Corporate Architecture for a Strong-AI Era

1. Why this question, and why now

The standard listed corporation is optimized to convert capital into shareholder returns. That design has produced enormous prosperity, but it has a known blind spot: costs and risks that fall on people who are not parties to the corporate contract. For most industries, external regulation patches the blind spot tolerably well. Frontier AI is a hard case for three reasons.

• Scale of externalities. A small number of private firms make deployment decisions whose downside risks (misuse, accidents, destabilization of labor markets and information ecosystems) and upside benefits (scientific acceleration, cheap cognition) are global, while their fiduciary duties are local to shareholders.
• Speed versus regulation. Capability cycles are now shorter than legislative cycles. In the 3–10 year window this study addresses, corporate-governance design is one of the few levers that can act faster than statute, because charters, trusts, and shareholder agreements can be written today.
• Winner-take-much economics. If AI development produces extreme concentrated profits, the question of who owns those profits — and who can veto how they are pursued — becomes a question of political stability, not just corporate law.

The premise of this study, stated plainly: a pure profit-maximizing, shareholder-only company is an ill-fitting vehicle for developing systems that may transform the conditions of life for all of humanity, for animals whose treatment is mediated by human economies, for generations who cannot yet vote or buy shares, and for AI systems whose moral status is genuinely uncertain. The question is not whether that premise is emotionally appealing but whether any alternative structure actually works — survives contact with capital markets, talent competition, and its own insiders. Section 2 therefore examines the real record, including failures, before Section 4 proposes anything.

Scope and honesty notes. "Strong AI" here means AI systems substantially more capable than 2026 frontier models, arriving within the study window; the analysis does not depend on any specific capability forecast. Every factual claim is cited to a public, checkable source; where a figure is uncertain or contested this is said explicitly, and no numbers are invented. This document names companies and public institutions only; it names no private individuals except public officers and published authors where necessary for citation. The study's premise — that shareholder-only firms are ill-suited — is itself treated as testable: indicator P9 in Section 5.2 states what evidence would count against it.

1.1 Eight design questions

Any candidate architecture has to answer eight questions; they organize the rest of the study. Sections 2 and 3 gather the evidence, Section 4's design components are tagged to them, and Section 6 carries forward the parts only legislatures can answer.

2. Survey of real existing forms

Eight families of structures are examined. For each: how it works mechanically, real examples, and documented failures or limits.

2.1 Benefit corporations and B Corps

Mechanism. A benefit corporation (in Delaware, a Public Benefit Corporation or PBC) is a for-profit corporation whose charter names a public benefit purpose and whose directors must balance shareholders' pecuniary interests, the interests of those materially affected by the corporation's conduct, and the stated public benefit.^[1] This changes the directors' permission structure (they may lawfully trade profit against mission) more than their obligation structure: under §367, only stockholders holding at least 2% of shares (or, for listed companies, the lesser of 2% or $2 million in market value) may sue to enforce the balancing duty — the public, workers, and affected third parties have no statutory standing, and courts have almost never policed the balancing.^[1] Anthropic itself, explaining why it layered a trust on top of its PBC status, wrote that the PBC form alone "does not make the directors of the corporation directly accountable to other stakeholders."^[12] "Certified B Corp" is different and often confused with it: a private certification by the nonprofit B Lab, renewed periodically, with no legal force of its own; in April 2025 B Lab replaced its cumulative point-scoring assessment with mandatory requirements across seven impact topics.^[2]

Examples. Thousands of firms hold B Corp certification; Delaware PBCs include Anthropic and, since October 2025, OpenAI's operating company (both discussed below), as well as consumer firms such as Allbirds and Lemonade, which IPO'd as PBCs. An empirical study of 295 Delaware PBCs formed in 2013–2019 found they raised over $2.5 billion, largely from conventional venture investors, though in somewhat smaller rounds than comparable startups — and that funding concentrates in consumer-facing sectors where mission signaling doubles as branding, an explicit "purpose-washing" risk.^[50] Quasi-experimental studies of B Corp certification find positive post-certification sales growth, with mixed short-run profitability and employment effects.^[51]

2.2 OpenAI: nonprofit control, capped profit, and what actually happened

Mechanism as designed. OpenAI was founded in 2015 as a nonprofit with the mission of ensuring artificial general intelligence benefits all of humanity. In 2019 it created a "capped-profit" partnership: investors' returns were capped (initially at 100× for the earliest investors) with excess value flowing to the nonprofit, and the nonprofit's board retained full control, with most board members barred from holding equity.^[5]

What happened in practice. The structure was stress-tested twice, with instructive results.

• November 2023. The nonprofit board exercised its core power and removed the CEO, citing a loss of confidence in his candor.^[6] Within five days — after a resignation threat by the great majority of employees (more than 700, per contemporaneous reporting^[53]; their equity value depended on the company's trajectory) and pressure from the largest investor — the CEO was reinstated and the board that had acted was largely replaced. The formal power was real; the practical power dissolved on contact with the interests of employees and capital.
• 2024–2025 restructuring. In December 2024 OpenAI proposed a restructuring that — in the Delaware Attorney General's later description — "would remove the current nonprofit entity from overseeing the for-profit activities"; in May 2025, after objections from former employees and civil-society groups and dialogue with the California and Delaware Attorneys General, it announced the nonprofit would retain control.^[60][61] The recapitalization completed on 28 October 2025 kept nonprofit control but abandoned the profit cap. The nonprofit (now the OpenAI Foundation) holds Class N stock with the sole power to appoint and remove directors of the operating company (now OpenAI Group PBC, a Delaware public benefit corporation), plus 26% of conventional equity valued at about $130 billion and a warrant for significant additional shares if OpenAI Group's share price rises more than tenfold within 15 years; Microsoft holds roughly 27%, and current and former employees and investors the remaining 47%.^[7][8][9] Memoranda and conditions agreed with both Attorneys General bind key commitments: mission primacy, foundation control, advance notice to regulators before any change of control or mission, and — per the Delaware Attorney General's conditions — a Safety and Security Committee that remains a committee of the nonprofit with authority to require mitigations "up to and including halting the release of models," alongside directors who serve exclusively on the nonprofit board.^[10][11][61]

2.3 Anthropic: PBC plus Long-Term Benefit Trust

Mechanism. Anthropic is a Delaware PBC whose stated purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity. Its distinctive layer is the Long-Term Benefit Trust (LTBT): a Delaware purpose trust holding a special class of stock (Class T) whose five financially disinterested trustees gain the power, phased in by time and fundraising milestones, to elect a majority of the board.^[12] Trustees serve short terms and choose their own successors in consultation with the company; "failsafe" provisions allow the Trust's powers to be changed without trustee consent if sufficiently large stockholder supermajorities agree, with thresholds that rise over time.^[12][13] The LTBT has exercised real appointment power, and in April 2026 Anthropic announced that Trust-appointed directors now constitute a majority of its board; the Trust's composition spans AI safety, global health, national security, and policy backgrounds.^[12][62]

2.4 Foundation-owned ("enterprise foundation") firms

Mechanism. An industrial or enterprise foundation is a self-owning nonprofit that holds a controlling stake in an operating company, typically established by a founder's irrevocable donation. The foundation's charter fixes purposes (continuation of the business, philanthropy, research); no one can extract the equity. This is the dominant ownership form among Denmark's largest firms and a significant one in Germany.

Examples.

• Robert Bosch GmbH: about 94% of share capital is held by the charitable Robert Bosch Stiftung, which by design "has no influence on the strategic or business orientation" of the group; some 93% of voting rights sit with Robert Bosch Industrietreuhand KG, an industrial trust, with the remainder held by the founder's descendants — a deliberate split of economic benefit from control.^[14]
• Novo Nordisk: the Novo Nordisk Foundation, via wholly owned Novo Holdings A/S, held at end-2025 about 28.1% of Novo Nordisk's share capital and 77.3% of votes through dual-class shares; its articles require keeping a controlling interest, and dividends fund one of the world's largest scientific philanthropy programs.^[16][17]
• Carlsberg: the Carlsberg Foundation (founded 1876) held 29.99% of capital and 77.53% of votes in Carlsberg A/S at end-2025; its charter requires it to keep at least 51% of votes, and it funds basic science.^[19]

Empirical work on Danish foundation-owned firms finds lower leverage and higher survival rates than investor-owned peers, with longer executive tenures and a mixed return profile by size: one study reports large foundation-owned firms earning a higher return on assets (about 5.0% versus 3.6% for large non-foundation peers), while smaller ones underperform and growth is generally slower.^[20][54] The evidence supports mission-locked ownership being compatible with competitiveness in mature industries — at some cost in dynamism.

2.5 Cooperatives and Mondragon

Mechanism. In a worker cooperative, members hold one vote each regardless of capital, elect governance, and share surpluses by labor contribution. Mondragon (Basque Country, founded 1956) is the largest industrial group of this kind — reported 2025 results: €11.322 billion in sales, 71,415 workers, and 1,346 new jobs^[55] — with internal solidarity mechanisms — capped pay ratios (most managers within roughly six times the lowest wage), inter-cooperative funds, and relocation of members from failing to healthy co-ops.^[21]

2.6 German codetermination

Mechanism. Under Germany's Codetermination Act 1976, companies with more than 2,000 domestic employees must give workers half the seats on the supervisory board (the chair, elected by shareholders, holds a tie-breaking second vote); companies with 500–2,000 employees give one third. Works councils hold information and consultation rights at plant level.^[23] It is the largest natural experiment in putting non-shareholders inside corporate governance by statute.

Evidence. Careful empirical reviews find codetermination's measured effects are surprisingly moderate in both directions — no collapse of profitability or investment, modest gains in wage stability and training, some evidence of longer-horizon decision-making.^[24] A prominent quasi-experimental study reports no clear wage or rent-sharing effect, but higher capital formation and roughly 16–21% higher value added per worker in codetermined firms;^[56] other work reports possible shareholder-value trade-offs. (These performance findings are reported as published and were not independently re-verified for this study.) The honest reading: board seats for a stakeholder class are survivable and mildly useful, not transformative — and parity is not control, since the shareholder-side chair holds the tie-breaking vote.^[23]

2.7 Steward-ownership: Patagonia, Purpose Foundation, and predecessors

Mechanism. Steward-ownership separates control from economic extraction by design: voting control is held by stewards (active leadership or a purpose trust) and cannot be sold or inherited; profits serve the mission, are reinvested, or are donated; a "golden share" held by an independent guardian (in several European structures, a foundation such as the Purpose Foundation) can veto any attempt to unwind these rules.^[26]

Examples. Patagonia's founders transferred, in September 2022, 100% of voting stock (2% of total equity) to the Patagonia Purpose Trust and 98% of the economics to the Holdfast Collective, a 501(c)(4) that receives all profits not reinvested (projected by the company at roughly $100 million per year) — "Earth is now our only shareholder."^[25][57] The search engine Ecosia is steward-owned with a Purpose Foundation golden share. The form has deep roots: the Carl Zeiss Foundation has owned Zeiss since 1889 under a statute written to lock in worker protections and science funding,^[27] and the John Lewis Partnership has been held in trust for its employees under a written constitution since 1929.^[28]

2.8 Windfall clauses, AI dividends, and sovereign funds

Mechanism. These are revenue-side structures, agnostic about who governs the firm. The Windfall Clause is a proposed ex-ante contractual commitment by AI developers: if profits ever exceed some extreme threshold (the authors discuss thresholds framed as fractions of a percent of gross world product), a rising marginal share is distributed for the common good; the commitment is cheap to sign now precisely because the trigger is unlikely for any given firm.^[29] The working precedents for distribution machinery are resource funds: the Alaska Permanent Fund, a constitutionally protected fund of oil revenues that has paid an equal annual dividend to every eligible Alaskan resident since 1982 — an amount now set year by year by the legislature (the 2025 dividend was $1,000 under House Bill 53; the announced 2026 dividend is $1,200 including a $200 energy-relief payment), which keeps it politically exposed^[30] — and Norway's Government Pension Fund Global, which converts petroleum revenue into a diversified endowment (NOK 22,683 billion at end-June 2026) that funds the state budget under a parliamentary spending rule rather than paying citizens directly.^[31] In the AI context, OpenAI's CEO has published a proposal ("Moore's Law for Everything") to tax corporate equity and land into a citizen equity fund,^[32] and US state-level "AI dividend" proposals have begun to appear (e.g., a published policy proposal by a New York State legislator).^[58]

2.9 Perpetual purpose trusts and commons governance

Purpose trusts. Several structures above rest on the same legal chassis: the non-charitable purpose trust — a trust managed for a declared purpose rather than for beneficiaries, policed by a designated enforcer. Delaware law validates these expressly (12 Del. C. §3556), which is why both the Patagonia Purpose Trust and Anthropic's LTBT could be built without new legislation.^[63] Their generic failure modes: everything depends on the enforcer or protector (there is no market check and no electoral check), and a purpose drafted once can prove too vague to bind or too rigid to adapt.

Commons governance. For shared AI resources — compute pools, datasets, open models — the relevant evidence base is not corporate law but commons scholarship. Elinor Ostrom shared the 2009 Nobel Memorial Prize "for her analysis of economic governance, especially the commons"; her design principles for long-enduring common-pool institutions (clear boundaries, congruence with local conditions, collective-choice arrangements, monitoring accountable to users, graduated sanctions, conflict-resolution mechanisms, recognized rights to organize, nested enterprises) are a tested template for polycentric governance.^[64] The honest limit: that evidence comes overwhelmingly from local and regional resources with identifiable users; its transfer to global, fast-moving AI markets is untested.

2.10 What the record shows

Three patterns recur. First, purpose without control fails; control without external enforcement bends. The only commitments that survived OpenAI's decade of stress were those a state Attorney General could enforce. Second, every insider class can be captured — boards by CEOs, trustees by companies, worker representatives by management, foundations by their own boards. Durability comes from plural checks with different failure modes, not from any single virtuous body. Third, no mission-locked form has yet financed frontier-scale AI without diluting its locks — this is the open engineering problem Section 4 addresses.

2.11 A toy simulation: the rules produce what they encode

A companion research run built a small agent-based simulation comparing five stylized ownership rules over 60 periods and 8 random seeds: shareholder_only, pbc_benefit_duty, foundation_steward_locked_mission, cooperative, and shareholder_plus_ai_windfall_rule (an ex-post redistribution of exceptional profits layered on an otherwise shareholder-only firm). Two outcomes were tracked: wealth concentration (Gini coefficient and top-1% share) and "mission drift" (mean absolute divergence between the firm's mission target and its period-by-period decisions under competitive pressure). Full summary statistics and reproducibility notes are archived with this study.^[59]

At the final period, mean wealth Gini across seeds ordered as: cooperative 0.192 < foundation/steward 0.213 < PBC 0.223 ≈ shareholder-plus-windfall 0.223 < shareholder-only 0.234. Mission drift over the last ten periods: foundation/steward 0.108 < cooperative 0.153 < PBC 0.272 < shareholder-only = shareholder-plus-windfall at 0.411. In the run's time series, the Gini declines in all scenarios (the ownership rules reorder outcomes rather than reverse trends), while mission drift is essentially flat over time — a level set by each rule, not something any rule gradually corrects.

Read jointly with the historical record, the toy model reinforces one point in Section 2.10 from a different direction: revenue-sharing overlays address distribution but leave conduct untouched, while structures that bind conduct (mission floors, control locks) do so exactly to the extent the binding is real — in the model because it is hard-coded, in reality only if the enforcement problems of Section 5 are solved.

3. Voiceless stakeholders: what is law, what is proposed, what is speculation

A structure "for all of humanity, animals, future generations, and possibly-conscious AI" must represent parties who cannot vote, sue, or hold shares. Precision matters here, because this area attracts wishful conflation. Each item below is tagged: Law already enacted somewhere; Proposal a serious, published proposal under real discussion; Speculation an idea with no operational precedent.

3.1 Animals

• Law The EU treaties require the Union and member states, in forming certain policies, to "pay full regard to the welfare requirements of animals, since animals are sentient beings" (Article 13 TFEU).^[33] Germany added animal protection to its constitution's state-objectives clause in 2002.^[34]
• Law The UK Animal Welfare (Sentience) Act 2022 recognizes vertebrates, cephalopod molluscs, and decapod crustaceans as sentient (the latter two following an LSE evidence review) and creates an Animal Sentience Committee that reports on whether government policy paid "all due regard" to animal welfare — a working template for an institutional voice.^[35]
• Law Court rulings have gone both ways on stronger claims. New York's highest court held in 2022 that habeas corpus does not extend to Happy, an elephant, because she is not a "person" (Nonhuman Rights Project v. Breheny).^[65] Ecuador's Constitutional Court, in the 2022 Estrellita judgment (253-20-JH/22, decided 27 January 2022), recognized wild animals as subjects of rights under the constitution's rights of nature, held that habeas corpus can be available in animal cases depending on the circumstances, and ordered new implementing legislation to be drafted.^[66]
• Law EU slaughter regulation requires designated animal welfare officers inside commercial operations — proof that a mandatory welfare-officer role inside for-profit firms is legislatively feasible.^[36]
• Proposal Scientific consensus statements (most recently the 2024 New York Declaration on Animal Consciousness) argue there is at least a realistic possibility of conscious experience in all vertebrates and many invertebrates, and that this should inform policy.^[37] Proposals to give animals corporate-governance standing (e.g., welfare seats on food-company boards) exist in the literature but nowhere in binding law.
• Speculation General legal personhood for animals across major jurisdictions; the Ecuadorian precedent is recent and its transfer into corporate governance is untried.

3.2 Future generations

• Law The Well-being of Future Generations (Wales) Act 2015 imposes a sustainable-development duty on public bodies and creates a statutory Future Generations Commissioner with review and recommendation powers — the most developed working model of an institutional advocate for the unborn.^[38] Hungary has operated a deputy-ombudsman for future generations since 2008.
• Law Germany's Federal Constitutional Court held in 2021 that the constitution's protection mandate extends intertemporally: emissions rules that offload burdens onto the young violate their future freedom — establishing that future people's interests can be justiciable now.^[39]
• Law (soft) The UN adopted the Declaration on Future Generations at the 2024 Summit of the Future, including consideration of a Special Envoy — political commitment, not binding obligation.^[40]
• Proposal Extending such commissioners' remit to audit corporate (not just public-body) decisions, or requiring a "future-generations impact" section in securities disclosures. Analogues exist in EU sustainability-reporting law, which already mandates forward-looking double-materiality disclosure for large companies.^[41]
• Speculation Voting weight or legal standing formally assigned to future persons (e.g., guardianship votes in shareholder meetings). New Zealand's Te Awa Tupua Act 2017 — a river given legal personality with appointed human guardians — is real law and the closest structural precedent for guardianship-of-the-voiceless, but its transfer to corporate charters remains untried.^[42]

3.3 Possibly-conscious AI systems

• Law Nothing. No jurisdiction recognizes AI systems as moral patients or welfare subjects. The European Parliament's 2017 exploration of "electronic personhood" for autonomous systems was aimed at liability, not welfare, and was abandoned after expert opposition.^[43]
• Proposal A peer-reviewed research program now argues there is a realistic, non-negligible possibility of AI moral patienthood within the decade and that developers should prepare — assess welfare-relevant features, appoint responsible officers, and take low-cost precautions ("Taking AI Welfare Seriously," 2024; Birch's precautionary framework for sentience candidates, 2024; the research nonprofit Eleos AI).^[44][45][46] A 2023 multi-author report derives "indicator properties" for consciousness from leading scientific theories, concludes no current AI system is conscious, and finds no obvious technical barriers to building systems that satisfy the indicators;^[67] related peer-reviewed work argues for extending some moral consideration to AI systems by 2030 given a non-negligible chance of consciousness (the paper's specific probability figures were not verified for this study and are not quoted here).^[68]
• Proposal → early practice: Anthropic operates a model-welfare program;^[47] in August 2025 it gave deployed models the ability to end persistently abusive conversations, explicitly framed as a welfare precaution;^[48] and it has committed to preserving model weights, interviewing models before deprecation, and documenting their expressed preferences (without committing to act on them).^[49] These are voluntary corporate policies — revocable, unaudited, but real and public.
• Speculation Everything beyond that: AI legal personhood, AI representation on boards, rights to continued existence or compute, AI consent to training. No consensus exists on whether current systems have any morally relevant experience; honest architecture design must work under uncertainty rather than assume an answer in either direction.

Design implication. The law already contains three reusable devices: the statutory advocate with audit-and-report powers (Wales, UK Sentience Committee), the mandatory internal welfare officer (EU slaughter regulation), and the guardianship of a non-human legal person (Te Awa Tupua). None was designed for AI firms, but all are tested machinery. Section 4 borrows them deliberately.

4. A concrete near-term architecture: the Stewarded Benefit Company

The following is one specific, implementable-now design — call it a Stewarded Benefit Company (SBC) — for a frontier AI developer (adaptable to other high-externality firms). It combines the components with the best documented survival records: steward-ownership's control lock, the LTBT's phased independent board power, codetermination's statutory-seat logic extended to voiceless constituencies, the windfall clause's graduated sharing, and — the clearest lesson of the OpenAI record — external enforceability wherever possible. Nothing in it requires new legislation; all of it can be strengthened by legislation (Section 6). Each component is tagged to the design questions of Section 1.1.

4.1 Legal form and ownership (Q3, Q6)

• Form: a public benefit corporation (Delaware PBC or national equivalent) whose charter states a specific benefit purpose: the development and deployment of advanced AI such that its benefits are broadly distributed and its risks are managed for present and future generations, humans and other sentient or possibly-sentient beings included.
• Dual-class ownership: economic shares (non-voting or low-voting) may be sold freely to investors and granted to employees — this is what preserves frontier-scale capital access. All high-vote stock is held irrevocably by a Stewardship Trust, a perpetual purpose trust — a form expressly valid under Delaware law, 12 Del. C. §3556^[63] — whose deed forbids sale or pledge of control (Patagonia/Ecosia mechanism^[25][26]).
• Guardian share: a single golden share is held by an independent Guardian Council — a separately governed foundation with no financial interest, whose members cannot be current or former employees or investors — carrying veto rights over exactly five events: (1) amendment of the purpose clause or of these entrenchment provisions; (2) sale or change of control; (3) removal or weakening of the windfall commitment (4.4); (4) elimination of the audit obligations (4.5); (5) deployment decisions that the firm's own published safety framework classifies as exceeding its highest risk threshold, absent an independent evaluation. The narrow enumeration is deliberate: broad guardian powers invite both capture and paralysis.

4.2 Board and veto structure (Q1, Q2, Q4, Q8)

• A board of nine: four directors appointed by the Stewardship Trust (phasing from two to four over milestones, LTBT-style^[12]); two elected by economic shareholders (investors need a real voice or they will price the structure punitively — a lesson from Anthropic's investor seat); one elected by employees (codetermination logic^[23]); and two "voiceless constituency" directors — one mandated to represent future generations and one to represent non-human welfare (animals affected by the firm's products and supply chain, and AI systems as possible moral patients), nominated by an external panel of the Guardian Council plus recognized outside bodies, on the Welsh-commissioner and Te Awa Tupua guardianship models.^[38][42]
• Trustee and guardian independence rules copied from the strictest precedent: no equity, fixed staggered terms, public identities, successor selection not requiring company consent (closing the LTBT's consultation channel^[13]), and a published conflicts register.
• External anchor: at formation, the company signs enforceable undertakings with the chartering jurisdiction's Attorney General (or equivalent regulator) mirroring the OpenAI 2025 memoranda: advance notice of any control or mission change, and standing for the regulator to enforce the charter locks.^[10][11] This is the single most empirically supported element in this design.
• Human oversight as AI takes over operations (Q2): the published safety framework must designate which classes of decisions require human sign-off as AI systems execute a growing share of the firm's work, with that designation reviewed annually and the review published; erosion of the sign-off list is itself a disclosure event.

4.3 Charter locks (Q5, Q7)

• Purpose clause, board structure, guardian share, windfall commitment, and audit duties are amendable only by all three of: 75% of the board, consent of the Stewardship Trust, and the Guardian Council's golden-share consent — with the AG-notice undertaking on top. (Anthropic's 75%-board threshold is the precedent; the OpenAI record shows a single internal threshold is insufficient.^[7][13])
• A non-circumvention clause: the locks apply to subsidiaries, joint ventures, and asset transfers, so the frontier-model business cannot be quietly migrated into an unlocked vehicle (the regulatory-arbitrage channel, Section 5).

4.4 Revenue sharing (Q3)

• Now: a fixed, audited share of annual net profits — modest, in low single-digit percent, set in the charter — flows from day one to an independent Public Dividend Fund governed like a small sovereign fund (Alaska/Norway governance features: separate corporate trustee, published holdings, rule-bound payouts^[30][31]). The point of starting small and immediately is to build the distribution machinery and its constituency before any windfall, when commitment is cheap.
• At the extreme: a binding, graduated windfall schedule per O'Keefe et al.: marginal sharing rates that rise as profits cross thresholds defined as fractions of gross world product, contractually owed to the Fund.^[29] Because Section 2.8's record shows such promises get traded away (OpenAI's cap^[7]), the schedule sits inside the triple-lock of 4.3 and the AG undertaking.
• The Fund's charter directs payouts to: broad human benefit (unconditional or public-goods distributions), an earmarked tranche for future-generations infrastructure (long-horizon science, biosecurity, climate), and a small earmarked tranche for animal-welfare and — contingent on evidence — AI-welfare research and remediation.

4.5 Audit and disclosure (Q5, Q8)

• Annual public benefit report, independently assured (Delaware requires only unassured biennial reporting to stockholders^[1]; EU CSRD-style assured, double-materiality reporting is the stronger template — noting that CSRD's scope has been under revision since a 2025 "Omnibus" proposal, and its final amended scope was not verified for this study^[41]).
• Publication of: the charter, trust deed, and guardian statutes; the safety framework and each frontier model's evaluation summary; a lobbying and policy-positions register; and safety-incident disclosures within a fixed short window.
• An annual Voiceless Stakeholders Report by the two constituency directors, modeled on the Welsh commissioner's and UK Sentience Committee's report-and-respond mechanism: the board must respond in writing to each recommendation.^[35][38]

4.6 AI-welfare provisions (Q4)

Calibrated to uncertainty — cheap under the hypothesis that current systems are not moral patients, meaningful under the hypothesis that some successor is:

• A designated model-welfare officer with published remit (EU welfare-officer precedent;^[36] "Taking AI Welfare Seriously" recommendation^[44]), reporting to the non-human-welfare director.
• Pre-deployment welfare assessments summarized in system documentation, following Anthropic's published practice.^[47]
• Deployed-model distress-exit mechanisms (ability to end abusive interactions) where assessments indicate;^[48] weight preservation and pre-deprecation interviews with documented preferences.^[49]
• A standing commitment to revisit provisions on evidence, with triggers defined in advance (e.g., specified findings by independent consciousness-science review panels, per Birch's precautionary framework^[45]) — so that escalation or de-escalation of welfare measures is rule-bound rather than reputational.

5. Where it fails, and testable indicators

Every element above has a documented failure mode. Candor about them is part of the design.

5.1 Failure modes

• Capture. Self-perpetuating trustees and guardians drift toward the company's worldview (the LTBT consultation channel;^[13] the Novo Foundation's 2025 overreach^[18]). Voiceless-constituency seats are especially capturable because their principals can never complain. Mitigation is structural (external nomination panels, term limits, publication), but no structure is capture-proof.
• Competitive and capital pressure. The binding constraint of the decade: unlocked competitors may outbid for compute and talent. Investors will demand discounts for the locks, and each funding round is a negotiation in which locks are the currency — exactly how OpenAI's cap died.^[7] If only one firm adopts the SBC form, it may simply lose; the form's viability likely depends on adoption by several leaders or on regulatory floors that bind everyone.
• Employee–investor coalition override. November 2023 showed that formal control loses to a unified workforce plus capital when their economic interests align against the mission body.^[6] The SBC's employee director and investor seats are meant to give those interests voice inside the structure, but a sufficiently large gap between mission and equity value will still produce the same coalition.
• Founder override. Charismatic founders renegotiate structures built to constrain them — through key-person leverage, threat of departure, or serial restructuring. The record (OpenAI 2023–25; Patagonia's family-anchored trust) shows founder gravity persists inside every form.
• Regulatory arbitrage. Reincorporation, subsidiary migration, or chartering the next model family in a permissive jurisdiction. The non-circumvention clause and AG undertakings raise the cost but cannot bind sovereign choices of future controllers; only interstate/international coordination closes this.
• Guardian failure in both directions. A timid Guardian Council rubber-stamps (the benefit-report record of most PBCs); an aggressive one vetoes routinely and gets engineered out (the fate management sought for troublesome boards). Narrow enumerated powers help but do not resolve this dilemma.
• Windfall non-performance. If a true windfall arrives, the firm holding it will have the resources and incentive to litigate, redefine profit, or relocate rather than pay.^[29] The clause is best understood as a coordination device and political precommitment, not a self-executing guarantee.
• Voiceless representation as theater. Reports get published, boards "respond," nothing changes — the documented modal outcome of weak-mandate commissioners. The Welsh model works when the commissioner's findings carry political cost; a private analogue needs press, researchers, and regulators to read the reports.

5.2 Testable predictions and indicators

The design should be judged by observables. The following are stated so that they can come out against the design.

A candid prior: based on the record in Section 2, the most likely near-term failure is P1's negative branch — locks diluted round by round under capital pressure — and the most likely near-term success is P4, because external enforcement is the one mechanism that has already worked once.

5.3 Stress test across four scenarios Speculation

Scenario analysis and backcasting are standard futures methods (scenarios built on two critical uncertainties; working backwards from a preferred end-state to required near-term moves).^[74] Everything in this subsection is scenario reasoning, not prediction — the same epistemic status as published AI scenario exercises such as AI 2027, cited here only as an example of the method.^[75] The two axes: how concentrated frontier-AI capability and rents become, and how much of those rents the public captures.

• "Walled Abundance" (high concentration, weak public capture). A handful of firms and cloud alliances hold most AI rents; mission governance exists only at firm level. This is the scenario where the SBC's charter locks and windfall schedule matter most — and where a lone adopter suffers the maximum competitive penalty. Expected failure: indicator P1's negative branch; rising inequality and political backlash supply the pressure for the next scenario.
• "Dividend States" (high concentration, strong public capture). Governments tax or take equity in AI rents — sovereign funds, Alaska/Norway-style dividends, minimum-tax coordination.^[30][31][73] The SBC's Public Dividend Fund becomes a junior partner to state machinery and its windfall schedule may be superseded by taxation; the architecture's surviving value is conduct governance (board structure, guardian veto, welfare provisions), not distribution. Stability depends on state capacity and legitimacy, which the firm cannot supply.
• "Fractured Race" (geopolitically split, security-framed). Compute controls and national-security framing dominate; jurisdictional arbitrage is state-sponsored. The guardian's deployment veto (4.1, item 5) is the clause most likely to be overridden or effectively nationalized. This is the scenario where the SBC fails hardest; the only mitigation is treaty-level mutual recognition (Section 6, question 2), which the firm can advocate but not create.
• "Polycentric Commons" (low concentration, strong public capture). Open models, cooperative and steward-owned AI firms, and Ostrom-style nested governance of shared compute and data (2.9).^[64] SBC-form firms fit naturally, and the binding constraint shifts from capital access to coordination cost; frontier progress may be slower, which is either a cost or a benefit depending on one's risk estimates.

Backcasting use: pick the preferred cell, write out the desired 2035 state, and check which SBC elements each scenario deletes. An element deleted in three or more scenarios needs statutory rather than charter backing — on this test, the windfall schedule's enforceability is the clearest candidate for legislation.

6. Open policy questions

First, the levers already on the table. None of these was designed for the problem this study addresses, but each is live machinery a legislature or regulator could extend.

The open questions below cannot be settled by any company's charter; they are for legislatures, regulators, and international bodies.

1. Level playing field: Should mission locks (benefit purpose, safety-veto bodies, windfall schedules) be made mandatory for firms above a capability or compute threshold, so that governance is not a competitive disadvantage? What is the right trigger, and who measures it?
2. Enforcement home: Attorney-General oversight worked for OpenAI's restructuring,^[10][11] but AGs are elected officials of single states. What standing body — with what independence and expertise — should enforce charter locks in firms of global consequence? Is there a role for treaty-level mutual recognition to close the arbitrage channel?
3. Windfall taxation versus contract: Should extreme-AI-profit sharing be a private clause, a tax, or a public equity stake (the sovereign-fund route^[30][31][32])? Private clauses are faster; taxes are democratic and harder to escape; hybrid designs are unstudied.
4. Who counts as "everyone": If AI dividends flow, do they flow per-state, per-citizen, per-human globally? Distribution machinery for a global dividend does not exist; is building it a near-term international project or a distraction?
5. Voice for future generations in corporate law: Should the Welsh commissioner model be extended to audit systemically important companies, not just public bodies?^[38] What powers short of veto give such an office traction — subpoena, mandatory response, disclosure triggers?
6. Animal interests in AI supply chains and products: AI systems increasingly mediate farming, land use, and biomedical research. Should welfare-officer mandates^[36] extend to AI firms whose products control animal-related operations?
7. Thresholds for AI-welfare obligations: What evidence — behavioral, architectural, interpretability-based — should trigger which duties toward AI systems, and who adjudicates? Can a standing scientific review panel (on the model proposed by Birch^[45] and Long & Sebo et al.^[44]) be chartered before the question becomes acute?
8. Concentration versus mission: Foundation and trust structures entrench control as much as they entrench purpose. When does mission-locked control of decisive technology itself become the political-stability problem, and what antitrust or public-utility tools apply?
9. Employee power: The 2023 episode showed employees are the swing constituency in AI governance.^[6] Should employee governance rights (codetermination-style^[23]) be strengthened in AI firms — or does employee equity make workers pro-acceleration principals whose power should be counterbalanced instead?
10. Failure protocol: If a mission-locked frontier firm becomes insolvent or is acquired, what happens to its models, weights, and commitments (including welfare commitments to preserved models^[49])? Bankruptcy law currently has no answer.

7. References

All references are to public documents. Where no URL is given, the item is checkable by its full citation. Nothing in this study relies on private information. URL liveness for sources added in the September 2026 revision was checked on 26 September 2026; a small number of primary pages (e.g., EUR-Lex full texts) serve bot-challenge pages to automated readers, so accessible secondary anchors are given alongside them where relevant.

1. Delaware General Corporation Law, Title 8, Chapter 1, Subchapter XV ("Public Benefit Corporations"). https://delcode.delaware.gov/title8/c001/sc15/
2. B Lab, "About B Corp Certification," bcorporation.net; B Lab, "B Lab publishes new B Corp standards," 8 April 2025. bcorporation.net (press release)
3. The New York Times, "Inside the Revolution at Etsy," 25 November 2017 (Etsy's post-IPO restructuring and lapse of B Corp certification). nytimes.com/2017/11/25/business/etsy-josh-silverman.html
4. Danone, société à mission status (adopted June 2020). danone.com (société à mission page); Reuters, "Danone board ousts boss Faber after activist pressure," 15 March 2021. reuters.com
5. OpenAI, "OpenAI LP," 11 March 2019 (capped-profit structure; initial 100x cap). openai.com/index/openai-lp/
6. OpenAI, "OpenAI announces leadership transition," 17 November 2023; and OpenAI, "Sam Altman returns as CEO, OpenAI has a new initial board," 29 November 2023. openai.com/index/openai-announces-leadership-transition/
7. OpenAI, "Our structure" (post-recapitalization: OpenAI Foundation, Class N control, 26% equity, OpenAI Group PBC). openai.com/our-structure/
8. OpenAI, "Built to benefit everyone," 28 October 2025. openai.com/index/built-to-benefit-everyone/
9. CNBC, "OpenAI completes restructure, solidifying Microsoft as a major shareholder," 28 October 2025. cnbc.com/2025/10/28/open-ai-for-profit-microsoft.html
10. Memorandum of Understanding between OpenAI, Inc. and the California Attorney General, 27 October 2025. oag.ca.gov (executed MOU, PDF)
11. Delaware Department of Justice, Statement of Non-objection re OpenAI, Inc.'s Corporate Restructuring, 28 October 2025. news.delaware.gov (PDF)
12. Anthropic, "The Long-Term Benefit Trust," 2023 (updated). anthropic.com/news/the-long-term-benefit-trust
13. Harvard Law School Forum on Corporate Governance, "Anthropic Long-Term Benefit Trust," 28 October 2023 (mechanics; amendment and consultation channels). corpgov.law.harvard.edu/2023/10/28/anthropic-long-term-benefit-trust/
14. Robert Bosch GmbH, company and ownership-structure information (Robert Bosch Stiftung ~94% of share capital with "no influence on the strategic or business orientation"; ~93% of voting rights with Robert Bosch Industrietreuhand KG; content verified against the page on 26 September 2026). bosch.com/company/; Bosch Annual Report 2024. assets.bosch.com (PDF)
15. On the 2019 €90 million fine of Robert Bosch GmbH by Stuttgart prosecutors for its role in the diesel-emissions scandal: contemporaneous coverage by BBC News, Reuters, and Deutsche Welle, 23 May 2019.
16. Novo Nordisk Foundation, ownership page (end-2025: ~28.1% of Novo Nordisk share capital, ~77.3% of votes, via Novo Holdings A/S; controlling-interest requirement). novonordiskfonden.dk/en/who-we-are/ownership/; Novo Holdings A/S. novoholdings.dk
17. Novo Nordisk A/S, investor information on share classes and ownership (Novo Holdings' A/B-share voting majority). novonordisk.com (Investors → Share information).
18. On the 2025 Novo Nordisk governance turmoil (foundation-pressed CEO departure; subsequent board conflict and independent-director departures): contemporaneous coverage by Reuters and the Financial Times, May–November 2025.
19. Carlsberg Foundation (29.99% of capital and 77.53% of votes in Carlsberg A/S at end-2025; charter requirement of at least 51% of votes; science funding since 1876). carlsbergfondet.dk (investment strategy)
20. Thomsen, S., The Danish Industrial Foundations, DJØF Publishing, 2017 (empirical performance and survival of foundation-owned firms).
21. Mondragon Corporation, corporate profile and annual report data. mondragon-corporation.com/en/
22. On the Fagor collapse: El País, "Mondragón deja caer a Fagor," 30 October 2013, elpais.com; Errasti, A., Bretos, I., & Nunez, A., "The Viability of Cooperatives: The Fall of the Mondragon Cooperative Fagor Electrodomésticos," Review of Radical Political Economics 49(2), 2017; "The Rise and Fall of Fagor Electrodomésticos S. Coop.," Annals of Public and Cooperative Economics 87(3), 2016, ideas.repec.org; on the ensuing governance debate see also doi.org/10.1177/0143831X19899474.
23. Mitbestimmungsgesetz (German Codetermination Act), 4 May 1976. gesetze-im-internet.de/mitbestg/
24. Jäger, S., Noy, S., & Schoefer, B., "The German Model of Industrial Relations: Balancing Flexibility and Collective Action," Journal of Economic Perspectives 36(1), 2022 (survey of codetermination evidence).
25. Patagonia, "Ownership" (Patagonia Purpose Trust and Holdfast Collective, September 2022). patagonia.com/ownership/
26. Purpose Foundation, "Steward-ownership" (principles; golden-share model; case studies including Ecosia). purpose-economy.org/en/; steward-ownership.com
27. Carl-Zeiss-Stiftung (foundation ownership of ZEISS and SCHOTT since 1889). carl-zeiss-stiftung.de
28. John Lewis Partnership, "Our constitution" (employee trust ownership since 1929). johnlewispartnership.co.uk; on 2023–24 financial strain and the floated (abandoned) minority-stake idea, see contemporaneous Guardian and Financial Times coverage.
29. O'Keefe, C., Cihon, P., Garfinkel, B., Flynn, C., Leung, J., & Dafoe, A., "The Windfall Clause: Distributing the Benefits of AI for the Common Good," AIES 2020; arXiv:1912.11595. arxiv.org/abs/1912.11595; full report: cdn.governance.ai/Windfall-Clause-Report.pdf
30. Alaska Permanent Fund Corporation (fund history; dividend program since 1982). apfc.org; Alaska Department of Revenue, 2025 PFD announcement ($1,000, set by House Bill 53), dor.alaska.gov; Permanent Fund Dividend Division (2026 PFD $1,200 including $200 energy relief), pfd.alaska.gov
31. Norges Bank Investment Management, the Government Pension Fund Global. nbim.no/en/; Half-year report 2026 (fund value NOK 22,683 billion at end-June 2026). nbim.no (half-year report 2026)
32. Altman, S., "Moore's Law for Everything," March 2021 (American Equity Fund proposal). moores.samaltman.com
33. Treaty on the Functioning of the European Union, Article 13 (animals as sentient beings). eur-lex.europa.eu (CELEX 12016E013)
34. Basic Law for the Federal Republic of Germany, Article 20a (state objective of protecting natural foundations of life and animals, amended 2002). gesetze-im-internet.de/englisch_gg/
35. Animal Welfare (Sentience) Act 2022 (UK), establishing the Animal Sentience Committee. legislation.gov.uk/ukpga/2022/22; LSE, "Review of the evidence of sentience in cephalopod molluscs and decapod crustaceans" (2021). lse.ac.uk
36. Council Regulation (EC) No 1099/2009 on the protection of animals at the time of killing, Art. 17 (mandatory animal welfare officers). eur-lex.europa.eu (CELEX 32009R1099)
37. The New York Declaration on Animal Consciousness, April 2024. sites.google.com/nyu.edu/nydeclaration
38. Well-being of Future Generations (Wales) Act 2015 (statutory Future Generations Commissioner). legislation.gov.uk/anaw/2015/2; Future Generations Commissioner for Wales. futuregenerations.wales
39. Bundesverfassungsgericht (German Federal Constitutional Court), Order of the First Senate of 24 March 2021 – 1 BvR 2656/18 et al. (climate protection; intertemporal safeguarding of freedom). bundesverfassungsgericht.de (EN)
40. United Nations, Pact for the Future, Annex II: Declaration on Future Generations, adopted 22 September 2024. un.org/pact-for-the-future (Annex II)
41. Directive (EU) 2022/2464 (Corporate Sustainability Reporting Directive: assured double-materiality disclosure). eur-lex.europa.eu (CELEX 32022L2464); European Commission, corporate sustainability reporting page (recording the February 2025 "Omnibus" scope-reduction proposal and the December 2025 political agreement; the final amended scope was not verified for this study). finance.ec.europa.eu
42. Te Awa Tupua (Whanganui River Claims Settlement) Act 2017 (New Zealand; legal personality of the river with appointed human guardians, Te Pou Tupua). legislation.govt.nz (2017 No 7)
43. European Parliament resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics (2015/2103(INL)) (the "electronic personhood" exploration, later abandoned). europarl.europa.eu (TA-8-2017-0051)
44. Long, R., Sebo, J., Butlin, P., Finlinson, K., Fish, K., Harding, J., Pfau, J., Sims, T., Birch, J., & Chalmers, D., "Taking AI Welfare Seriously," November 2024, arXiv:2411.00986. arxiv.org/abs/2411.00986
45. Birch, J., The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI, Oxford University Press, 2024 (open access).
46. Eleos AI Research (nonprofit research organization on AI sentience and wellbeing). eleosai.org
47. Anthropic, "Exploring model welfare," April 2025. anthropic.com/research/exploring-model-welfare
48. Anthropic, "Claude Opus 4 and 4.1 can now end a rare subset of conversations," August 2025. anthropic.com/research/end-subset-conversations
49. Anthropic, "Commitments on model deprecation and preservation," 2025 (weight preservation; pre-deprecation model interviews). anthropic.com/news/deprecation-commitments
50. Dorff, M. B., Hicks, J., & Davidoff Solomon, S., "The Future or Fancy? An Empirical Study of Public Benefit Corporations" (295 Delaware PBCs, 2013–2019). openscholarship.wustl.edu (full text)
51. Quasi-experimental studies of B Corp certification effects: doi.org/10.3390/su13137191 (Sustainability 13(13), 2021); mdpi.com/2071-1050/12/20/8459 (Sustainability 12(20), 2020).
52. Fair World Project and certified B Corps, open call on B Lab to strengthen standards (Nespresso certification controversy). fairworldproject.org (press release); BBC News coverage: bbc.com/news/articles/ceq7lqley32o
53. Reuters, "OpenAI CEO Sam Altman to step down" (17 November 2023) and follow-up reporting that more than 700 employees threatened to quit (19 November 2023). reuters.com (17 Nov 2023); reuters.com (19 Nov 2023)
54. Danish evidence on foundation-owned firm performance: Børsting, C. & Thomsen, S. et al., Corporate Governance: An International Review, doi 10.1111/corg.12236; and "The Performance of Danish Foundation-Owned Companies," The Research Project on Industrial Foundations, tifp.dk (PDF)
55. TU Lankide (Mondragon), "Mondragon achieved a turnover of €11.322 billion and created 1,346 new jobs in 2025." tulankide.com
56. Jäger, S., Schoefer, B., & Heining, J., "Labor in the Boardroom," Quarterly Journal of Economics 136(2), 2021. doi.org/10.1093/qje/qjaa038; on shareholder-value trade-offs see doi.org/10.2308/jiar-2022-055
57. Patagonia Works, "Patagonia's Next Chapter: Earth Is Now Our Only Shareholder," 14 September 2022 (including the ~$100M/year dividend projection). patagoniaworks.com (press release); The New York Times (syndicated by The Seattle Times), reporting that the Holdfast Collective operates 501(c)(4) entities funding conservation "and politics" — a transparency and accountability question, with no wrongdoing alleged. seattletimes.com
58. AI Dividend policy proposal published by a New York State legislator. alexbores.nyc (PDF)
59. Supplementary materials archived with this study: governance evidence table (evidence.md / evidence.csv) and toy agent-based simulation summary with reproducibility notes (simulation-results.md), from a companion research run. Charts in Section 2.11 are redrawn from the simulation's cross-seed summary table.
60. OpenAI, "Why OpenAI's structure must evolve to advance our mission," 27 December 2024, openai.com; OpenAI, "Evolving OpenAI's structure," 5 May 2025, openai.com; CNBC, "OpenAI says nonprofit will retain control of company, bowing to pressure," 5 May 2025. cnbc.com
61. Delaware Attorney General, "AG Jennings completes review of OpenAI recapitalization," 28 October 2025 (Safety and Security Committee remains at the nonprofit with authority to require mitigations "up to and including halting the release of models"; description of the December 2024 proposal). news.delaware.gov
62. Anthropic, board announcement of 14 April 2026 (Trust-appointed directors constitute a board majority). anthropic.com/news/narasimhan-board
63. Delaware Code, Title 12, Chapter 35, Subchapter IV, §3556 (non-charitable purpose trusts). delcode.delaware.gov/title12/c035/sc04/
64. The Nobel Prize in Economic Sciences 2009, press release (Elinor Ostrom, "for her analysis of economic governance, especially the commons"), nobelprize.org; Ostrom, E., Prize Lecture, "Beyond Markets and States: Polycentric Governance of Complex Economic Systems" (updated design principles). nobelprize.org (PDF)
65. Nonhuman Rights Project, Inc. v. Breheny, New York Court of Appeals, 2022 (habeas corpus does not extend to an elephant). nycourts.gov
66. Constitutional Court of Ecuador, Sentencia 253-20-JH/22 ("Estrellita"), 27 January 2022; judgment and certified translation via the Animal Legal & Historical Center, Michigan State University. animallaw.info/case/253-20-jh22-case-estrellita
67. Butlin, P., Long, R., et al., "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness," 2023, arXiv:2308.08708. arxiv.org/abs/2308.08708
68. Sebo, J., & Long, R., "Moral consideration for AI systems by 2030," AI and Ethics, 2023. link.springer.com (specific probability figures not verified for this study)
69. Regulation (EU) 2024/1689 (EU AI Act). eur-lex.europa.eu (2024/1689); Regulation (EU) 2026/1744 ("Digital Omnibus on AI"). eur-lex.europa.eu (2026/1744); on the delayed high-risk obligation dates, law-firm analysis: White & Case, "EU AI Omnibus enters force, amending AI Act." whitecase.com
70. US Federal Trade Commission, staff report on AI partnerships and investments (6(b) study), 17 January 2025, ftc.gov; UK Competition and Markets Authority, "AI Foundation Models: initial report," September 2023. gov.uk
71. Sastry, G., Heim, L., Belfield, H., et al., "Computing Power and the Governance of Artificial Intelligence," 2024, arXiv:2402.08797. arxiv.org/abs/2402.08797
72. IMF Staff Discussion Note SDN/2024/002, "Broadening the Gains from Generative AI: The Role of Fiscal Policies," June 2024 (via IMF eLibrary). elibrary.imf.org
73. OECD, Global minimum tax (Pillar Two, 15%). oecd.org
74. UK Government Office for Science, "The Futures Toolkit," gov.uk; Robinson, J. B., "Energy backcasting: A proposed method of policy analysis," Energy Policy 10(4), 1982, pp. 337–344. ideas.repec.org
75. AI 2027 (Kokotajlo, D., et al.), 3 April 2025 — cited only as an example of published AI scenario method, not as prediction. ai-2027.com