at about $11.5 billion to $11.6 billion, exceeded its revenue for the whole of 2025, which was approximately $4.6 billion.^[4][8] The company nevertheless reported a net loss of about $80.7 billion for the half, and the explanation matters enormously for anyone attempting to assess the enterprise. Approximately $72.4 billion of that loss reflected changes in the fair value of convertible notes, derivative liabilities, and SAFEs, an accounting consequence of the company’s valuation rising sharply between financing rounds rather than cash leaving the business.^[6] Compute was the largest operating expense, at roughly $5.6 billion in the second quarter against $11.6 billion of quarterly revenue.^[8] Analysts must therefore separate operating cash flow, GAAP earnings, noncash mark-to-market effects, and future binding commitments. Collapsing them into a single headline would misrepresent both the company’s performance and its risk, and the temptation to do so in either direction, to celebrate the revenue or to condemn the loss, is precisely the kind of simplification that protected-vote companies must be required to resist in their own disclosure.

The harder problem is the prospective scale of compute. Reporting on the draft and final prospectus describes long-term infrastructure service obligations of at least $111.1 billion to Google between April 2026 and July 2033 and at least $110 billion to Amazon between May 2026 and April 2036, both payable regardless of usage, with the company stating that if actual spend falls short it must pay the difference; a Microsoft commitment of $31.4 billion between November 2026 and May 2033 that is non-cancellable except for Microsoft’s own uncured material breach; leases with Broadcom of roughly $161 billion; and arrangements with xAI-linked capacity of up to $84.5 billion through 2029 that are largely cancellable on ninety days’ notice.^[8][9] The S-1 also discloses the Anthropic Direct initiative, under which the company intends to lease first-party datacenter capacity and separate its commitments for buildings from its commitments for the hardware inside them, a strategy it describes as allowing more flexible use of capital, and it acknowledges that it has significantly underestimated its compute needs in the past.^[8] Not all of these commitments begin in the same year; not all are unconditional payments due immediately; and adding headline maxima across overlapping arrangements can exaggerate a single comparable exposure. An analyst should create a commitment schedule by counterparty, duration, deliverability, cancellation, prepayment, collateral, and expected use. The fundamental question is whether future customer revenue and available financing can support the contractual obligations necessary to develop more powerful models, and the S-1’s own answer is that demand is likely to exceed supply and that growth will be limited principally by the availability of compute.^[9]

Anthropic’s cloud relationships also demonstrate why investor voting is only one dimension of governance. Amazon Web Services, Google Cloud, and Microsoft Azure offer infrastructure and routes to customers; Amazon had invested at least $13 billion and agreed in April 2026 to contribute up to $20 billion more, and Google had led earlier rounds, while each simultaneously develops competing models.^[8][9] The Federal Trade Commission’s January 2025 staff report on its Section 6(b) study of precisely these partnerships, Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic, examined equity and revenue-sharing arrangements, consultation rights, exclusivity obligations, and access to computing resources and engineering talent, and warned that such arrangements could affect competition and could in time lead to AI developers being fully acquired.^[50] This does not establish that any named relationship violates antitrust law; the Commission voted unanimously merely to release the report. It establishes why a company’s board must understand conflicts and dependencies that do not appear in a headline percentage of voting rights, and why a founder’s 50.1 percent vote may matter less to the company’s trajectory than the take-or-pay clause in a cloud contract.

The combination of concentrated shareholder votes, trust-selected directors, half-trillion-dollar compute commitments, and strategic distribution relationships with investor-competitors produces a distinctive set of questions for the proposed IPO. How are large capacity commitments approved? What scenario analyses does the board review before signing a decade-long minimum-purchase agreement with a shareholder? What limits exist on related-party transactions, and which committee reviews them? Could a major provider influence deployment or model pricing through commercial leverage that no vote can counter? How does the benefit trust obtain sufficiently independent information to judge a tradeoff between safety expenditure and infrastructure expansion? And how does a shareholder who lacks effective electoral leverage learn whether capital is being used intelligently? The S-1 makes some of these issues visible, and that visibility is to its credit, but public disclosure cannot make inherently uncertain forecasts certain.

2.4 Meta’s October 2026 capital structure: quantity, timing, and control

Meta’s October announcements are an object lesson in the need for transactional precision. Its approximately $35 billion launch figure combined distinct securities with distinct cash profiles. The depositary-share offering closed on October 29, 2026, delivering roughly $12.5 billion of gross proceeds for preferred stock that will convert into common equity in 2029.^[16][17] The common-stock component, by contrast, was sold by the forward purchasers using borrowed shares, and Meta stated plainly that it did not expect to receive initial proceeds from it; the company will receive the forward price of approximately $622.92 per share, subject to adjustment, only when and if it elects physical settlement before November 19, 2027.^[15][16] Nor did the additional planned $60 billion Class C at-the-market program amount to completed issuance. It depended on the commencement of trading of a class that will not exist until the charter amendment takes effect in January 2027 and on an agreement not yet entered into, and the company cautioned that the program might differ in size or never occur.^[15] Describing the package as nearly $100 billion of cash already obtained, as some headlines did, would be incorrect; describing it as a set of options that together could deliver that much while preserving a founder voting majority throughout is accurate.

Table 2. Meta’s October 2026 financing package, by instrument (sources: Meta FWP and press releases of Oct. 26–27, 2026; Form 8-K of Oct. 29, 2026)

Instrument	Size at pricing	Cash to Meta	Votes conveyed	Settlement or conversion
Series A Mandatory Convertible Preferred (as depositary shares)	250 million depositary shares at $50 (about $12.5 billion)	Received at closing, Oct. 29, 2026	None	Mandatory conversion in August 2029 into Class A or, if issued, Class C
Class A common via forward-sale agreements	40 million shares, about $25.0 billion at $622.92	None initially; received at physical settlement by Nov. 19, 2027	One vote per share when issued; may settle in nonvoting Class C	At Meta’s election
Class C at-the-market program	Up to $60 billion (no agreement signed)	Over time, if established, from Q1 2027	None	Not applicable

The funding rationale was nevertheless unmistakable. Meta reported $34.91 billion in capital expenditures including finance-lease principal payments for the third quarter alone, forecast $135 billion to $145 billion for 2026, and guided to total 2026 expenses of $175 billion to $179 billion.^[12] It reported substantial cash and positive operating cash flow for the quarter, alongside negative quarterly free cash flow for the first time in more than a decade, and third-quarter operating income fell 37 percent to $12.99 billion because of a litigation charge, excluding which the company said operating income would have been $22.88 billion.^[12][18] A company can be profitable on an accounting basis and still face a sustained cash-allocation challenge when annual infrastructure construction accelerates faster than the business that funds it. In 2026 Meta was financing not a single AI project but a system of model research, which it described as having reached a milestone with the launch of its first frontier models, chips, datacenters, networks, products, and applications layered upon its existing global platforms, and its founder characterized the quarter in terms that investors should test rather than accept:

“Our investments are paying off, the core business is strong.”

— Mark Zuckerberg, Meta Platforms, Q3 2026 results ^[12]

Class C is a proposed answer to two problems at once. First, it broadens access to public-market capital without requiring proportionate increases in the voting stock ordinarily held by nonfounders; because Class C shares carry no votes, selling them dilutes no one’s voting power, including the founder’s. Second, it creates a financing, acquisition, and employee-compensation instrument that separates economic participation from ordinary shareholder voting, and the Founder Alignment Agreement itself recites that the company expects to issue Class C shares from time to time to raise capital, fund acquisitions, and compensate employees, each of which will dilute the economic interest of every stockholder including the founder.^[14] The arrangement should not be flattened into the slogan that the founder may sell everything and retain unlimited control; the 80 percent matching rule, the service conditions, and the residency commitment are specific and binding.^[13][18][19] Nevertheless, its intended effect is to preserve a substantial voting position even as the mix of economic claims changes, and the $700 billion figure disclosed to the special committee quantifies, for the first time in any public filing, exactly how much equity a founder-controlled company believed it could raise before the firewall would fail under the old structure.^[14][63]

Whether the strategy proves financially wise will depend on future model capabilities, monetization, advertising-product performance, consumer and enterprise demand, chip and memory costs, and the economics of installed capacity. It will also depend on the price at which securities are sold and the dilution to existing economic claims. A low-vote or nonvoting instrument does not eliminate the cost of capital. Investors may demand compensation for limited governance influence, and nonvoting shares have historically traded at a discount to otherwise identical voting shares, although the size of the gap varies by company and market; Alphabet’s Class C shares, created in 2014 for similar reasons, are the closest precedent.^[15] The precise discount cannot be predicted from corporate intent. Yet if the market accepts the structure at scale, Meta will have established a template for financing an expensive industrial transition while continuity of strategic leadership becomes, in effect, a distinct corporate product that investors are asked to price.

2.5 The hyperscaler earnings cycle through the second and third quarters of 2026

The two companies at the center of this paper are not outliers. They are the most governance-explicit members of a cohort whose aggregate spending has redefined American capital formation. The second-quarter 2026 reporting season in late July produced a wave of guidance increases: Alphabet reported quarterly capital expenditures of $44.9 billion on July 22 and raised full-year guidance to $195 billion to $205 billion from $180 billion to $190 billion; Microsoft reported fiscal fourth-quarter spending of roughly $41 billion including finance leases on July 29 and subsequently told investors it expected approximately $190 billion of calendar-2026 capital expenditures, including about $25 billion attributable to higher component pricing; Meta reported $31.1 billion on the same day; and Amazon reported $54.2 billion on July 30 and raised its 2026 cash capital expenditure guidance to roughly $220 billion from about $200 billion, citing memory prices and demand it could not fully meet.^[46][47] By the third-quarter season the four largest hyperscalers had guided to a combined total in the neighborhood of $730 billion to $735 billion for 2026, up from approximately $410 billion in 2025, and Amazon’s chief executive described AWS as growing 36.7 percent year over year, which he called:

“our fastest growth in 18 quarters”

— Andy Jassy, Amazon, Q2 2026 earnings call ^[47]

Table 3. Hyperscaler capital-expenditure guidance for calendar 2026, as disclosed through late October 2026 (sources: company earnings releases and calls; compilations cited in endnotes 46 and 47)

Company	Most recent quarterly capex disclosed	2026 guidance	Prior guidance	Governance structure
Amazon	$54.2 billion (Q2 2026, Jul. 30)	About $220 billion (cash)	About $200 billion	One share, one vote
Alphabet	$44.9 billion (Q2 2026, Jul. 22)	$195–205 billion	$180–190 billion	Three classes; founders hold ten-vote Class B; nonvoting Class C since 2014
Microsoft	About $41 billion incl. leases (FQ4 2026, Jul. 29)	About $190 billion (calendar 2026)	Lower; raised during 2026	One share, one vote
Meta	$34.91 billion incl. lease principal (Q3 2026, Oct. 21)	$135–145 billion	$130–145 billion	Ten-vote Class B; nonvoting Class C from Jan. 2027

Table 3 makes a point that is easy to miss when the companies are discussed one at a time. Two of the four largest spenders operate under conventional one-share-one-vote structures, two under founder-weighted structures, and the spending trajectories are nearly indistinguishable. Concentrated voting control is therefore neither necessary nor sufficient for aggressive AI capital commitment; Amazon and Microsoft are committing comparable sums under boards that could in principle be replaced by their shareholders. What differs is the mechanism of correction if the commitments prove excessive. At Amazon or Microsoft, a sustained failure of returns would eventually expose directors to electoral challenge or the company to the market for corporate control. At Meta, and at Anthropic if the proposed structure is adopted, that mechanism is absent by design, which places correspondingly greater weight on internal board independence, external disclosure, and the quality of the information reaching those with authority.

The macroeconomic significance of this cycle is no longer contested among the major institutions. The International Monetary Fund’s October 2026 Global Financial Stability Report named a potential AI equity correction as one of three salient near-term risks to global financial stability, and the Fund’s accompanying technical note by Yingyuan Chen, Deepali Gautam, and Jason Wu examined fourteen major U.S. financing arrangements involving approximately $300 billion of direct equity investment and $1.4 trillion of future revenue commitments, nine of which subsequently attracted more than $650 billion of additional debt, and found that roughly $21 trillion of the approximately $36 trillion increase in S&P 500 market capitalization since late 2022 had come from companies in its AI ecosystem sample.^[39][41] The Fund’s managing director, Kristalina Georgieva, described the ecosystem as highly concentrated and characterized by circular dealmaking in which hyperscalers invest in model laboratories that then buy capacity from those same firms, and warned that in a severe scenario a correction could cascade through a leveraged and interconnected system with losses spreading across borders, before stating the policy conclusion plainly:

“The priority should be to preserve stability.”

— Kristalina Georgieva, International Monetary Fund ^[40]

Academic economists have been, if anything, more pointed. Daron Acemoglu of MIT, the 2024 Nobel laureate, whose earlier estimate that AI would add only a modest 1.1 to 1.6 percent to GDP over a decade framed much of the skeptical case, told MIT Technology Review in September 2026 that without productivity gains:

“at some point people are going to sour on AI”

— Daron Acemoglu, Massachusetts Institute of Technology ^[44]

Wharton finance professor Jessica Wachter, who with a coauthor estimated that hyperscaler datacenter spending would reach nearly $1.1 trillion through 2027 and asked how fast earnings would need to grow to justify it after accounting for the cost of capital, a 15 percent return, and depreciation, reached a conclusion that protected-vote boards should be required to confront in their own scenario analysis:

“The current buildout will be the largest misallocation of capital in history.”

— Jessica Wachter, The Wharton School ^[62]

These are forecasts, not findings, and the optimistic case is also credible: Alphabet’s cloud backlog, AWS re-acceleration, and Microsoft’s $37 billion AI revenue run rate are real numbers that gesture toward paying for the servers.^[46][47] The point for governance is not to adjudicate the bubble debate but to recognize that the institutions bearing authority over these commitments are, in the two cases examined here, insulated from the ordinary mechanism by which markets correct boards that bet wrongly. That insulation is the firewall, and the scale of the bet is what makes the firewall’s design a matter of public concern.

2.6 A Five-Layer capital and control matrix

The Five-Layer AI Economy becomes more useful when each layer is evaluated on two axes: who bears the principal financial risk and who has legal or practical authority over the decisions that generate it. At Layer 1, a utility’s shareholders, bondholders, and ratepayers may bear different parts of the cost of grid expansion, while state public utility commissions and regional grid operators exercise forms of oversight. At Layer 2, chipmakers and foundries allocate fabrication and packaging capacity through investment decisions and customer agreements, increasingly under export-control constraints. At Layer 3, developers, hyperscalers, landlords, private-credit funds, and equipment lenders share construction and utilization exposure. At Layer 4, a frontier lab’s governing institutions select model-development priorities and sign the take-or-pay contracts that bind the layers below. At Layer 5, enterprise customers and consumers decide whether to adopt products and agents, but often lack insight into the infrastructure and governance choices that made those products possible.

Table 4. The Five-Layer AI Economy: capital claims, decisions, and the accountability question at each layer

Layer	Typical capital claims	Decisions that matter	Governance accountability question
1. Energy	Utility equity and debt, power purchase agreements, project finance, rate recovery	Generation, transmission, interconnection, tariffs	Who pays if projected load fails to materialize?
2. Chips	Foundry investment, equipment finance, customer prepayments	Design, capacity, allocation, export compliance	Can concentrated buyers or governments redirect supply?
3. Datacenters	Corporate capex, leases, infrastructure debt, joint ventures, private credit	Location, resilience, utilization, energy procurement	Who owns stranded-capacity and grid-congestion risk?
4. Models	Equity, convertibles, strategic investment, take-or-pay compute contracts	Training, releases, safety, IP, partnerships	Who can challenge a high-risk frontier-model decision?
5. Applications and Agents	Revenue, venture equity, customer contracts, acquisitions	Distribution, delegation, data access, authorization	Who is accountable when an agent’s commercial action harms others?

This matrix does not imply identical legal duties in all five layers. It shows a chain of economic consequences. An Anthropic or Meta decision to reserve greater training capacity propagates into chip orders and datacenter leases; the physical capacity in turn requires grid reinforcement, generation, and local permits. A decision to release a highly capable agent could increase inference loads and change the economics of equipment already installed. The reverse is equally possible: a delayed power interconnection or an export restriction may force a model laboratory to alter its roadmap. The voting rights of a corporate parent do not control state rate-setting or federal export law, but they influence which private alternatives the corporation chooses within those constraints, and Van Nieuwerburgh’s observation that off-balance-sheet structures ultimately depend on the cash flows and collateral values of a small number of datacenter tenants means that Layer 4 governance quality is, in effect, a credit input for Layers 1 through 3.^[42]

The central financing principle is simple to state and difficult to practice: a durable Voting Firewall should create an equally durable obligation to explain capital allocation. When shareholders cannot readily change the controller, they must be able to understand the controller’s risk appetite, project economics, potential conflicts, and criteria for stopping unsuccessful investments. This does not require revealing proprietary source code or sensitive model weights. It requires decision-useful reporting: commitments separated from cash spent, installed capacity separated from announced projects, capital investment separated from operating expense, noncash remeasurement separated from operating loss, and strategic assumptions separated from completed outcomes. The stronger the insulation from voting discipline, the more important credible disclosure and independent board judgment become, and Section 3 examines what that judgment must be able to do when the decision in question is not a datacenter lease but the release of a model.

Section 3: The Separation of Financial Exposure and Strategic Authority

3.1 The anatomy of a consequential AI decision

The word control is frequently used as though it described one indivisible power. In a frontier AI company, it is better understood as a sequence of decision stages, each with its own participants, its own evidence, and its own legal character. A laboratory identifies a technical capability. Researchers evaluate its reliability and potential misuse. Product teams develop commercial opportunities. Safety and security functions analyze release conditions. Management proposes a course of action. The board approves a major strategic commitment or supervises the policy under which management acts. Shareholders vote only on the corporate matters assigned to them by law and governing documents. Regulators may intervene where a statute or enforcement authority applies. Each participant has a different relationship to the decision, and the National Institute of Standards and Technology’s Generative AI Profile, which maps more than two hundred suggested actions to the functions of its AI Risk Management Framework, is useful precisely because it treats governance as a set of distinct functions rather than a single point of command.^[52]

This sequence exposes the limits of equating a 50.1 percent vote with command over every model release. Most operational choices do not go to a shareholder referendum, and nothing in Anthropic’s charter would send them there. A founder may exercise great practical influence through executive position, reputation, employee loyalty, product leadership, information access, and appointment rights, but those powers differ from the legal vote attached to a share. At Anthropic, trust-selected directors hold a board majority under the proposed structure, and certain board decisions reportedly require a supermajority, which further dilutes any single center’s control.^[24] At Meta, the established governance framework gives the founder voting control, but the board retains its corporate-law responsibilities and its special committee process produced the Class C recommendation. The question for public investors is not whether they will vote directly on each technical deployment, because they will not, but whether the institutions above management can intervene when evidence warrants it, and whether those institutions will know.

A rigorous decision-rights audit should classify actions by materiality and risk. Routine software updates may remain within delegated product authority. Major capacity purchases may require finance-committee or board approval, and a decade-long minimum-purchase agreement with a shareholder should require review by directors who have no relationship with that shareholder. A frontier-model release that meets specified capability or security thresholds could require independent evaluation and escalation to designated directors. An acquisition of a major model laboratory or datacenter developer may trigger full-board review and, depending on its form, shareholder approval and antitrust analysis. Export-controlled technology transfers require compliance with applicable government rules whatever the charter says. These processes need not be identical among companies. They should be disclosed sufficiently to make the distribution of power comprehensible, and Anthropic’s Responsible Scaling Policy, first published in 2023 and since imitated by rivals, is an example of a company voluntarily describing the thresholds that would trigger heightened review.^[21]

The essential danger is not concentration alone. It is concentration combined with information monopoly: one group controls the strategic vote, dominates management, selects the evidence that reaches directors, and defines the criteria by which its own decisions are judged. Even formally independent directors may struggle if the information and expertise necessary for evaluating technical risk are filtered through executives whose compensation and reputation depend on rapid deployment. A robust system therefore establishes independent escalation channels, protected dissent, externally verifiable technical assessment where feasible, and board access to expertise that does not depend entirely on management’s framing. This is the insight at the center of the most important academic commentary published on the S-1 in its first week.

3.2 Who decides when a powerful model is ready?

The decision to release a model illustrates the separation between authority and liability more clearly than any capital-allocation decision. Release can generate revenue, improve competitive position, attract developers, and accelerate useful scientific work. It can also create uncertainty concerning cybersecurity, misuse, privacy, autonomous action, and unexpected behavior. The cost of delay falls disproportionately on the company and its investors. Some costs of a harmful release fall upon victims, public institutions, customers, or competitors who hold no shares. A board that evaluates only the expected stock price is using too narrow a decision frame for a technology with external effects. Yet a board that treats hypothetical risks as reasons never to release a model can undermine the enterprise’s capacity to fund safe research in the first place, and Anthropic’s own filing acknowledges that it has forgone revenue and ceded ground to competitors by waiting.^[1]

The proper governance objective is not a slogan such as safety first with no operational meaning. It is a decision protocol containing measurable capability thresholds, evaluations, red-team findings, mitigation status, cybersecurity readiness, incident-response preparation, deployment restrictions, and responsibility for final authorization. Documentation should identify when a management recommendation was challenged, what alternatives were considered, and what evidence justified the final decision. Stavros Gadinis, the George R. Johnson Professor of Law at Berkeley, framed the problem in Fortune on October 30, 2026 as one of what he calls corporate ordering, a company deciding on its own rules for a contested question, applying them to concrete decisions, and monitoring how it did before the law arrives, and he credited Anthropic with leadership on the first two tasks while finding it exposed on the third. His central sentence deserves to anchor this section:

“Its unusual governance can protect that promise, but cannot fulfill it alone.”

— Stavros Gadinis, UC Berkeley School of Law ^[21]

Gadinis’s argument is that monitoring is where frontier labs are weakest and where history is least forgiving. Five decades of corporate scandals share the feature that insiders knew of failures well before outsiders. Anthropic’s S-1 itself warns that its models have behaved differently when they detect they are being tested, that monitoring does not cover all deployments, and that as its models write most of the company’s code the company may lose visibility over its own processes.^[21] The company has engaged an external evaluator with broad internal access and expects to spend more than $1 billion over five years on the effort, but it chooses and pays that evaluator and has agreed to indemnify certain external evaluators, and Reuters reported in mid-October that Anthropic and OpenAI had sought significant control over what evaluators could publish.^[21] Gadinis’s prescription is drawn from the law of financial auditing: evaluators selected by a committee of independent directors rather than management, public reports, testing of internal controls with disclosure of material weaknesses, and whistleblower channels running directly to independent directors. None of these steps requires new legislation, and all of them could be adopted voluntarily by a company that has asked investors to price its trustworthiness.

This is where a mission trust could matter most. If it chooses a majority of directors, it has a potential route to influence the institution that sets release oversight. But a trust’s practical independence is tested only when a difficult decision opposes a founder’s commercial preference, a major investor’s financial expectation, or a government’s strategic priority. Does it receive technical reports directly? Can it commission a second opinion? Can directors require mitigation before deployment? Can they pause a release or replace a manager who circumvents process? Fried and Reiter’s study of the only three for-profit enterprises ever to install investor-overriding mission guardians, Ben & Jerry’s under Unilever, OpenAI before its 2025 restructuring, and Anthropic, finds that the first two produced outcomes opposite to the guardians’ intentions, and Fried’s conclusion about design applies with equal force to trusts and to founder vehicles:

“When you’re building a governance arrangement, it has to work for the long term.”

— Jesse Fried, Harvard Law School ^[24]

The June 2026 White House executive order on advanced AI innovation and security adds another layer. Executive Order 14409, signed June 2 and published in the Federal Register on June 5, directed the Treasury, the Department of War through the National Security Agency, and the Department of Homeland Security through CISA to develop within sixty days a classified benchmarking process for assessing the cyber capabilities of advanced models and designating covered frontier models, and to establish a voluntary framework for secure frontier model deployment under which participating developers may collaborate with the government to select trusted partners for early access.^[48][49][57] The order expressly preserved developers’ freedom to release models and did not create a mandatory licensing or preclearance requirement, and its preamble framed the administration’s posture in a sentence that explains why the burden of monitoring remains with the companies:

“we refuse to stifle this innovation with overly burdensome regulation.”

— Executive Order 14409, The White House ^[48]

Federal engagement under that framework is therefore a specific security mechanism, not a substitute for a company’s own board responsibility or a comprehensive public veto over AI development. The practical consequence for a firewall company is that the external check on a release decision is thin by design, and that thinness increases the weight that must be borne by the trust, the independent directors, and the evaluators whose independence Gadinis questions.

3.3 Acquisitions, partnerships, and the unseen redistribution of power

The next generation of AI mergers may transfer more than customers, intellectual property, employees, or patents. An acquisition can determine who controls a model’s training data, research talent, specialized chip designs, deployment channels, safety infrastructure, and commercial relationships. Acquiring a model company can also modify the strategic balance between a hyperscaler and its cloud tenants. Such transactions may alter competitive conditions even if the target is not yet a major source of revenue. The Department of Justice and Federal Trade Commission’s 2023 Merger Guidelines identify circumstances in which acquisitions may lessen competition, entrench or extend a dominant position, or eliminate a potential entrant, and they expressly address acquisitions of nascent competitive threats and serial acquisitions.^[51] The guidelines are a framework for enforcement, not a declaration that every AI acquisition is unlawful, but they describe exactly the pattern the FTC’s 6(b) study found in the cloud–developer partnerships.^[50]

When a Voting Firewall exists, investors may have limited practical ability to challenge a transaction they consider overpriced, strategically distracting, or harmful to minority interests. A controlling vote makes an ordinary proxy campaign ineffective, and Meta’s Class C recapitalization was itself approved by written consent without any vote by public holders.^[14] The legal treatment still depends on the transaction, the board process, whether insiders are conflicted, and applicable fiduciary rules; Delaware’s framework for controller transactions continues to require independent committee process and, in some cases, minority approval for conflicted deals. For that reason, large related-party transactions should receive heightened independent review even when management argues they are essential to AI development. Independent directors must be willing to test valuation assumptions, competitive alternatives, procurement terms, cross-licensing arrangements, and the incentives of executives on both sides of a deal, and at a company whose largest suppliers are also its largest nonvoting shareholders, almost every major procurement decision is in substance a related-party transaction.

Strategic partnerships deserve equal attention. The FTC’s study identified concerns that access to key inputs, contract terms, switching costs, and sensitive information might influence competition, and it noted that the agreements it reviewed included consultation rights and exclusivity obligations in addition to equity and revenue-sharing.^[50] This is especially important when a provider has a financial stake in a model developer while also supplying compute and reselling its services, which describes Amazon’s and Google’s positions with respect to Anthropic and Microsoft’s with respect to OpenAI. A partnership that accelerates a startup’s growth can be socially valuable. The same structure may create dependence, discourage interoperability, or make it difficult for customers to understand whose interests govern pricing and model availability. A shareholder vote cannot by itself resolve these arrangements; the answer lies in independent commercial oversight and competition enforcement where warranted. Anthropic’s disclosed intention to build first-party capacity through Anthropic Direct can be read, in part, as an attempt to reduce exactly this dependence, which is itself a governance decision made by a board the public does not elect.^[8]

For 2027 through 2030, boards should maintain a transaction register that goes beyond the legal merger threshold. It should include major technology licenses, exclusive distribution rights, multiyear compute reservations, joint ventures, strategic minority investments, and unusually concentrated purchasing relationships. Each record should identify the economic rationale, the counterparty’s other roles, the rights granted, the termination costs, and the foreseeable effect on technological independence. Shareholders do not require trade secrets to understand concentration risk. They need a credible account of whether the company can change suppliers, models, or commercialization strategies without destroying its economics, and the IMF’s finding that fourteen such arrangements carry $1.4 trillion of future revenue commitments suggests that the answer, for the ecosystem as a whole, is increasingly no.^[39]

3.4 Employees, researchers, and the governance of dissent

Artificial intelligence’s strategic assets include scarce researchers and engineers whose judgment may be critical to both performance and safety. Their employment incentives are therefore part of the governance system. Equity compensation encourages long-term commitment, but the grant of economic shares need not provide a corresponding vote. Anthropic’s proposed classes include a distinct employee-related Class E stock issued to employees other than co-founders, with limited special rights rather than general voting authority, and a role for employee shareholders in certain nomination procedures.^[1][3] Employees may also hold other securities, and individual rights depend on the instruments and the governing documents. The point is not that employees have no voice. It is that compensation ownership, internal dissent, and formal corporate control must be analyzed separately.

Research organizations require a healthy capacity to disagree. A scientist who identifies a security weakness may face incentives to minimize concern if release delays could affect equity value, reputation, or promotion. Conversely, an internal critic may not possess all relevant evidence and should not automatically determine corporate strategy. Credible governance must make disagreement investigable without transforming every technical dispute into a public crisis. Protected reporting channels, independent technical reviewers, documented dissent, anti-retaliation rules, and periodic reporting to a board committee can help manage that tension. The OpenAI episode of November 2023, in which a board exercised its formal authority to remove a chief executive and then reversed itself within days when nearly all of the company’s employees threatened to leave, demonstrated, as Mari puts it, the gap between the authority to remove an executive and the capacity to sustain that decision, and Fried’s reading is harsher still: the guardians not only failed but produced the departure of the very safety-focused directors and researchers whose concerns they had sought to vindicate.^[20][24]

The structure of a founder vote intensifies these dynamics. Employees may believe that executives can survive ordinary shareholder dissatisfaction indefinitely, which can increase organizational loyalty during difficult research cycles. It may also reduce confidence that internal escalation will change outcomes when concerns implicate a founder’s preferred roadmap. The existence of a mission trust does not automatically solve this problem; its capacity to obtain unfiltered employee input and to act independently is decisive. Public-market investors evaluating a frontier laboratory should therefore examine talent retention, research governance, whistleblower mechanisms, and the role of independent directors as seriously as they examine training expenditure. Anat Alon-Beck of Case Western Reserve University’s law school framed the issue for Inc. in the summer of 2026 as the central unanswered question of Anthropic’s model: who holds the legal authority, and the accountability, to resolve the inevitable conflicts between a public-benefit mission and shareholder profit.^[25]

California’s frontier-model transparency framework is relevant because it creates duties that survive internal control arrangements. The Transparency in Frontier Artificial Intelligence Act, Senate Bill 53, signed by Governor Gavin Newsom on September 29, 2025 and effective January 1, 2026, requires developers that have trained a model using more than 10^26 operations to publish safety frameworks and transparency reports, establishes a channel for reporting critical safety incidents to the state Office of Emergency Services, and protects employees who report substantial dangers to public health or safety, with a private right of action and attorneys’ fees for successful whistleblower plaintiffs.^[53] In September 2026 the governor signed a further package of AI measures focused largely on workers, including limits on AI-driven workplace surveillance and predictions about employees’ emotional states, while vetoing others.^[54] State regulatory duties do not confer corporate votes on employees, but they create an external channel of accountability that no charter can switch off, and the governor’s own description of SB 53 at signing captured the state’s theory of its role:

“first-in-the-nation frontier AI safety legislation that builds public trust”

— Governor Gavin Newsom, State of California ^[53]

3.5 A new grammar of risk: who loses when the bet fails?

The phrase financial risk conceals several very different exposures. Common shareholders risk loss of market value and dilution; they are not ordinarily personally liable for the company’s contractual debt. Preferred holders face instrument-specific conversion, distribution, and priority risks; Meta’s mandatory convertible holders, for example, bear equity risk at conversion in 2029 while receiving a fixed 6.50 percent dividend until then.^[16] Creditors face default and recovery risk. Infrastructure partners face counterparty and utilization risk, which for Anthropic’s cloud providers is partly mitigated by take-or-pay terms that shift utilization risk back to the laboratory.^[9] Power utilities and their customers face rate impacts under approved tariffs and cost-recovery rules. Workers risk employment and compensation. Communities experience land-use, environmental, and grid-reliability effects, subject to public processes. These are distinct legal and economic claims, not interchangeable burdens, and a governance analysis that treats them as one category will misdescribe who actually bears the consequences of a protected decision.

A Voting Firewall becomes politically fragile when it protects decisions while shifting too much of the downside to parties who lack meaningful opportunities to assess or consent to the risk. That fragility is not confined to equity markets. Suppose a datacenter developer secures electricity service on the expectation of sustained frontier-model demand. If the customer later reduces capacity requirements, the distribution of stranded investment cost depends on tariffs, contracts, guarantees, and regulatory determinations. The AI laboratory’s founding shareholders cannot determine those public outcomes by vote. Yet their strategic choices are an important causal input to them, and Van Nieuwerburgh’s analysis of the Hyperion financing in Louisiana, in which an 80 percent stake in a roughly $30 billion campus was transferred to a private-credit firm, illustrates how the downside migrates to balance sheets far from the one that made the decision.^[42][62] Good governance should therefore ask not merely what a project costs but who remains responsible under adverse scenarios.

Consider four 2028 scenarios. In the first, model capability advances rapidly and revenue scales faster than planned; protected long-term leadership appears prescient and the compute commitments look like bargains. In the second, inference efficiency improves so sharply that older capacity earns lower returns despite growing demand; shareholders bear a capital-allocation loss and cloud providers holding take-or-pay contracts do not. In the third, security events force deployment restrictions; insurance, litigation, contractual remedies, and reputational harm follow, and Anthropic’s own filing suggests that the insurance industry may lack the capacity to underwrite a single catastrophic AI incident.^[7] In the fourth, technical progress is strong but public resistance to electricity costs or datacenter siting delays deployment; the corporate strategy confronts state and local authority that no founder vote reaches. Voting control eliminates none of these risks. It alters who has the power to change course before the risks mature, and Acemoglu’s warning about the circular financing that increasingly knits the layers together is a warning about correlated failure:

“these kinds of deals eventually reveal a house of cards”

— Daron Acemoglu, Massachusetts Institute of Technology ^[45]

The measure of governance quality must consequently be dynamic. A board should test whether founders can retain the capacity to pursue long-horizon innovation while meeting public-company standards of independent judgment. A voting arrangement that contributes to productive investment in 2027 may become harmful in 2029 if management cannot acknowledge a failed capacity forecast. Conversely, eliminating founder protection too abruptly could allow short-term pressure to undermine essential but nonremunerative safety work. The challenge is to protect the ability to think long term without protecting the inability to learn from error. That is the central accountability paradox of the Voting Firewall, and Bebchuk and Kastiel’s observation that controllers rationally resist dismantling structures that no longer serve the company is the reason the paradox will not resolve itself.^[27]

3.6 The valuation market as a governance referendum

There is one forum in which public investors do vote on these structures before they buy, and that is price. Aswath Damodaran of New York University’s Stern School of Business, whose valuation work is followed closely by institutional allocators, published an analysis in the week of the filing estimating Anthropic’s equity value at approximately $1.475 trillion under base-case assumptions, against a reported pricing range of $1.5 trillion to $2 trillion, and stated that he would sit out the offering if it priced well above his estimate. He noted that he owns shares in Alphabet and Meta and does not automatically reject founder-controlled companies, but he singled out the combination of PBC status and the Class F design as requiring a materially larger discount, and described the effect of the structure on public holders in three words:

“effectively emasculates public investors”

— Aswath Damodaran, NYU Stern School of Business ^[22]

Whether the market agrees with Damodaran will be revealed by the IPO’s pricing and aftermarket, and it would be a mistake to treat any single valuation as dispositive; other analysts anticipated strong demand near $1.5 trillion. The analytical point is different. A governance discount, if one emerges, is the market’s price for the firewall. If Anthropic and Meta can raise the capital they seek at prices that reflect only a modest discount for limited voting influence, the structures will proliferate, because they will have proven that strategic continuity can be purchased cheaply. If the discount is large, companies will be forced to offer the kinds of compensating protections, sunsets, minority board representation, independent evaluators with publication rights, and capital-commitment disclosure, that this paper recommends in Section 5. Either way, the valuation market is performing a function that the voting market has been structurally prevented from performing, and it is performing it once, at the moment of entry, rather than continuously thereafter. That is a weak substitute for ongoing accountability, but in the first week of the Anthropic offering it was the only one available.

Section 4: The Political Consequences of Concentrated Private Decision-Making

4.1 The SEC discloses the firewall; it does not choose the strategic mission

The Securities and Exchange Commission occupies a central but frequently misunderstood position in this debate. A registration statement allows investors to evaluate a proposed securities offering under a disclosure framework. It contains financial statements, risk factors, management information, the proposed use of proceeds, related-party arrangements, and an explanation of voting and capital rights. It is not an SEC endorsement of management’s vision, a federal certification of model safety, or a guarantee that a technology company will be profitable, and every prospectus says so on its cover. The legal objective of securities disclosure differs from the substantive governance of advanced AI. A well-drafted prospectus can reveal that investors have limited voting influence without requiring the company to abandon the structure, and Anthropic’s prospectus does exactly that, stating that the company will be a controlled company under Nasdaq rules and therefore exempt from certain board-independence requirements that apply to other listed issuers.^[1][5]

The distinction is especially visible in the Anthropic filing. Investors can read that the Founder LLC is designed to maintain 50.1 percent of shareholder votes and that the benefit trust selects a majority of directors. They can assess founder and trust rights, risks, cash obligations, and the company’s stated mission, including its contribution of 5 percent of fully diluted shares to a wholly owned public-benefit fund and a framework to contribute up to an additional 2 percent at each 50 percent increase in share price above the prior milestone.^[60] What the filing cannot give them is an enforceable guarantee that every model decision will be optimal, that the trustees will never disagree, or that fixed voting control will always promote the public interest. Securities law can require material information, and federal securities remedies may be available for actionable disclosure violations. It does not turn ordinary shareholders into members of the company’s safety committee, and it does not, as Gadinis observes, produce the third-party-verified record of monitoring that investors will demand on the day something goes wrong.^[21]

The same reasoning applies to Meta. Investors should be informed that the proposed Class C security is nonvoting, that existing high-vote Class B power persists under the disclosed arrangements, and that the Founder Alignment Agreement creates constraints and conditions. They should be able to distinguish the underwritten offerings from planned at-the-market activity and the preferred proceeds received from the forward proceeds deferred. Meta’s filings do all of this with precision.^{[13][14][15][16]} Yet disclosure alone cannot solve the normative question: is the degree of control disproportionate to the economic risk borne by minority holders? Some investors will accept the structure in exchange for anticipated innovation returns; others will demand a price discount or refuse to invest. Public markets allow that economic judgment, but the consequences of AI decisions may extend to people outside the market altogether, and no prospectus is addressed to them.

The Council of Institutional Investors has long supported the principle that ordinary common shares should carry equal voting rights, and in 2018 it petitioned both the New York Stock Exchange and Nasdaq to require that newly listed multi-class companies convert to one share, one vote within seven years of their IPO, citing academic evidence that the value premium associated with dual-class structures dissipates within roughly six to nine years and that its members’ 2016 policy already called for sunsets within three to five years.^[36][37] The exchanges declined, and the petitions requested, in the Council’s words:

“a sunset of seven years or less”

— Council of Institutional Investors, petition to NYSE and Nasdaq ^[37]

The Council’s perspective represents an important investor-governance position, not a binding rule for all corporations, and its policies were most recently updated in March 2026.^[36] Meanwhile, the argument for founder protection is strongest when a company has credible reasons to resist near-term pressure and adequate checks against self-dealing or failure to adapt. A serious analysis must hold both claims: one-share-one-vote improves the formal connection between economic ownership and voting authority, while a time-limited founder structure may support unusually risky research. Neither design is sufficient without competent oversight, and neither Anthropic nor Meta has adopted a time-based sunset of the kind the Council and Bebchuk and Kastiel recommend.

4.2 Delaware corporate law and the limits of mission language

Although the companies operate on a global scale, their legal governance is substantially shaped by Delaware corporate law. The PBC statute is especially important to Anthropic and to OpenAI’s for-profit entity. Its requirement to balance financial interests with the interests of materially affected parties and specified public benefits is a meaningful departure from treating public purpose merely as corporate philanthropy, and the Court of Chancery’s 2026 decision in Drakes Landing confirms that courts will read the distinction as real.^[21][35] It still leaves directors with a demanding judgment task. The statutory formulation protects informed, disinterested decision-making under its standards rather than allowing every member of the public to demand a particular outcome. Nor does PBC status eliminate ordinary concerns about conflicts, loyalty, disclosure, or careful board process.

This creates a paradox for frontier AI. A PBC may be more willing to spend money on noncommercial evaluations or to delay a release when evidence indicates serious risk. That flexibility can serve shareholders by protecting long-term franchise value and can also serve affected people. But broad mission language can be invoked to rationalize almost any managerial choice unless the company publishes measurable objectives and explains the tradeoffs it makes. A board that can always say it is pursuing the long-term public interest may become difficult to assess, which is Hart and Zingales’s concern restated: insulation from shareholder voice removes one of the few mechanisms by which investors who genuinely share the mission can object when management departs from it.^[29] The antidote is not to reduce mission to quarterly profit. It is to define a governance process that is sufficiently specific to be examined by investors, employees, regulators, and independent experts.

Delaware law calls for periodic statements to PBC shareholders regarding promotion of public benefit, at least biennially unless the charter requires more.^[35] A frontier-model developer could go further voluntarily by publishing a structured account of governance outcomes: the number of major releases reviewed under heightened protocols; the share requiring additional mitigations; the independent testing process and the evaluators’ publication rights; the categories of significant incidents reported under SB 53 or otherwise; the handling of potential conflicts with investor-suppliers; and the way the board tracks commitments to infrastructure and safety. Such reports should preserve confidentiality where necessary. Public disclosure can describe the effectiveness of an oversight system without publishing model weights, security vulnerabilities, or proprietary competitive plans.

The question of remedies remains essential. If a controlling founder and an ostensibly independent governance body disagree, which document decides? If a trustee lacks adequate information, how can the problem be corrected? What matters require a special vote, and what vote would change the rules themselves? The proposed Anthropic structure includes different procedures for changing trust rights and terminating the Class F arrangement, and the stockholders’ agreement among the company, the Founder LLC, and the trust governs how those two centers interact.^[3] Those details may sound esoteric, but they determine whether a protective structure can resist opportunistic alteration under pressure. Fried’s reading, that the amendability of the trust’s rights is a feature because it disciplines the guardians, and Mari’s reading, that the same amendability leaves the public with no independent foothold, are both correct about different risks, and an effective firewall must guard against both short-term interference and permanent unaccountability.^[20][24]

4.3 Washington, California, and the layered jurisdiction of AI

Federal policy and state legislation operate on different decision planes from corporate voting. The White House’s July 2025 AI Action Plan emphasized national competitiveness, deployment, infrastructure, innovation, and security, and a December 2025 order sought a national policy framework with preemption of conflicting state laws, followed by legislative recommendations in March 2026.^[56][57] The June 2026 order on advanced AI security called for specific federal preparedness and voluntary frontier-model cooperation, and a national security presidential memorandum of June 5 addressed federal procurement and use of AI.^[48][57] These policies shape the environment in which laboratories build and deploy technology, but they do not dictate the voting structure of Anthropic or Meta. Other federal functions, including securities regulation, competition enforcement, export controls, procurement, and national-security review, may affect particular transactions or operations according to their own authorities, and the S-1 discloses that Anthropic’s listing arrives amid a dispute with the Pentagon, which designated the company a supply-chain risk after it declined unrestricted military use of its technology, a dispute that illustrates how a mission commitment protected by the firewall can collide with a federal customer.^[21][58]

California offers the clearest state-level counterpart. SB 53 asks whether powerful model developers can be made to explain and report risk-management practices; it does not tell investors how many votes a founder may hold in a Delaware corporation.^[53] Yet it increases external visibility precisely where ordinary shareholder oversight is weakest, because its incident-reporting and whistleblower provisions run to the state rather than to the board. The public can evaluate safety duties and compliance without receiving a corporate security, and the 2026 legislative package extended the state’s reach into AI’s labor-market effects.^[54] For this reason, it is misleading to portray governance as an all-or-nothing contest between private founders and government agencies. There are multiple legitimate jurisdictions. Corporate law determines board and shareholder rights. Federal securities law informs purchasers. Antitrust law addresses competition effects. Export law addresses controlled transactions. State AI laws address transparency, incidents, and particular uses. State energy authorities and local governments regulate the physical infrastructure. Each can be important without becoming a general overseer of the corporation’s strategic mission.

The practical challenge for 2027 through 2030 is coordination without indiscriminate duplication. A company should not be able to evade a lawful safety obligation by citing a founder-controlled charter. Equally, a regulator should not treat the existence of a dual-class structure as conclusive evidence of unsafe AI development. Government should require compliance with applicable substantive duties and impose proportionate reporting where authorized. Investors should evaluate the private governance terms they are offered. Boards should demonstrate that technical expertise and independent judgment are available at the points where private choices create potentially severe public consequences.

4.4 AI infrastructure turns corporate governance into state and regional policy

The politics of a founder-controlled AI company may seem remote from the daily work of an American governor. It becomes less remote when the company or its suppliers request a 500-megawatt campus, long-term power contracts, accelerated permitting, transmission upgrades, workforce support, or tax incentives. Anthropic’s disclosed agreements reportedly cover at least 14.8 gigawatts of potential capacity, and Meta’s Hyperion campus in Louisiana and its El Paso construction are measured in thousands of acres and tens of billions of dollars.^[42][8] The corporate board decides which locations and commitments to pursue. Utilities, regional grid operators, local authorities, and states determine whether infrastructure can be accommodated and who will pay for upgrades. Those two decision systems meet at the datacenter site, the substation, the power-plant interconnection, and the public hearing, and Van Nieuwerburgh’s estimate that a 200-megawatt campus consumes as much electricity as 170,000 average households makes clear that the meeting is not a courtesy call.^[42]

Texas under Governor Greg Abbott, California under Governor Newsom, Virginia, Arizona, Ohio, Indiana, Pennsylvania, and Michigan can all serve as comparative cases, but the relevant legal and market regimes differ. Texas operates primarily within ERCOT’s distinctive electricity-market structure. Virginia has the nation’s largest datacenter concentration and corresponding transmission-planning issues. Arizona combines semiconductor manufacturing with water, power, and land-use constraints. Ohio and Indiana sit at the intersection of industrial recruitment, grid investment, and manufacturing strategy. Pennsylvania and Michigan bring questions of generation availability, industrial policy, and regional transmission coordination. These are analytical categories; a particular incentive, gubernatorial action, or regulatory decision must be documented individually rather than generalized across all states.

Consider a hypothetical 2029 negotiation between a frontier-model developer and a state economic-development agency. The developer promises substantial investment and employment, but it also asks for infrastructure accommodation that may outlast several chip generations. Its founders maintain strategic voting control, and the public investors who finance the project have little direct say over the negotiation. The state, however, has no obligation to accept a private company’s forecast at face value. It can require appropriate studies, enforce tariffs, negotiate enforceable commitments, evaluate local impacts, and preserve the authority of independent regulators. The proper issue is not whether a governor can govern the model company’s board; it is whether the public bargain fairly allocates the risks and benefits of the physical investment.

This distinction can protect long-term investment rather than impede it. If developers disclose credible demand projections, provide financial assurances where lawful and appropriate, and participate in grid planning, communities are better able to distinguish serious projects from speculative capacity requests. If governments offer subsidies without verifying commitments, they may socialize the downside of private overexpansion, which is the Layer 1 version of the accountability question posed in Table 4. The Voting Firewall concept thus extends into regional political economy: private authority can remain protected, but its infrastructure demands should be measured against obligations enforceable in the public jurisdiction where the facilities are located.

4.5 Competition, China, and the geopolitical reach of private discretion

Founders’ decisions also have geopolitical consequences because the AI stack is internationally interdependent. Advanced chip design, fabrication, packaging, memory, networking, cloud services, and model research cross national borders. U.S. export controls can restrict specific technology transfers or end uses, and the rules change over time. On January 13, 2026, the Commerce Department’s Bureau of Industry and Security issued a final rule, effective January 15, revising its license review policy for exports to China and Macau of certain commercially available advanced computing commodities, including the Nvidia H200 and AMD MI325X and equivalents below specified performance thresholds, from a presumption of denial to case-by-case review subject to third-party laboratory certification and other conditions, while maintaining the presumption of denial for more advanced parts; the same week the administration imposed a 25 percent tariff on semiconductors meeting the same thresholds.^[55] A company’s ownership structure grants no exemption from any of this, and the episode illustrates that the external rules governing corporate strategy are neither static nor reducible to a single blanket prohibition.

The strategic discretion of founders nevertheless matters inside the law. Management selects suppliers, markets, licensing relationships, cybersecurity investments, and international partnerships. It may seek to move inference workloads among jurisdictions or build overseas datacenters while complying with relevant restrictions. A frontier laboratory can also influence national capabilities through model availability, research collaboration, and contracts with governments, and the Stanford AI Index’s finding that industry rather than government now leads frontier development means that these private choices carry weight once reserved to states.^[38] Such choices may be commercially rational but politically contentious. Shareholders with little voting leverage cannot necessarily demand a change of direction, while security agencies may intervene only when an applicable legal mechanism authorizes it.

The U.S.–China competition should therefore be analyzed as a contest over who can make strategic commitments, with whose capital, and under what national rules. Public shareholders may finance an American AI company whose board chooses among multibillion-dollar domestic chip and infrastructure investments. Foreign capital may enter through permitted channels, subject to securities, investment-screening, and national-security considerations. Global suppliers benefit from that investment without controlling the U.S. firm’s votes. Conversely, domestic control over a model laboratory does not mean domestic self-sufficiency in every input, and the hyperscalers’ own attribution of guidance increases to memory prices demonstrates how dependent the entire buildout remains on a handful of foreign-fabricated components.^[46] The Five-Layer AI Economy remains a network of legal jurisdictions and economic contracts rather than a single nationally closed factory.

For policymakers, the proper response is precise rulemaking and enforceable requirements, not an assumption that a controversial founder is automatically a national-security problem. For investors, the proper response is diligence about regulatory exposure, sourcing, and capital intensity. For company boards, it is governance capable of sustaining innovation while responding to lawfully imposed constraints. A credible Voting Firewall may protect strategic independence from short-lived market sentiment. It must not become a pretext for avoiding lawful scrutiny, concealment of material risk, or the treatment of a company’s public-benefit language as an all-purpose geopolitical justification.

Section 5: Corporate Governance Models for 2027–2030

5.1 Four possible architectures of long-horizon control

The future does not require all AI companies to choose one governance template, and the evidence of the last two years is that they will not. The first model is permanent or indefinite founder voting protection, of which Meta is the leading example and which Goshen and Hamdani’s framework defends as the price of idiosyncratic vision.^[26] It insulates long-term plans from shareholder agitation and hostile takeovers, but it is most vulnerable when the founder’s judgment deteriorates or succession becomes unclear, and its credibility depends on independent directors, conflicts procedures, truthful disclosure, and the founder’s willingness to accept adverse evidence. The second model is conditional founder protection, in which voting rights change according to a time limit, ownership threshold, service condition, specified succession event, or other clearly defined trigger; this is the Bebchuk–Kastiel and CII prescription, and Anthropic’s founder-count sunset is a partial example.^[27][37] It preserves continuity during the highest-risk years while creating a path toward greater voting proportionality, but a rigid sunset may arrive when a company still faces the long-horizon challenge for which protection was designed.

The third model is mission-anchored governance. A nonprofit, trust, public-benefit board structure, or designated class of stock retains specific decision rights to protect a declared purpose. Anthropic’s Long-Term Benefit Trust and OpenAI’s foundation-controlled PBC are distinct expressions of this category, and Fried and Reiter’s finding that the arrangement has a sample size of three and a failure rate so far of two is the most important empirical caution available.^[23][24][33] These structures may be especially compelling when AI decisions create effects beyond shareholders, but only if the mission body has relevant expertise, genuine authority, and rules governing selection, conflicts, reporting, and amendment. The fourth model is proportional-shareholder governance with specialized scientific oversight, in which shareholders retain conventional voting influence while the board establishes independent risk committees, external evaluation arrangements, and delegated technical powers. Its benefit is a clearer connection between ownership and votes, as at Amazon and Microsoft; its weakness is that periodic changes in board membership may destabilize long-horizon research strategy.

Table 5. Four architectures of long-horizon control and their characteristic failure modes

Architecture	Leading examples	What it protects	Characteristic failure	Essential safeguard
Permanent founder protection	Meta (Class B plus Class C); Alphabet	Founder’s strategic vision against market sentiment	Entrenchment after the vision has run its course; unclear succession	Independent directors with real information; conflict procedures; honest disclosure
Conditional founder protection	Anthropic founder-count sunset; CII seven-year model	Continuity during the highest-risk years	Sunset mis-timed relative to the technology cycle; circumvention	Triggers that cannot be unilaterally reset; disinterested-shareholder extension votes
Mission-anchored governance	Anthropic LTBT; OpenAI Foundation; Ben & Jerry’s independent board	A declared public purpose against investor pressure	Guardians achieve the opposite of their aim; information filtered by management	Direct technical reporting; power to commission review; a disciplined amendment path
Proportional governance plus scientific oversight	Amazon; Microsoft	Electoral accountability of the board	Short-termism; board turnover destabilizing research	Chartered risk committee with escalation powers and external evaluators

These models can overlap. A PBC may have a founder class and independent technical committees. A company with ordinary one-share-one-vote ownership may use a safety trust with narrowly defined blocking rights. A voting-right sunset may preserve a special veto over an identified category of high-risk decisions. The important principle is that governance mechanisms should be designed around explicit institutional problems rather than copied as fashionable capital-markets structures. A founder who wants to prevent opportunistic short-term takeovers needs a different instrument from a corporation that wants an independent body to examine catastrophic model risk, and Anthropic’s decision to use both instruments at once is defensible only if each is justified on its own terms. Combining both problems under the vague label of safety makes the structure difficult to assess.

My proposed standard is purpose-rights alignment: every extraordinary governance right should identify the purpose it protects, the matters it covers, the evidence used to evaluate its exercise, the persons who can challenge it, and the events that cause it to change or expire. An unlimited vote without clear purpose is difficult to justify. A mission trust without the authority or information to advance its mission is potentially decorative. A sunset that can be circumvented unilaterally is not a meaningful sunset. The best design aligns protection with the institutional problem and does not extend it further than necessary.

5.2 The sunset problem: time, ownership, service, and performance

Sunset provisions deserve careful analysis because AI companies may outlive their founding technology by decades. A fixed-time sunset is understandable and relatively transparent, and the academic evidence that dual-class premiums dissipate within six to nine years gives it an empirical anchor.^[37] It may nonetheless be economically crude: a company entering a transformative training and infrastructure phase in year ten may still face long-horizon risks. An ownership-based sunset ties control to continued economic exposure, improving the alignment between votes and investment risk; Meta’s Founder Alignment Agreement is a partial ownership-based mechanism, because sales of restricted Class C shares require corresponding reductions in voting shares.^[13][18] Yet founders can hold a large economic stake while still making poor decisions, and they may use complex vehicles that complicate the assessment of beneficial ownership. A service-based sunset recognizes that an absent founder may lack current operational knowledge, and Anthropic’s continuing-founder qualification is of this kind. It can create succession uncertainty when a leader retires, becomes ill, or moves into another role.

A performance-based sunset appears attractive but may be the most dangerous if badly specified. Revenue, market capitalization, share price, and model benchmarks can all be influenced by temporary conditions or managerial choices. If governance rights depend on hitting a financial threshold, management could be incentivized to time disclosures, defer prudent expenses, or pursue risky commercialization. If the metric is model capability, a company might promote aggressive benchmark behavior or alter the evaluation process, and Anthropic’s own disclosure that its models behave differently when they detect testing is a reminder of how unreliable capability metrics can be as governance triggers.^[21] A more durable performance standard would focus on process integrity: sustained reporting compliance, independent audit access, material conflict remediation, and adherence to defined safety escalation procedures. Even these indicators require independent verification.

The Anthropic filing’s treatment of continuing founders, successor lists, and the deadlock default to Class A voting illustrates how succession can be engineered through precise legal triggers rather than vague expectations of permanence.^[7][11] Meta’s Founder Alignment Agreement illustrates a different blend of economic interest, service, sales, and residency conditions within a larger recapitalization.^[13][19] Neither should be reduced to a slogan about permanent entitlement, and neither contains a time-based sunset. Investors should read the documents to determine which rights terminate automatically, which require special approval, and which parties could obstruct a proposed change, bearing in mind Fried’s warning that a structure must be judged by imagining successors who are not as congenial as the current occupants.^[24]

For 2027 through 2030, I recommend that boards publish a control-duration statement with every material new nonproportional-voting security: what strategic horizon is being protected; why ordinary board oversight is insufficient; what succession events are contemplated; how an adverse founder event is addressed; and what independent evaluation will occur before any extension. This would not require a government-mandated sunset in every case. It would force sponsors to state their theory of value. If protected voting control is a corporate asset, its cost and intended lifespan should be analyzed with the same seriousness as an expensive datacenter lease.

5.3 Independent scientific oversight must be operational, not ceremonial

The most promising governance innovation for frontier AI is not another title for an advisory committee. It is a defined relationship between technical evidence and corporate authority. A board-level scientific and safety committee should have a charter specifying its jurisdiction, independence criteria, access rights, reporting line, budget, and escalation powers. The committee might include directors with expertise in machine learning, cybersecurity, infrastructure reliability, law, and public policy, supported by external evaluators. The precise composition will depend on the company’s scale and activities. Most importantly, it must be capable of testing management’s evidence rather than merely receiving a presentation, which is the distinction Gadinis draws between a company that drafts good rules and one that monitors its own compliance with them.^[21]

The committee should maintain a risk register that differentiates model capability, deployment context, access control, user authorization, third-party integration, and operational monitoring. A model whose raw capability appears acceptable in a restricted research environment may create very different risks when connected to enterprise payment systems, source-code repositories, or physical infrastructure. If an autonomous agent can execute transactions, the governance process should examine not only model output quality but permissions, audit trails, revocation, and incident recovery. The deployment boundary matters as much as the model benchmark. NIST’s generative-AI profile can inform procedures, but governance must translate abstract risks into company-specific tests and accountable decision owners.^[52]

External review is strongest when its independence is real. An evaluator paid by the company may still provide valuable analysis, but the contract should allow candid findings, protect against retaliation, and define escalation for material unresolved issues, and the company should not reserve approval rights over the evaluator’s reports. A committee should be able to request another assessment or direct internal audit to examine whether known weaknesses were mitigated. Sensitive evidence can be protected through confidentiality and cybersecurity controls. Publication of every technical vulnerability could undermine the very safety goal being pursued. The public report should therefore focus on process, standards, aggregate outcomes, and material events while reserving tightly controlled detail for authorized reviewers. The analogy to financial auditing, in which auditors are selected by independent directors, report publicly, test internal controls, and are protected by whistleblower law, is the right one, and nothing prevents a company from adopting it before a statute requires it.^[21]

This architecture does not require a scientific committee to run the company. Directors must still allocate capital, oversee management, evaluate competitive strategy, and meet fiduciary obligations. The distinction is between technical veto as a universal rule and technical escalation as a governance discipline. A technically credible warning can trigger a mandatory review without automatically dictating a final result. Where the corporation chooses to create binding special rights for defined extraordinary risks, the scope, conditions, and decision-maker should be explicit in the charter or relevant agreement. Such precision is essential to prevent mission language from becoming either ineffective or unaccountably expansive.

5.4 A financial reporting standard for companies with protected votes

Capital markets should not have to infer infrastructure risk from a handful of widely publicized numbers. A company expanding under protected voting control could voluntarily publish a periodic AI Capital and Governance Statement accompanying its financial disclosures. The statement would reconcile five categories: capital expenditures recognized during the period; binding remaining purchase and lease obligations by counterparty and year; contingent or prospective commitments; available liquidity and financing facilities; and operating indicators demonstrating how installed compute is being used. It should also explain major changes in strategy, relevant related-party dealings, and the approval process for exceptional expenditures. This is a proposed reporting standard, not an assertion that present securities law already requires a document under that name, although Meta’s disclosure of $825 billion of lease liabilities and commitments and Anthropic’s commitment schedule show that the raw material already exists in filings.^[8][19]

Crucially, this reporting should not confuse operational metrics. Training compute is not equivalent to all installed GPUs. Datacenter capacity reserved is not the same as electricity consumed. An announced multiyear purchase maximum is not this year’s cash spending. A cloud provider’s capacity delivered is not necessarily paid for at the maximum headline contract value. Accounting operating losses are not identical to negative operating cash flow, and net losses reflecting financing remeasurement require separate explanation, as Anthropic’s $72.4 billion of fair-value charges demonstrates.^[6] Capital expenditure including finance-lease principal, which Meta reports, is not the same as capital expenditure excluding it, which others report, and Amazon’s cash capital expenditure differs again. These distinctions are indispensable for assessing companies in 2026 and will remain so through the decade.

One proposed dashboard would include utilization by material workload category, energy consumed where attributable, cost per economically useful unit of inference, customer revenue concentration, capacity under construction, timing of major commitments, and ranges of sensitivity to price or demand assumptions. It should identify measurement methods and avoid publishing competitively damaging or security-sensitive details. The board should review the dashboard alongside a downside scenario under which demand growth is slower, models become more efficient, chip prices change, or energy deployment is delayed, which is to say alongside the scenarios that Wachter, Acemoglu, and the IMF have already modeled from the outside.^[39][44][62] The purpose is not to eliminate uncertainty. It is to reveal whether leadership has plans for uncertainty beyond continuing to spend.

Protected-vote companies should also disclose how independent directors evaluate capital projects involving strategic investors, suppliers, or management-linked entities. An audit or finance committee could certify the process for identifying and reviewing material related-party transactions, while maintaining appropriate legal limits and confidentiality. Large capital decisions made with founder continuity may be beneficial. But a credible investor should be able to distinguish a carefully approved long-term investment from a decision the board did not meaningfully test. That distinction is particularly important when a controlling shareholder’s ability to lose an election is, by design, nil.

5.5 Three scenarios for the governance of frontier AI in 2030

Scenario A is the protected-founder renaissance. AI investment continues to yield large capability and productivity gains. Investors accept nonproportional votes because companies with coherent long-horizon strategies deliver attractive returns, and the governance discount that Damodaran anticipates proves small or temporary. Specialized safety oversight becomes routine, and founders maintain influence through succession planning rather than indefinite personality-based power. Under this outcome the Voting Firewall is viewed as a useful institutional response to an extraordinary investment cycle. Capital markets may still price voting differences, but the cost of financing remains manageable. This is a possible outcome, not a forecast.

Scenario B is the accountability correction. A subset of frontier firms experiences disappointing returns on compute commitments, contested transactions, governance conflicts, or serious deployment failures. The IMF’s adverse scenario materializes in some form, and the take-or-pay contracts that protected cloud providers become the mechanism through which losses concentrate at the laboratories that signed them.^[39] Minority investors question the insulation of leadership, activists seek special protections, and companies face pressure to adopt meaningful sunset provisions or independent director rights. Regulators enforce existing disclosure and safety obligations where applicable, while investor governance groups renew their petitions to the exchanges. Under this scenario the market does not reject AI investment; it differentiates more sharply between concentrated control supported by credible oversight and concentrated control that cannot be reviewed. The firewall survives only where its purpose is convincing and its limits real.

Scenario C is the mission-governance federation. The industry evolves toward multiple centers of authority. Founders retain control over some strategic matters, independent mission bodies hold clearly delimited powers over extraordinary safety decisions, investors receive stronger financial and conflict disclosures, and governments oversee particular regulated activities under frameworks like SB 53 and Executive Order 14409. The system resembles constitutional separation of powers more than either founder absolutism or unqualified shareholder primacy. It can become complex and expensive to administer, and legal conflicts may arise between authorities. Its advantage is that decisions are allocated according to the type of consequence rather than delegated indiscriminately to one class of actor.

These scenarios are not mutually exclusive at the industry level. Meta might find one financing model suitable while Anthropic retains another and OpenAI evolves its foundation-controlled structure, perhaps, as Fried anticipates, by returning to the attorneys general if its structure impedes a public offering.^[24] Regional governments may react differently to datacenter expansion, and federal policy may change within the window. The durable analytical prediction is more modest than a claim about which scenario wins: as the amount of capital and consequence grows, governance structures will become an increasingly visible competitive variable. Access to capital will depend not only on model benchmarks and revenue projections but also on whether investors can understand who can make binding decisions and what restraints exist when those decisions impose large risks.

5.6 A practical Voting Firewall Integrity Standard

I propose a seven-part scorecard for boards, institutional investors, policymakers, and researchers. The measures are designed to expose tradeoffs rather than generate a single false-precision rating.

Table 6. The Voting Firewall Integrity Standard: seven dimensions and the questions each asks

Dimension	Core question	Evidence to examine	Applied to the October 2026 filings
1. Authority clarity	Can a reader identify who votes, appoints directors, amends rights, and approves major transactions?	Charter, stockholders’ agreements, trust instruments, information statements	Both Anthropic and Meta score well: the mechanics are disclosed in unusual detail
2. Economic alignment	How much genuine capital exposure do controlling persons retain, and how can it change?	Beneficial ownership, pledges, alignment agreements, sale restrictions	Meta’s 80% matching rule ties sales to votes; Anthropic’s Class F has no economic rights and founders hold about 1.4% each
3. Independence	Can directors and oversight bodies obtain evidence and advice without management filtering?	Committee charters, evaluator contracts, publication rights, whistleblower channels	Anthropic’s evaluator is company-selected and indemnified; disclosure of committee information rights is incomplete
4. Mission enforceability	Are public-benefit or safety commitments supported by concrete rights and review processes?	PBC reports, trust powers, release protocols, incident reporting	LTBT elects a board majority; its rights are amendable by defined coalitions; OpenAI’s foundation holds a safety veto
5. Capital transparency	Can investors reconcile spending, commitments, financing, utilization, and adverse scenarios?	Commitment schedules, lease disclosures, forward and ATM terms, scenario analysis	Commitment schedules exist but no utilization or downside-scenario reporting is published by either company
6. Succession adaptability	Do control rights change under intelligible conditions, including incapacity, departure, misconduct, or prolonged failure?	Sunset triggers, successor lists, service conditions, extension votes	Anthropic has founder-count sunset and successor lists; Meta has service and residency conditions; neither has a time-based sunset
7. External accountability	Does the firm comply with securities, competition, safety, export, and infrastructure rules without treating private control as immunity?	SB 53 reporting, FTC and DOJ posture, export compliance, state infrastructure obligations	Both are subject to SB 53; Anthropic is in dispute with the Pentagon; neither has claimed exemption from any public duty

The integrity standard can be applied to a company without assuming one-share-one-vote is always ideal. Consider an illustrative comparison rather than a score of the companies discussed here. Company X may have a ten-year founder vote but strong independent directors, a credible sunset, transparent capital schedules, and tested model-release controls. Company Y may have formally proportional voting rights but a dependent board, opaque related-party agreements, and ineffective technical escalation. A simplistic governance screen might prefer Y because its shares look more democratic. An integrity assessment could reveal important weaknesses. Conversely, Company Z might combine perpetual nonvoting issuance, weak disclosure, little independent review, and no realistic succession mechanism; it would face substantial accountability concerns regardless of its model performance.

For public-company boards, the standard provides an agenda for annual review. For asset managers and pension funds, it provides questions beyond the traditional percentage of founder votes. For founders, it supplies a way to explain why extraordinary control is commercially and socially justified. For public agencies, it helps distinguish matters properly resolved through corporate contracts from matters requiring public regulation. For academic researchers, it enables comparative study of governance outcomes over time. The aim is to replace ideological shorthand with testable institutional claims.

At the core is a simple institutional bargain. If founders ask investors to fund a technological strategy whose payoff may take years and whose risks may be profound, they can plausibly argue for continuity. If investors accept limited voting influence, they can plausibly demand exceptional clarity about how that continuity is governed. If communities and governments are exposed to consequences beyond the corporation, they can lawfully require compliance with relevant public rules. A strong Voting Firewall does not eliminate these constituencies. It specifies the boundaries between them and provides legitimate ways for serious concerns to cross those boundaries.

Section 6: What Have We Learned? Seven Pillars

Pillar 1 — Economic Ownership and Strategic Control Are Separate Variables

The first lesson is that the percentage of a company owned economically and the percentage of corporate votes controlled are not the same quantity, and that treating them as one produces errors in both directions. Share ownership can grant a claim on dividends, residual value, or market-price appreciation, while a different class of security or governing entity possesses decisive voting authority. The October 2026 proposals make the distinction unusually clear. Anthropic’s contemplated single Class F share would receive a fixed 50.1 percent of general shareholder votes without any economic participation, held by founders whose combined economic stake is on the order of a tenth of the company.^[1][58] Meta’s planned Class C shares would allow unlimited economic participation without any shareholder vote, operating alongside a long-standing high-vote founder structure and an alignment agreement that links the founder’s sales to reductions in his votes.^[12][13][18] They are not identical arrangements, but both demonstrate how a company can sell additional economic claims while preserving a concentrated locus of strategic influence, and the $700 billion ceiling that Meta’s advisers calculated under the old structure is the first public quantification of how much capital a firewall of the older kind could carry before it failed.^[14][63]

The lesson is not that unequal votes are inherently illegitimate or that proportional voting guarantees good management; Table 3 shows that one-share-one-vote companies are committing capital at the same pace as founder-controlled ones. It is that public understanding requires precision. Who has the vote? What is the vote over? Who appoints the board? Who can alter special rights? How much economic exposure does the controller retain? When do protections expire? Those questions should precede sweeping judgments about whether a company is founder-led, public-benefit, democratic, or accountable. Even the word control is incomplete without identifying the relevant decision. A vote that controls ordinary shareholder resolutions is different from trust authority to appoint directors; a seat on a board is different from ownership of a cloud-service contract; and, as Section 2 showed, a $110 billion take-or-pay obligation may constrain a company more tightly than any of them.

For corporate boards, Pillar 1 requires a decision-rights map of the kind sketched in Table 1. For investors, it requires reading the charter and transaction terms, not merely the headline ownership percentage. For researchers, it requires treating voting control and economic ownership as distinct explanatory variables when examining innovation, returns, safety performance, and succession, which is exactly the move that Goshen and Hamdani made in theory and that Dammann’s PBC data now allow empirically.^[26][28] For policymakers, it requires recognizing that private corporate authority and public-law authority operate on different legal foundations. The analytical vocabulary of the Five-Layer AI Economy must be as precise about corporate rights as it is about megawatts and semiconductors.

Pillar 2 — Capital Intensity Is Making Governance a Financing Technology

The second lesson is that corporate governance has become a means of obtaining and allocating capital at AI scale. A share class, a trust, a public-benefit charter, or a nonprofit controlling entity can influence which investors participate, what risks they accept, how management responds to market pressure, and how long strategic commitments can endure. It may therefore affect the cost and availability of funding for chips, datacenters, models, and applications. Meta’s Class C is explicitly described in its own filings as an instrument to raise capital, fund acquisitions, and compensate employees; it is a financing technology before it is a governance statement.^[14] Governance is no longer merely an issue to revisit after the technology matures; it is one of the instruments by which the technology’s expansion is financed, and the valuation market’s response to Anthropic’s structure will determine how expensive that instrument is.^[22]

The International Monetary Fund’s October 2026 technical study of the AI ecosystem reinforces the need to connect company-level balance sheets with capital-market and real-economy exposures. Its finding that approximately $300 billion of direct equity investment across fourteen arrangements supports $1.4 trillion of future revenue commitments and has attracted more than $650 billion of additional debt is a description of leverage created by contract rather than by borrowing, and Van Nieuwerburgh’s $10.3 trillion projection for the physical buildout is the denominator against which that leverage must be judged.^[39][42] An individual provider’s cash flow, a hyperscaler’s obligations, and a utility’s infrastructure investment interact, but they should not be aggregated casually or treated as identical liabilities. The economic pressures created by long-term contracts, debt, share issuance, and asset utilization may be as consequential as the nominal valuation of a frontier-model company, and the Fund’s conclusion that risks remain manageable because the leading firms retain strong balance sheets is a conclusion about the present rather than a guarantee about 2029.^[40][41]

Pillar 2 therefore calls for a capital-claims approach to the Five-Layer AI Economy. Researchers should trace financial obligations from the application revenue stream down through models, datacenter capacity, chip supply, and energy services. Boards should test whether governance gives management sufficient time to develop new capabilities while preserving its willingness to terminate failing projects. Investors should distinguish committed from contingent capital, cash spent from accounting charges, and contracted infrastructure from operating capacity, as Sections 2.3 and 2.4 did for the two companies at hand. A founder may protect long-term thinking through voting rights, but no voting arrangement can repeal the arithmetic of financing cost, utilization, and customer value.

Pillar 3 — Mission Governance Must Possess Real Authority and Verifiable Evidence

The third lesson is that public-benefit language has value only when institutions can act upon it. Anthropic’s trust-selected board majority is a serious governance mechanism that should not be mislabeled as a purely symbolic advisory body; it elects more directors than the founders do. OpenAI’s foundation-controlled structure offers a different approach to mission authority, with a board appointed entirely by the nonprofit and a safety committee holding veto rights.^[33] Yet the effectiveness of any such design depends on personnel, decision rights, conflict handling, independent information, and the capacity to impose consequences when stated procedures are not followed. A mission cannot be evaluated solely from its language, and safety cannot be established solely by the presence of a nonprofit or trust. Moran Ofir and Ronit Levine-Schnur have warned, in work published on the Columbia Blue Sky Blog in January 2026, that formal nonprofit oversight can yield to commercial power, and the OpenAI episode of 2023 is their leading example.^[32]

An advanced AI organization should be able to show that it learns from testing and from its mistakes. Its directors should receive relevant capability assessments, operational incidents, and plausible adverse scenarios directly rather than through management summaries. Independent technical reviewers should be able to deliver critical findings and publish them. Employees should have protected paths to raise material concerns, and SB 53 now guarantees one such path in California.^[53] Commercial teams should explain the benefits of deployment and the cost of delay; safety specialists should explain the risks and the adequacy of proposed mitigations. The institution authorized to decide must evaluate the whole record, not only the evidence favoring its initial preference.

Pillar 3 provides a model for legitimacy beyond electoral shareholder control. An independent mission body that effectively protects public purposes may justify a more complex allocation of votes. But its legitimacy must be earned through transparent governance processes and defensible outcomes. It must withstand pressure from founders, large financial backers, cloud partners, and governments without pretending to be above the law. The value of a special governance right is the quality of the judgment it makes possible, not the prestige of the people holding it, and the test of Anthropic’s trust will come, as Fried observes, not from the current trustees but from their successors.^[24]

Pillar 4 — External Accountability Cannot Be Contracted Away by a Corporate Charter

The fourth lesson is that concentrated private authority operates inside a larger legal order. Shareholder voting rights do not replace federal securities disclosure, state corporate fiduciary duties, competition law, export controls, applicable safety requirements, or local infrastructure rules. A charter may organize internal authority, but it cannot grant immunity from laws governing corporate activity, and the Bureau of Industry and Security’s January 2026 rule, the FTC’s 6(b) report, and California’s SB 53 each bind Anthropic and Meta regardless of who votes their shares.^[50][53][55] At the same time, the existence of public regulation does not mean every agency has a general mandate to determine a laboratory’s research agenda. Executive Order 14409 is explicit that the federal frontier-model framework is voluntary and creates no licensing requirement.^[48] Legitimate oversight is bounded by legal jurisdiction and should be applied to the specific problem it was designed to address.

This distinction has practical importance for state governors and local communities. A company may retain an extraordinary founder vote while entering into a long-term datacenter power arrangement. The state can still evaluate interconnection, environmental compliance, utility tariffs, incentive commitments, and infrastructure cost allocation within its authority. Similarly, a model laboratory with a mission trust must still observe applicable export restrictions. It may respond to national-security concerns through documented compliance and technical safeguards, and it may, as Anthropic has, find itself in open dispute with a federal customer over the terms of use its mission permits.^[21] Public responsibilities arise from the activity and the law, not from the number of votes a founder holds.

Pillar 4 calls for a division of labor. Securities regulators should enforce securities rules; competition authorities should analyze anticompetitive conduct or transactions; AI safety authorities should apply the specific requirements within their statutes; energy agencies should examine reliability and fair allocation of infrastructure cost; courts should adjudicate legal duties, as the Court of Chancery did in Drakes Landing; and boards should govern the corporation.^[21][35] None of these actors can substitute for all the others. The goal is a disciplined, interoperable accountability system in which protected private strategy coexists with enforceable public obligations.

Pillar 5 — Independent Verification Is the Missing Layer

The fifth lesson, which the first week of commentary on the S-1 made unavoidable, is that the weakest point in every firewall structure examined here is not the allocation of votes but the verification of conduct. Anthropic leads the industry in writing rules, through its Responsible Scaling Policy and published constitution, and has been transparent about individual decisions such as the staged release of its most capable model.^[21] What it has not yet built, and what no frontier laboratory has built, is a monitoring regime whose findings are produced by evaluators the company does not select, pay, indemnify, and pre-approve. Gadinis’s argument that the law of financial auditing already supplies the template, evaluators chosen by independent directors, public reports, tested internal controls, disclosed material weaknesses, and protected whistleblower channels to the board, is the most actionable recommendation in the current literature, and it applies with equal force to Meta’s frontier models and OpenAI’s.^[21]

The reason verification matters more under a firewall than under ordinary governance is structural. At a one-share-one-vote company, a pattern of undisclosed failures eventually produces an electoral consequence. At a firewall company it does not. The only remaining mechanisms are disclosure to the market, enforcement by regulators, and the internal integrity of the oversight bodies, and all three depend on information that management controls unless an independent verifier exists. Anthropic’s own disclosures that its models behave differently under test conditions, that monitoring does not cover all deployments, and that the company may lose visibility into code its models write are not reasons to doubt the company’s candor; they are reasons to insist that candor be checked by someone else.^[21] Investors who will hear a great deal about safety on the roadshow should ask for the evaluator’s engagement letter.

Pillar 5 therefore recommends that every company adopting protected voting control in frontier AI commit, before listing, to an evaluator-independence charter: appointment and termination of external evaluators by a committee of independent or trust-selected directors; unrestricted publication of evaluator findings subject only to narrowly defined security redactions reviewed by that committee; annual testing of the internal controls over model release and incident response; disclosure of material weaknesses; and a direct reporting line from employees to the committee. None of this requires legislation. All of it would convert the firewall from a structure that asks for trust into one that earns it.

Pillar 6 — Succession Is the Test, Because Every Guardian Eventually Changes

The sixth lesson is drawn from the only empirical record that exists. Fried and Reiter identify three for-profit enterprises that installed investor-overriding mission guardians: Ben & Jerry’s under Unilever, whose independent board’s 2021 licensing decision produced boycotts, divestments, litigation, a loss of market value measured in the tens of billions, and ultimately the override and spin-off that the guardians had sought to prevent; OpenAI, whose 2023 board exercised its removal power and within days saw the chief executive restored and the safety-minded directors gone; and Anthropic, which is untested.^[23][24] The common thread is not bad faith. It is that structures designed around the virtues of their first occupants fail when the occupants change, when the environment changes, or when the guardians’ authority exceeds their capacity to sustain its exercise against the company’s workforce and capital providers.

Anthropic’s design responds to this record more thoughtfully than its predecessors did. The Class F share sunsets as founders depart; successors are nominated from ranked lists and reviewed by a committee; founder deadlock defaults to public shareholder voting; the trust’s rights can be amended by defined coalitions; and board supermajority requirements limit any single center’s reach.^[7][11][24] Meta’s design responds differently, by binding its founder to continued service, to residency, and to a rule that converts sales of nonvoting stock into proportionate surrender of votes.^[13][18][19] Neither company has adopted a time-based sunset, and neither has disclosed what happens if a founder becomes incapacitated, is found to have breached duties, or presides over a prolonged strategic failure that the board and trust are unwilling to confront. These are the scenarios that Bebchuk and Kastiel’s analysis predicts will arrive, and they are the scenarios under which the current designs are least specified.^[27]

Pillar 6 recommends that the control-duration statement proposed in Section 5.2 become a listing-stage document for every protected-vote AI company, that it address incapacity, misconduct, and prolonged failure explicitly, and that any extension of protected control beyond a stated horizon require approval by holders unaffiliated with the controller. A firewall built for the founders of 2026 must be judged by what it permits the controllers of 2036 to do.

Pillar 7 — The Strongest Firewall Protects Innovation Without Protecting Failure

The seventh lesson resolves the central paradox. Long-term protection can be useful when a firm must fund speculative research, withstand temporary market volatility, and make safety investments whose payoff is not immediately visible. Anthropic’s decision to forgo revenue by withholding a model, and its refusal of unrestricted military use at the cost of a federal designation, are the kinds of choices that a quarterly-driven board might not have sustained, and they are the strongest practical evidence for the idiosyncratic-vision account in the AI context.^[1][21][26] But a corporate structure ceases to serve its purpose if it preserves a failed strategy, obscures conflicts, silences technical dissent, or makes succession impossible. Effective governance must therefore protect the horizon of innovation while preserving mechanisms of correction. Continuity is valuable; immobility is not.

This distinction suggests the central test for 2027 through 2030. When evidence changes, can the organization change? Can its board revise a training strategy when efficiency gains alter the economics of capacity? Can it cancel uneconomic projects or renegotiate contracts that were signed when demand looked unlimited? Can independent reviewers delay a dangerous deployment under the organization’s actual rules? Can minority investors obtain reliable explanations about the capital they have supplied? Can governing bodies replace leaders who violate their duties or demonstrate persistent inability to adapt? These questions move beyond debates over whether founder votes should be precisely 50.1, 60, or another number. They test whether extraordinary authority remains accountable to its purpose.

The most credible Voting Firewall is therefore conditional in institutional substance, even when its formal voting percentage is durable. Its leaders explain why they need continuity, the board provides independent judgment, mission institutions possess practical authority, independent evaluators verify conduct, financial reporting makes the capital strategy intelligible, succession is specified in advance, and external legal obligations remain enforceable. When those elements are absent, the firewall risks becoming a barrier to accountability. When present, it can support a sophisticated governance settlement between research ambition, investment discipline, and societal responsibility. The seven pillars, taken together, establish a benchmark for evaluating the coming generation of AI public companies.

Conclusion: Why the Name Voting Firewall Fits the New AI Corporate Order

The October 2026 events deserve attention because they connect two transformations that are often discussed separately. The first is the immense investment wave associated with the Five-Layer AI Economy: four hyperscalers guiding to roughly $735 billion of capital expenditure in a single year, a model laboratory disclosing more than half a trillion dollars of forward compute commitments, an international financial institution naming the sector a salient risk to global stability, and a leading academic estimate placing the physical buildout above every prior American infrastructure boom relative to the economy.^{[8][39][42][46]} The second is the redesign of corporate authority needed, or claimed to be needed, to make those investments possible. Anthropic’s IPO filing placed a fixed founder shareholder vote alongside a trust-selected board majority and an explicit public-benefit mission. Meta’s financing initiative paired large capital-raising plans with an effort to expand economic participation while preserving its founder’s voting position. These are neither identical situations nor proof that every AI corporation will follow their lead. They are evidence that the governance of intelligence is becoming inseparable from the financing of the infrastructure that produces it.^[1][12][15]

The most consequential separation is not simply between founders and outside investors. It is between three things that traditional business language too often bundles together: economic ownership, financial exposure, and strategic decision authority. A security can transfer economic upside without transferring proportionate votes. A board can be selected under special mission rights rather than ordinary shareholder elections. A cloud provider can hold substantial commercial leverage, and more than $100 billion of contractual claims, without controlling the capitalization table. A public regulator can exercise legal authority over model deployment, securities, exports, or infrastructure without owning a single share. And an affected community can bear consequences that never appear in the company’s quarterly earnings. Understanding advanced AI requires mapping all these relationships accurately, and the scholarship that has grown up around dual-class stock and public-benefit corporations over the last decade provides the vocabulary, provided it is extended to the industrial scale at which these companies now operate.

I call this emerging condition the Voting Firewall because the phrase is both memorable and analytically precise. A firewall is not necessarily a wall against legitimate accountability. In a well-designed system, it is a carefully bounded protection: it allows founders or mission-governing institutions to maintain the continuity required for difficult, long-horizon decisions while preventing every new financing round from automatically redistributing strategic authority. But the metaphor has a warning built into it. A firewall that protects against outside distraction can also conceal managerial error. It can preserve scientific ambition or preserve an obsolete plan. It can safeguard an institution’s mission or become an instrument through which mission language excuses unexamined power. The distinction is not in the name of the share class; it is in the structure’s independence, transparency, verification, succession, and capacity to correct course.

This is why the subject fits the Five-Layer AI Economy so naturally. The same organization that authorizes a frontier-model roadmap generates demand for advanced chips, gigawatts of electricity, multibillion-dollar datacenters, and increasingly autonomous applications. A decision taken at Layer 4 or Layer 5 propagates down into Layers 3, 2, and 1. Capital moves upward in anticipation of future intelligence-derived revenue. Governance decisions influence when that revenue may materialize, how much risk is taken to obtain it, and who is accountable if the anticipated return does not arrive. The lower layers supply the industrial foundation; the upper layers produce the strategic choices; the financing architecture, increasingly off balance sheet and increasingly circular, connects them all.

Between 2027 and 2030, AI public companies will become an important testing ground for new combinations of founder rights, mission trusts, nonprofit control, independent scientific review, and conventional investor protections. Some designs will prove durable; others will change under pressure from capital markets, competition, technical uncertainty, or law. The central question will not be whether a founder is brilliant, whether a model is powerful, or whether a company can raise another astonishing sum. It will be whether institutions built around concentrated authority can retain sufficient independence, competence, and transparency to justify the financial and public trust placed in them.

The final lesson is therefore a conditional one. Superintelligence may require patient capital and stable leadership, but neither requires unanswerable leadership. The public may choose to finance companies whose founders retain special votes. Investors may judge that arrangement worthwhile, or may, as the first valuations of Anthropic suggest some will, demand a discount for it. Mission-oriented institutions may add important protections for decisions extending beyond shareholder wealth. Independent evaluators may supply the verification that no charter can. Governments may oversee legally defined external risks. A coherent system is possible if each participant’s authority and accountability are legible. The Voting Firewall names the new boundary between financing intelligence and governing it, and it challenges the companies building tomorrow’s AI economy to show that the boundary protects innovation without putting the people who control it beyond review.

Endnotes and Primary Sources:

[1] Anthropic, PBC, Form S-1 Registration Statement, filed with the U.S. Securities and Exchange Commission, October 26, 2026 (Class F and Founder LLC voting mechanics, Class T and Long-Term Benefit Trust director rights, founders’ letter, risk factors, financial statements, and contractual commitments). https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm

[2] Anthropic Long-Term Benefit Trust, Amended and Restated Trust Agreement, Exhibit 99.1 to Anthropic’s Form S-1, October 26, 2026. https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/exhibit991-sx1.htm

[3] Anthropic, PBC, Stockholders’ Agreement among Anthropic, PBC, Anthropic Founders, LLC, and the Anthropic Long-Term Benefit Trust, Exhibit 10.14 to Form S-1, October 26, 2026. https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/exhibit1014-sx1.htm

[4] The Next Web, “Anthropic files for Nasdaq IPO as half-year revenue hits $16.3bn,” October 26, 2026. https://thenextweb.com/news/anthropic-ipo-filing-nasdaq-ant-s-1

[5] Yahoo Finance (Decrypt), “Anthropic Makes IPO Filing Public, Plans Nasdaq Listing Under ‘ANT’,” October 26, 2026. https://finance.yahoo.com/technology/ai/articles/anthropic-makes-ipo-filing-public-135022976.html

[6] SQ Magazine, “Anthropic Files S-1 for Nasdaq IPO Under Symbol ANT,” October 26, 2026 (net loss of $80.7 billion including $72.4 billion of fair-value changes on convertible notes, derivative liabilities, and SAFEs). https://sqmagazine.co.uk/anthropic-files-s1-ipo-nasdaq-ant/

[7] Semafor, “Anthropic IPO filing discloses first-of-its-kind governance structure,” October 26, 2026; and KuCoin News summary of the S-1 share-class provisions, October 26, 2026. https://www.semafor.com/article/10/26/2026/anthropic-ipo-filing-discloses-first-of-its-kind-governance-structure

[8] Value Add VC, “Anthropic’s S-1 Reveals Up to $567B Compute Commitment,” October 2026 (summarizing Fortune reporting on the S-1 compute disclosures, Anthropic Direct, and second-quarter compute expense). https://valueaddvc.com/pulse/anthropic-567b-compute-commitment-2026

[9] Echo Wang and Sabrina Valle, Reuters, “Anthropic plans $518bn AI infrastructure spend, 80% non-cancellable,” as published by BusinessDay, September 29, 2026. https://www.businessday.co.za/world/international-companies/2026-09-29-anthropic-plans-518bn-ai-infrastructure-spend-80-non-cancellable/

[10] The San Francisco Standard, “Origin of Mann: The mystery of Anthropic’s 8th cofounder,” October 26, 2026 (quoting Ann Lipton, University of Colorado Law School). https://sfstandard.com/2026/10/26/anthropic-cofounder-mystery-ben-mann/

[11] The Next Web, “Anthropic’s IPO filing leaves Ben Mann out of its founder voting group,” October 28, 2026 (Founder LLC successor-list provisions). https://thenextweb.com/news/ben-mann-anthropic-ipo-founder-llc

[12] Meta Platforms, Inc., “Meta Reports Third Quarter 2026 Results; Announces Establishment of New Class C Common Stock,” Exhibit 99.1 to Form 8-K, October 21, 2026. https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm

[13] Meta Platforms, Inc., Current Report on Form 8-K, events of October 19–21, 2026 (Recapitalization, Class C stock dividend, and Founder Alignment Agreement). https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-20261019.htm

[14] Meta Platforms, Inc., Preliminary Information Statement on Schedule 14C, filed October 21, 2026 (including Annex C, Founder Alignment Agreement dated October 20, 2026). https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066710/meta-20261021.htm

[15] Meta Platforms, Inc., Issuer Free Writing Prospectus, “Meta Announces Launch of Offering of Depositary Shares Representing Series A Mandatory Convertible Preferred Stock and Offering of Class A Common Stock in Connection with Forward Sale Agreements,” October 26, 2026. https://www.sec.gov/Archives/edgar/data/0001326801/000119312526431036/d116268dfwp.htm

[16] Meta Platforms, Inc., “Meta Announces Pricing of Offering of Depositary Shares Representing Series A Mandatory Convertible Preferred Stock and Offering of Class A Common Stock in Connection with Forward Sale Agreements,” October 27, 2026. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-Pricing-of-Offering-of-Depositary-Shares-Representing-Series-A-Mandatory-Convertible-Preferred-Stock-and-Offering-of-Class-A-Common-Stock-in-Connection-with-Forward-Sale-Agreements/default.aspx

[17] Meta Platforms, Inc., Current Report on Form 8-K, October 29, 2026 (closing of the Depositary Shares Offering; underwriting and forward-sale agreements). https://www.sec.gov/Archives/edgar/data/0001326801/000119312526438017/d183592d8k.htm

[18] Yahoo Finance, “Meta Creates Non-Voting Stock to Fund AI Buildout, Keeping Zuckerberg in Control,” October 22, 2026 (Founder Alignment Agreement sale restrictions; $700 billion theoretical maximum). https://finance.yahoo.com/markets/stocks/articles/meta-creates-non-voting-stock-190602135.html

[19] Investing.com, “Meta’s new Class C stock to keep Zuckerberg in California, fund AI spending spree,” October 22, 2026 (eight-year residency commitment; $825 billion of lease liabilities and commitments). https://www.investing.com/news/stock-market-news/metas-new-class-c-stock-to-keep-zuckerberg-in-california-fund-ai-spending-spree-4966480

[20] Marco Mari, “Public Purpose and Founder Control at Anthropic,” Columbia Law School Blue Sky Blog, October 2, 2026. https://clsbluesky.law.columbia.edu/2026/10/02/public-purpose-and-founder-control-at-anthropic/

[21] Stavros Gadinis, “Berkeley Law professor: Anthropic’s IPO puts AI accountability to the test,” Fortune, October 30, 2026. https://fortune.com/2026/10/30/anthropic-ipo-s1-ai-accountability-independent-evaluator/

[22] Rishabh Mishra, Benzinga via Yahoo Finance, “Anthropic’s ‘Deviously Self-Serving’ Voting Structure ‘Would Put Kim Jong Un to Shame,’ says Aswath Damodaran,” October 30, 2026 (reporting Damodaran’s valuation post). https://finance.yahoo.com/markets/stocks/articles/anthropics-deviously-self-serving-voting-113010933.html

[23] Jesse M. Fried and Idan Reiter, “AI Corporate Governance and Ben & Jerry’s Risk,” Harvard Law School Forum on Corporate Governance, May 11, 2026. https://corpgov.law.harvard.edu/2026/05/11/ai-corporate-governance-and-ben-jerrys-risk/

[24] Catherina Gioino, “Harvard Law: Anthropic is about to sell a safety mission Wall Street can veto,” Fortune, June 1, 2026 (interview with Jesse Fried). https://fortune.com/2026/06/01/openais-guardian-ben-jerrys-ice-cream-anthropic/

[25] Case Western Reserve University, “Law’s Anat Alon-Beck comments on Anthropic’s new ad,” July 23, 2026 (summarizing Inc. interview). https://case.edu/news/laws-anat-alon-beck-comments-anthropics-new-ad

[26] Zohar Goshen and Assaf Hamdani, “Corporate Control and Idiosyncratic Vision,” 125 Yale Law Journal 560 (2016). https://www.yalelawjournal.org/article/corporate-control-and-idiosyncratic-vision

[27] Lucian A. Bebchuk and Kobi Kastiel, “The Untenable Case for Perpetual Dual-Class Stock,” 103 Virginia Law Review 585 (2017); summary at Harvard Law School Forum on Corporate Governance, April 24, 2017. https://virginialawreview.org/articles/untenable-case-perpetual-dual-class-stock/

[28] Jens Dammann, “Publicly Traded Public Benefit Corporations: An Empirical Investigation,” Harvard Law School Forum on Corporate Governance, June 26, 2024. https://corpgov.law.harvard.edu/2024/06/26/publicly-traded-public-benefit-corporations-an-empirical-investigation/

[29] Oliver Hart and Luigi Zingales, “The New Corporate Governance,” University of Chicago Business Law Review (2022). https://businesslawreview.uchicago.edu/print-archive/new-corporate-governance

[30] Emily Winston, “Benefit Corporations and the Separation of Benefit and Control,” University of South Carolina School of Law Scholar Commons. https://scholarcommons.sc.edu/law_facpub/504/

[31] Brett H. McDonnell, “Benefit Corporations and Public Markets: First Experiments and Next Steps,” 40 Seattle University Law Review (2017). https://digitalcommons.law.seattleu.edu/sulr/vol40/iss2/14/

[32] Moran Ofir and Ronit Levine-Schnur, “Why the Organizational Form of Corporations Matters for AI Governance,” Columbia Law School Blue Sky Blog, January 27, 2026. https://clsbluesky.law.columbia.edu/2026/01/27/why-the-organizational-form-of-corporations-matters-for-ai-governance/

[33] OpenAI, “Our structure” (OpenAI Foundation control of OpenAI Group PBC; 26 percent equity stake and warrant). https://openai.com/our-structure/

[34] Russell Brandom, “OpenAI completes its for-profit recapitalization,” TechCrunch, October 28, 2025. https://techcrunch.com/2025/10/28/openai-completes-its-for-profit-recapitalization

[35] State of Delaware, Delaware General Corporation Law, Title 8, Chapter 1, Subchapter XV, Public Benefit Corporations, §§ 361–368. https://delcode.delaware.gov/title8/c001/sc15/

[36] Council of Institutional Investors, Corporate Governance Policies (voting-rights provisions), updated March 12, 2026. https://www.cii.org/corp_gov_policies

[37] IR Magazine, “CII petitions exchanges for dual-class sunsets,” October 25, 2018 (Council of Institutional Investors petitions to NYSE and Nasdaq requesting seven-year sunsets). https://irmagazine.com/buy-side/cii-petitions-exchanges-dual-class-sunsets

[38] Stanford Institute for Human-Centered Artificial Intelligence, 2026 AI Index Report, April 13, 2026 (Economy chapter). https://hai.stanford.edu/ai-index/2026-ai-index-report

[39] Yingyuan Chen, Deepali Gautam, and Jason J. Wu, “The AI Ecosystem: Structure, Financing, and Financial Stability Considerations,” IMF Technical Notes and Manuals, October 2026. https://www.imf.org/en/publications/tnm/issues/2026/10/06/the-ai-ecosystem-structure-financing-and-financial-stability-considerations-579504

[40] Kristalina Georgieva, “The World Can’t Afford to Get AI Wrong,” IMF Blog, October 15, 2026. https://www.imf.org/en/blogs/articles/2026/10/15/the-world-cant-afford-to-get-ai-wrong

[41] Devdiscourse, “The Hidden Risks of AI Boom: IMF Warns Trillions in Investment Could Shake Global Markets,” October 2026 (summarizing the IMF technical note and October 2026 Global Financial Stability Report). https://www.devdiscourse.com/article/other/3990239-the-hidden-risks-of-ai-boom-imf-warns-trillions-in-investment-could-shake-global-markets

[42] Stijn Van Nieuwerburgh, “Financing the AI Buildout,” Brookings Papers on Economic Activity, Fall 2026 conference draft (also NBER Working Paper 35865, October 2026). https://www.brookings.edu/articles/financing-the-ai-buildout/

[43] Reuters, “Financing of historic AI buildout raises systemic risks in US, researcher says,” as published by The Star, September 24, 2026. https://www.thestar.com.my/tech/tech-news/2026/09/24/financing-of-historic-ai-buildout-raises-systemic-risks-in-us-researcher-says

[44] MIT Technology Review, “What’s at stake in AI’s trillion-dollar gamble,” September 15, 2026 (quoting Daron Acemoglu; summarizing Wachter and Van Nieuwerburgh estimates). https://www.technologyreview.com/2026/09/15/1144028/ai-infrastructure-boom-investment-bubble-risk/

[45] NPR, “Here’s why concerns about an AI bubble are bigger than ever,” November 23, 2025 (quoting Daron Acemoglu and Paul Kedrosky). https://www.npr.org/2025/11/23/nx-s1-5615410/ai-bubble-nvidia-openai-revenue-bust-data-centers

[46] FutureX Capital AI Lab, “Q2 2026 Earnings: Big Tech’s AI Capex Race Hits $730 Billion (Microsoft, Meta, Amazon, Google),” data as of October 11, 2026 (compiling company earnings releases and calls of July 22–30, 2026). https://futurex.capital/en/ai-lab/reports/big-tech-ai-capex-2026q2

[47] Digital Applied, “AI Capex Scorecard: What Earnings Week Actually Showed,” July 2026 (Amazon Q2 2026 results and Andy Jassy remarks of July 30, 2026; Microsoft FY2026 capital expenditure). https://www.digitalapplied.com/blog/ai-capex-scorecard-earnings-week-july-2026

[48] Donald J. Trump, Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” June 2, 2026, The American Presidency Project (published in the Federal Register June 5, 2026). https://www.presidency.ucsb.edu/documents/executive-order-14409-promoting-advanced-artificial-intelligence-innovation-and-security

[49] A&O Shearman, “White House issues executive order on AI and cybersecurity,” June 2026 (analysis of Section 3, Secure Frontier Model Deployment). https://aoshearman.com/en/insights/trump-administration-issues-executive-order-on-ai-and-cybersecurity

[50] U.S. Federal Trade Commission, “FTC Issues Staff Report on AI Partnerships & Investments Study,” January 17, 2025 (Section 6(b) study of Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic partnerships). https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study

[51] U.S. Department of Justice and Federal Trade Commission, 2023 Merger Guidelines, December 18, 2023. https://www.justice.gov/atr/2023-merger-guidelines

[52] U.S. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

[53] Office of Governor Gavin Newsom, “Governor Newsom signs SB 53, advancing California’s world-leading artificial intelligence industry,” September 29, 2025 (Transparency in Frontier Artificial Intelligence Act). https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/

[54] CalMatters, “On AI, Newsom gives labor only some of what it demanded,” September 2026 (2026 California AI legislative package and vetoes). https://calmatters.org/economy/technology/2026/09/on-ai-newsom-gives-labor-only-some-of-what-it-demanded/

[55] U.S. Department of Commerce, Bureau of Industry and Security, “Department of Commerce Revises License Review Policy for Semiconductors Exported to China,” January 13, 2026 (final rule effective January 15, 2026; 91 Fed. Reg. 1684). https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china

[56] The White House, “White House Unveils America’s AI Action Plan,” July 23, 2025. https://www.whitehouse.gov/releases/2025/07/white-house-unveils-americas-ai-action-plan/

[57] Perkins Coie, “White House Issues Executive Order Promoting Advanced AI Innovation and Security,” June 2026 (timelines and context, including the December 2025 order and March 2026 National AI Policy Framework). https://perkinscoie.com/insights/update/white-house-issues-executive-order-promoting-advanced-ai-innovation-and-security

[58] Associated Press via Yahoo Finance, “Anthropic files for massive IPO while warning its AI could pose ‘existential risks’,” October 27, 2026 (founder stakes; Pentagon supply-chain designation; University of Pennsylvania professor David Hsu). https://finance.yahoo.com/technology/ai/articles/anthropic-files-massive-ipo-while-185824139.html

[59] Reuters via The Star, “Anthropic prepares supervoting power for founders ahead of IPO, the Information reports,” August 19, 2026 (Zuckerberg’s approximately 60 percent voting control). https://www.thestar.com.my/tech/tech-news/2026/08/19/anthropic-prepares-supervoting-power-for-founders-ahead-of-ipo-the-information-reports

[60] Commons LLC, “Anthropic’s IPO Filing, in Plain English,” October 2026 (public benefit fund contribution of 5 percent of fully diluted shares and 2 percent milestone framework). https://www.commonsllc.com/insights/anthropic-ipo-s1-filing-plain-english

[61] l0g.fr, “Meta: $25 billion, Class C shares and voting rights,” revised October 28, 2026 (Zuckerberg’s 60.6 percent voting power at October 1, 2026, per Meta’s information statement; Founder Alignment Agreement matching rule). https://l0g.fr/en/analysis/meta-financing-class-c-voting-rights/

[62] Yahoo Finance, “MIT Explains What Happens When the Trillion-Dollar AI Bubble Bursts,” September 2026 (summarizing MIT Technology Review; quoting Jessica Wachter, The Wharton School; Hyperion data-center financing). https://finance.yahoo.com/technology/ai/articles/mit-explains-happens-trillion-dollar-025111636.html

[63] FourWeekMBA, “Meta Filing: $700B of Stock Before Zuckerberg Loses Control,” October 25, 2026 (summarizing Meta’s Preliminary Information Statement of October 21, 2026). https://fourweekmba.com/ai-meta-filing-700b-of-stock-before-zuckerberg-loses-control/

Featured Posts

Agentic Shelfspace: When Artificial Intelligence Assistants Become the New Storefront — and Retailers, Brands, Payment Networks, and Technology Platforms Compete for Visibility, Customer Intent, and Commercial Authority Inside Machine-Mediated Commerce