// analysis
Who will govern AI? Capital, mission and the state
Anthropic, OpenAI and the state: a source-based investigation of control rights, public equity, national security and accountability for AI decisions.
Sources and reading context
Topics
Contents11 sections
On 26 October 2026, Anthropic filed the prospectus for its initial public offering. Among the disclosures for prospective investors is an unusual distribution of power: a single share reserved for a founders’ company carries 50.1% of the votes on general matters, while another share class enables a trust to select a majority of directors. The offering remains proposed, with its price and number of shares still undetermined in this filing. Several of the mechanisms described are intended to take effect around the IPO. A01 A08
This structure opens a much larger debate than protecting founders from the demands of the stock market. It requires us to trace decisions to the people entitled to make them: who can authorise a military use, suspend a model, change a safety rule, negotiate with a computing provider or rewrite the company’s mission? At each stage, a different actor may have influence over the outcome.
The investigation leads to a central proposition: AI governance depends on the combination of rights, resources and remedies. An equity stake provides a claim on a share of economic value. Powers of appointment, access to information, veto and amendment draw a different map. Industrial contracts and the powers of the state complete the picture. Understanding who is in charge requires reading these maps together.
Government ownership would add another holder of rights to this arrangement. It could create a capacity for public action, give citizens a share of returns and stabilise funding. It would also introduce conflicts between the asset’s financial performance, protection of people, public procurement and military objectives. The strategic challenge is to establish how these conflicts will be resolved before a crisis imposes its own timetable.
This documentary investigation is current to 28 October 2026. It distinguishes legal provisions, company announcements, political proposals and interpretation. Scenarios and recommendations are identified as such; proposed transfers to the public retain their respective levels of confirmation.
Anthropic: two majorities within one company
Understanding the arrangement starts with separating two elections. Shareholders vote on certain company matters; different share classes elect different members of the board. A majority in the first arena can therefore coexist with a different majority in the second.
The F share is to be held by Anthropic Founders, LLC, whose initial members will be the seven co-founders. Its voting weight adjusts to maintain 50.1% on general matters, including the election of directors other than Class T directors. A shares carry one vote each and collectively account for the remainder of this general vote. The founders also retain the votes attached to their own A shares. The 50.1% figure therefore describes the power of the F share alone. It identifies an institutional mechanism for securing control, independently of the economic size of individual holdings. A03 A04
Within the LLC, an ordinary matter can be decided by four matching instructions from seven eligible members. If there is no majority meeting the agreement’s conditions, the F share follows the votes cast by A holders proportionally, subject to a quorum. A four-to-three split produces a majority; abstentions or dispersed instructions can produce a different outcome. Decisions requiring stronger approval have their own rules. Founder cohesion thus becomes an observable governance variable, with different consequences depending on the matter put to a vote. A04
The Long-Term Benefit Trust, or LTBT, holds another source of authority. Its T share allows it to select four of the board’s seven authorised seats. The other three are elected through the A and F vote. The prospectus names six directors; the seventh seat, a Class T seat, is vacant. The trust selects directors; those directors then serve on the board, subject to the duties of their office. The distinction matters: choosing the people who make decisions is a different form of power from day-to-day management. A01 A03 A05
Data and scope
Under the proposed IPO architecture, the F share alone carries 50.1% of general votes. The remaining 49.9% belongs to Class A, including founders’ individual holdings. Class T elects four of seven directors; Classes A and F elect three. The S-1 names six directors and leaves one T seat vacant. The “specified” T seat follows a special procedure. F and T carry no ordinary economic participation: their liquidation distribution is capped at par value of $0.00001 per share.
The protective logic is understandable. A market-funded company may devote time and resources to safety even when investors demand faster commercialisation. A separate body can retain the ability to appoint the people responsible for that judgement. The same structure also entrenches a private group’s power over choices with collective consequences. Both effects can occur at once.
An investor must therefore assess what comes with the economic exposure: the quality of a team, a mission that may sustain trust, and limits on the ability to replace decision-makers. A citizen faces a different question: how can someone outside the company obtain explanations, have harm examined or challenge a decision? The share price and public accountability answer different tests.
The employee role, including when it becomes effective
One T seat, designated the Specified Class T Director, follows a particular procedure. For certain appointments or removals without cause, the nominating committee can withhold approval. The trust may then withdraw its proposal or request a vote by E holders. Approval requires a majority of E votes cast, with a quorum consisting of a majority of outstanding E shares. The agreement also addresses the situation in which no E shares have yet been issued: the committee’s decision is then final. The prospectus specifies that there will be zero E shares at closing, with later issuance on a timetable still to be determined. A06 A01
This employee voice is weighted. The announced allocation combines one E share for each six-month period of employment with one E share for each A share pledged to the donation programme; co-founders are excluded. E shares participate in specified matters and cease to exist when F ends. Length of service, philanthropic commitments and the issuance timetable therefore help define the electorate. A01 A03
This is an institutional choice that deserves examination in its own right. The company involves some of its employees in safeguarding its mission and rewards a form of commitment to it. Their influence nevertheless depends on criteria chosen by the organisation. Assessing this counterweight will require knowing how many shares are actually issued, how they are distributed, turnout in votes and how the procedure is used. The legal mechanism creates a possibility; its exercise will provide evidence of its effectiveness.
Who can change the guardian’s rules?
The most revealing clause in a corporate constitution is often the one that allows it to be amended. It identifies whose agreement becomes necessary when the mission, the market and management cease to pull in the same direction.
Anthropic’s exhibits provide several routes to amend, merge or terminate the LTBT. One relies on agreement between the trust and the board, with enhanced approval thresholds and rounding rules. Others allow coalitions to override trustee opposition. During a window lasting until the third anniversary of the IPO, and ending sooner if F terminates, unanimous founder approval can combine with a majority of E votes cast under a 40% quorum, followed by a majority of A votes cast under the applicable quorum. Another route remains available for as long as F exists: two-thirds of founders, rounded up : five out of seven : 75% of E votes cast under the applicable quorum, and then the required A majority. Both routes operate in parallel during the initial window. After F ends, the prescribed threshold becomes 85% of outstanding A shares. A05 A03 A04
Data and scope
Routes to amend or terminate the trust without its consent combine separate class votes and different denominators. During the initial window, both the unanimous-founder route and the two-thirds route are available, subject to their respective conditions. The 7/7 and 5/7 figures illustrate seven eligible founders. After F sunsets, the 85% threshold covers all outstanding A shares. The figure omits the route involving consent from the trust and board, discussed in the article; other legal and charter requirements continue to apply.
These thresholds make revision demanding. They also reveal the arrangement’s underlying philosophy: the guardian of the mission can be reshaped by a coalition of founders, employees holding E shares and shareholders. Denominators matter as much as percentages. A majority of votes cast under a quorum and 85% of all outstanding shares require very different levels of mobilisation. The figure preserves that distinction so the procedure can be checked.
F control is itself durable and conditional. The agreement allows voluntary termination with enhanced majority approval. Having fewer than three qualified members for 90 consecutive days begins a 450-day transition. The tenth anniversary features in rules concerning, among other things, certain successors; F does not have a general automatic expiry on that date. The transfer of power therefore depends on conditions governing qualification, succession and share ownership. A04
Governance risk arises in two directions. High thresholds can protect a safety strategy against temporary commercial pressure. They can also prolong a flawed decision when the coalition needed to correct it cannot be assembled. More accessible amendment routes allow adaptation while providing a way to curtail the guardian’s powers when its intervention becomes costly. Analysis must examine the kind of conflict each safeguard is designed to resolve.
A public-interest mission, with remedies defined by private law
The public benefit corporation, or PBC, form imposes a duty on directors to balance shareholders’ financial interests, the interests of those materially affected and the public benefit specified in the charter. Delaware law also sets limits on standing to enforce that obligation: a holding of 2% of the shares or, for a company listed on a national securities exchange, the lesser of that threshold and shares with a market value of two million dollars. Other legal avenues may exist depending on the harm; this particular remedy depends on shareholder status. C02
Anthropic’s trust, meanwhile, is constituted without beneficiaries as defined by its agreement. Its terms can be enforced by the trustees, the company, the F holder or a group meeting a particular ownership threshold for one year. Members of the public whom the mission is intended to benefit remain outside that circle merely by virtue of their interest in the mission. A05
This arrangement calls for a distinction between independence, expertise and representation. Trustees may possess outstanding expertise and carefully protected financial independence. Selection through private mechanisms gives them a contractual mandate. Political representation requires a different chain: a public mandate, accountability to institutions and an opportunity to challenge decisions under shared rules.
The issue is particularly acute for a technology distributed across borders. People exposed to a model include users, workers, competitors, people represented in its data and populations affected by its governmental uses. A company’s mission may seek to take their interests into account. Giving those people the means to defend their interests themselves requires additional institutions and rights.
The power to stop must survive its exercise
OpenAI’s November 2023 crisis offers a useful historical test. The board announced Sam Altman’s removal on 17 November. On the 19th, Microsoft offered to welcome Altman, Greg Brockman and colleagues into a team with resources. A decision to bring him back followed on the 21st; on the 29th, OpenAI presented his return and a reconstituted initial board. Event dates and announcement dates differ, which explains some confused timelines. C06 C07 C08 C09
The institutional lesson that can be drawn from this sequence concerns the organisation’s continuity after a difficult decision. A board may have the right to dismiss a leader while depending on people, contracts and resources that make the decision difficult to sustain. A coalition capable of reassembling a team and its infrastructure elsewhere has an exit option that weighs on negotiations. This is an interpretation of the balance of power, distinct from the existence of a formal veto.
The review commissioned from WilmerHale, a summary of which OpenAI published on 8 March 2024, attributed the episode to a breakdown in trust between the former board and Sam Altman. That attribution and scope must be preserved: the available document is the summary published by the company. The documentary record does not establish that the dismissal resulted from a catastrophic technical discovery. C09
OpenAI’s current structure provides more specific answers to several of these difficulties. The Foundation appoints and can remove the directors of OpenAI Group PBC. The commitments accompanying the 2025 restructuring, recorded with the Attorneys General of California and Delaware, establish particular rights for the Foundation. On safety and security matters, they require the PBC board to consider only the mission. The Safety and Security Committee, or SSC, has an effective approval right and can halt a release, including when the applicable thresholds would otherwise allow it to proceed. C01 C03 C04
Material resources are part of the arrangement. The commitments provide access to information and reasonably necessary resources. In the disagreement over those needs contemplated by the services agreement, the Foundation’s determination prevails. Several changes to control, mission or special rights must be notified to the Attorneys General at least 21 days in advance. These authorities exercise public oversight of the commitments; the notification clause retains its specific scope. C03 C04
Data and scope
The October 2025 commitments connect Foundation control of the board, the Safety and Security Committee’s effective authority over releases, and access to information and reasonably necessary resources. The 21-day notice to California and Delaware authorities concerns specified structural changes; it serves a different purpose from an operational model-release decision. The diagram represents documented powers and does not certify their implementation.
This case shows why the label “foundation” alone says little about the actual strength of a safeguard. A right to halt a launch becomes more credible when accompanied by accessible evidence, independent expertise, a budget, the ability to enforce the agreement and operational continuity. OpenAI’s documents cover several of these dimensions. Their effectiveness in a future conflict will need to be established through observable evidence: decisions, resources mobilised, delays, corrections and reports.
Industrial dependence adds another layer. The FTC staff report published in January 2025 on the Microsoft–OpenAI, Amazon–Anthropic and Google–Anthropic partnerships describes cloud commitments, information exchanges and the costs of changing providers, among other features. Its contractual evidence runs through September 2024, supplemented by public information through January 2025; it must be used with that date in mind. Anthropic’s S-1 also describes possible conflicts involving partners that combine the roles of investor, supplier, distributor, customer and competitor. C10 A01
These relationships evolve. Microsoft’s 27 April 2026 announcement opens distribution of OpenAI products across multiple clouds and makes the intellectual-property licence non-exclusive through 2032. Other commercial and financial ties remain in the statement. A serious map of governance must therefore track agreements over time: the ability to move a computing workload or change partners can alter the available choices long before the board changes. C05
Plans to share AI capital
The idea of giving the public a stake in AI ownership has a documented history. In his personal essay of 16 March 2021, Sam Altman proposed an American Equity Fund, financed in part by an annual levy of 2.5% on the value of large companies, payable in shares. His stated argument linked wider ownership to support for growth. This earlier proposal provides useful context for reading those made in 2026. P01
OpenAI’s industrial-policy paper published in April 2026 proposes a Public Wealth Fund, with diversified assets and the possibility of distributing returns to citizens. Its endowment would be determined with public officials and companies. In its June economic framework, Anthropic considers universal capital accounts and, under a scenario of greater disruption, AI-linked sovereign wealth funds or equity-sharing mechanisms. These documents set out policy options and their justifications. The precise rights of a government shareholder remain to be defined in these proposals. P02 P04
A Financial Times report on 2 July, citing two people briefed on the matter, described preliminary discussions about transferring 5% of OpenAI to a public vehicle. No completed transfer of this stake was publicly confirmed in the material found by 28 October. P06
On the same day, Reuters reported, citing a source, that Anthropic and the administration had at that point held no discussions about an equity stake. That information is dated; later exchanges are a separate question. It calls for company-by-company attribution when discussing “AI leaders”. P07
Anthropic’s private fund: a different institutional arrangement
Anthropic introduced a Public Benefit Fund on 18 October. The prospectus of the 26th provides a more advanced account: it says N shares have been contributed to this LLC, which is 100% owned by the company. The amount represents 5% of a fully diluted base as of 30 June 2026, assuming full conversion of Amazon’s notes. That base differs from the final post-IPO share capital. The future contributions contemplated remain discretionary and conditional. P05 A01
This fund has a public-benefit purpose and economic resources. The LTBT holds the T governance rights. A sovereign wealth fund would be owned or controlled through public institutions. All three arrangements can pursue a sharing objective, but their decision-makers, beneficiaries and accountability mechanisms differ. A rigorous account must follow ownership of the vehicle as closely as its name.
Four questions reveal the substance of an equity-sharing proposal: where do the shares come from; who legally holds them; who exercises the votes; and what rule governs the distribution of income? A citizen may receive a payment without exercising voting rights. A fund may hold economic rights and delegate decisions. A state may obtain a specialised veto with very little financial exposure. The contract connects : or separates : these functions.
An invitation to the state can serve several purposes
The reasons publicly advanced by the laboratories are clear: sharing expected gains, helping people through changes in work and widening access to the technology’s benefits. Their merits will depend in part on the value actually created, its distribution and the costs borne by affected people. Forecasts of prosperity remain the forecasts of their authors. P02 P04
Our analysis identifies several additional possible incentives. Public ownership could strengthen a laboratory’s legitimacy and make its relationship with the state more durable. It could also give the authorities a financial interest in its success, facilitate discussions about infrastructure and create a coalition favourable to its development. Depending on the rights negotiated, this alliance could support the public interest or weaken the distance required for effective oversight.
These mechanisms are hypotheses to test against actual agreements. Establishing a particular consideration in return : a contract, guarantee, regulatory treatment or privileged access : would require documentary evidence. An offer to transfer shares creates a subject for investigation; the exact intentions, concessions and distribution of power remain open questions.
The same ambiguity runs through political statements. Interviewed by TIME on 28 September, in a conversation published on 1 October, Donald Trump ruled out nationalising the major laboratories and then expressed openness to taking a stake in OpenAI or Anthropic along the lines of Intel. The terms remained undetermined. P08
The state as shareholder: three very different ways to gain power
The word “nationalisation” becomes useful when the transaction is specified. Acquiring or receiving shares can pursue financial returns, influence over decisions or control of the company. A special share can create a veto over a limited set of actions. Legislation can mandate a transfer of ownership. Taking operational control requires further powers and safeguards. These choices involve different institutions, even when they concern the same laboratory.
The Sanders bill: new shares and an active public mandate
Bill S. 4825, introduced in the Senate on 18 June 2026 and referred to the Finance Committee, proposes an AI sovereign wealth fund at the Treasury. The version examined is an introduced bill; enactment has not been established in this investigation. Its scope covers AI, computing infrastructure, data centres and advanced robotics activities, subject to the definitions and revenue threshold specified. P09 P10
The mechanism deserves a precise reading. The tax would be paid through the issuance of new shares, bringing the Treasury’s holding to 50% after the transaction, proportionately across existing classes. The bill overrides certain internal restrictions on issuance and also provides for a tax on later issuances, including some equity compensation. Its ambition therefore concerns both an initial acquisition and maintaining a position in the share capital. P09
The text combines this mandatory transfer with structural separation. Section 4 would require the FTC to enforce the separation of covered activities, with restrictions on ownership by groups engaged in other businesses, financing, common executives and certain joint ventures. It provides a deadline of 90 days after the applicable starting point. For an activity integrated into a larger group, the implications would therefore extend beyond a new shareholder to the boundaries of the company and its industrial relationships. This proposed provision would require its own examination of implementation and effects; it could profoundly alter existing relationships between laboratories and suppliers. P09
A fictional example makes the difference tangible. A single-class company has 100 privately held shares. To give the fund half of the shares through a new issuance, it must create 100 for the fund: 100 out of 200 equals 50%. Issuing only 50 shares would give it 50 out of 150, or 33.33%. By contrast, donating five existing shares transfers a 5% interest within an unchanged total of 100. Share price, company value and special rights are outside this example.
Data and scope
Two fictional companies start with 100 private shares and a single class. Transferring five existing shares gives the recipient 5%. Issuing enough shares to give it 50% after the transaction requires 100 new shares, bringing the total to 200. Dots represent shares, with no price or valuation model. The 5% example illustrates a possible transfer; it does not reconstruct the terms of discussions reported by the Financial Times. The issuance mechanism in S. 4825 remains a legislative proposal.
Anthropic’s rules immediately show the limits of simply transplanting percentages. A structure in which certain shares carry class rights, separate elections and an adjustable control mechanism would require a specific legal examination. Calculating half of the shares is insufficient, by itself, to establish how many directors a fund could appoint at this company.
The Sanders bill nevertheless intends the fund to play an active role. A seven-member commission would exercise voting rights and seek to elect board representatives according to the rights held. The text assigns it objectives including public safety, labour, competition and the environment. A provision protects pursuit of these objectives in the exercise of fiduciary duties, even when they conflict with the company’s financial interest. This is a proposal to redirect the purpose for which a director’s powers are exercised. P09
A fair examination of the bill also requires looking at its safeguards. It provides for disclosure of votes, rules on conflicts and recusal, a prohibition on coordinating the competitive conduct of rival companies and a ban on using the fund to bail them out. These provisions address identifiable risks. Their effectiveness would depend on the mandate, implementation, oversight resources and available sanctions. P09
The relationship between public wealth and spendable income deserves the same attention. The bill authorises annual appropriations equal to 5% of the fund’s average market value, less reasonable administration costs, with Congress determining their allocation. It specifies that distributions must not require the sale of shares. Returns actually received as dividends, liquidity and authorised appropriations are different quantities. A higher valuation can increase reported wealth while leaving the financing of a cash payment unresolved. P09
Intel and U. S. Steel: read the rights before the size of the stake
The Intel agreement announced in August 2025 offers one precedent. It describes a stake corresponding to 9.9% under the terms of the transaction, with some shares held in escrow. That percentage must retain its date and scope: it is not a measure of ownership on 28 October 2026. The announcement describes a passive stake with no board seat. The agreement generally requires voting in line with the board’s recommendations, subject to exceptions relating in particular to the agreement and the government interests concerned. P11 P12 P13
The transaction also includes a warrant exercisable if Intel ceases to own at least 51% of its foundry, subject to the exceptions in the agreement. The right can be settled in cash or shares under the contractual terms. It can therefore make a future separation more costly and influence choices about the industrial structure alongside voting rights. P11
At U. S. Steel, a G share, or golden share, issued to the government under the national security agreement carries different rights: it is non-economic, allows the designation of an independent director and provides targeted consent rights over certain industrial decisions. The June 2025 documents describe these protections; financial statements for the first quarter of 2026 confirm issuance. Here, power comes from the special right attached to the share. P14 P15
Data and scope
Intel’s August 2025 announced agreement provides for a 9.9% stake, including escrowed shares, with a passive position, no negotiated board seat and votes generally aligned with the board. That percentage describes the terms announced in 2025. At U. S. Steel, one non-economic G share carries nomination of an independent director and targeted consent rights. The comparison concerns rights attached to the agreements, without deriving an overall ranking of public control.
These precedents provide a method for reading any future agreement with a laboratory. The relevant questions will concern the class received, reserved decisions, appointments, accessible information, exit conditions and events that trigger new rights. Two stakes of equal value can create opposing distributions of power. Conversely, an instrument with no claim on profits can profoundly alter freedom over industrial decisions.
National security already provides another route for intervention
The state also exercises power through procurement relationships and public authority. The conflict between Anthropic and the Department of War offers a documented example, which must be read along two separate judicial tracks.
In early 2026, the dispute concerned, among other things, restrictions on mass domestic surveillance and fully autonomous weapons. Directives targeted the administration’s use of the technology and its relationships with suppliers. Two statutory bases gave rise to two cases with different scopes. R01 R02
On 27 August 2026, federal judge Rita Lin, in the Northern District of California, largely ruled in Anthropic’s favour on the measures before her, including the designation under 10 U.S.C. § 3252 and the broader directives. Her findings included violations involving freedom of speech, procedural protections and administrative law. The judgment preserves the government’s choice of suppliers within the law; it examines the legality of extending sanctions beyond that choice. R01
On 25 September, in a separate proceeding, the D.C. Circuit rejected by two votes to one the challenge to the supply-chain exclusion under 41 U.S.C. § 4713. The majority held that the supplier’s restrictions and uncertainty over the model’s functional availability could fall within the risk addressed by that statute. The dissent defended a narrower reading of manipulation and the powers it authorises. R02
These decisions concern separate proceedings, rather than a judgment followed by its reversal on appeal. They reveal two routes, two statutory bases and different interpretations of the place of usage safeguards in a national-security relationship. A petition for rehearing of the second decision was filed on 13 October; the S-1 of the 26th mentions it. The material reviewed through 28 October establishes no subsequent outcome of that petition. R03 A01
Data and scope
The February and March 2026 measures lead to two proceedings with different scopes. The California judgment of 27 August addresses broad directives and the designation under 10 USC §3252. The D.C. Circuit’s 25 September ruling concerns procurement exclusion under 41 USC §4713 and upholds it by two votes to one. It does not overturn the California judgment. An en banc rehearing request was filed on 13 October; its later outcome has not been confirmed in the reviewed documents.
The case requires taking both conflicting responsibilities seriously. A supplier seeks to constrain the uses of a system it considers capable of causing harm. Military authorities want to rely on the resources they incorporate into operations and to define the lawful uses they authorise. The friction concerns the allocation of authority when software incorporates restrictions, contracts add to them and the state considers its operational autonomy impaired.
The September majority and dissent make a precise legal disagreement visible. How far can a power to protect the supply chain reach into usage limits chosen by a supplier? What procedure must precede exclusion? What distinction should be maintained between a commercial disagreement, subversive conduct and a national-security risk? The institutional stakes extend beyond which party wins: they concern the criteria for deciding and the authority entitled to review the decision.
The documents also require a technical clarification. The case distinguishes restrictions encoded in the model from remote control after delivery to classified systems. A supplier’s power depends on the deployment model, updates, access to weights and interfaces. A narrative attributing a universal secret kill switch to the laboratory would misrepresent the documents examined. R02
The California judgment also records discussion of the Defense Production Act before the challenged measures. This reference documents an instrument being considered. Its actual invocation, legal scope and judicial review remain separate questions. Above all, the episode shows the range of powers a government may seek to use when dealing with a strategic technology. R01
For a laboratory’s governance, the consequence is concrete: a mission’s resilience also depends on revenues, the ability to sustain litigation, industrial alternatives and the legal security of its contracts. A safety body may have a strong mandate while facing pressures that originate entirely outside share ownership.
The state as owner, customer and referee
Government investment could bring several of these powers within the same authority. The finance ministry would want to preserve the value of its holding. A public buyer would seek a capable and available service. The regulator would have to enforce requirements concerning safety, competition and rights. The executive might pursue military or industrial objectives. These mandates can converge; their divergence is precisely the situation for which governance must be designed.
The analytical risk is capture in both directions. A laboratory might receive favourable treatment because its value forms part of public wealth. An executive might use a stake, an appointment or a contract to steer uses to its advantage. Financial proximity creates incentives and opportunities for intervention. Establishing capture in a particular case requires evidence about decisions, procedures and reciprocal benefits.
It would be equally reductive to overlook the capabilities public ownership can provide. Well-defined information rights can improve decisions. A long-term investor can finance infrastructure whose benefits are widely dispersed. Revenues can support transition policies. A public mandate can give institutional weight to affected people. Outcomes will depend on the chosen structure and the quality of implementation.
The OECD specifically recommends a clear separation between ownership, regulation and policymaking, alongside impartial remedies and objective procurement conditions. The Santiago Principles propose explicit objectives, defined governance and operational autonomy with accountability for sovereign wealth funds. These standards provide useful foundations; applying them to AI requires specifying the conflicts particular to the technology. P16 P17
Data and scope
This analytical view separates the state’s roles as customer, owner and regulator. Each exchanges different resources or information with the same laboratory, pursuing objectives that can come into tension. Separate mandates, traceable decisions and impartial review are proposed safeguards informed by the cited public-governance principles. Arrows represent possible mechanisms without attributing intent to any actor.
A credible public structure should therefore make the author of each decision identifiable. The portfolio manager would exercise shareholder rights under a published mandate. The regulator would retain its powers, resources and separate route for review. Procurement would follow verifiable criteria. Conflicts of interest would lead to recusals and an auditable record. Safety would have a reporting and expert-review channel capable of bringing forward information unfavourable to the valuation.
These proposals demand more than a promise of dividends. They take seriously the possibility that a safety measure could temporarily reduce the fund’s income. A body whose political legitimacy rests on an annual payment could come under pressure to preserve that payment. Governance must therefore establish who bears the financial cost of a pause, how uncertainties are disclosed and what procedure governs reconsideration of a decision.
Concentrating competing assets within one fund raises another question. Representation of a common long-term interest can support research or shared infrastructure. Simultaneous access to strategic information and commercial coordination require precise limits. The Sanders bill recognises this problem by prohibiting certain forms of coordination; an effective organisation would have to make that boundary enforceable. P09
Finally, the nationality of the public owner defines the mandate’s scope. A fund charged with paying American citizens represents interests defined by American law. Models can affect people elsewhere. International governance must therefore address externalities, information and avenues for cooperation alongside the domestic distribution of returns. National public ownership resolves part of the distribution question while leaving open the question of whose interests are represented.
Europe has powers that need the means to work
The European perspective calls for combining law, expertise and material capacity. The AI Act’s chapter on general-purpose AI models imposes specific obligations. For models presenting systemic risk, Article 55 requires evaluation, risk assessment and mitigation, serious incident reporting and cybersecurity protection. Article 88 assigns supervision and enforcement of the chapter to the Commission, implemented through the AI Office and subject to procedural safeguards. The consolidated text consulted incorporates the July 2026 amendments. R04 R05
Timetables must remain tied to their respective categories. The general-purpose model regime has applied since 2 August 2025. Article 111(3) gives providers of models already placed on the market before 2 August 2025 until 2 August 2027 to comply. R08 Other provisions, including those for certain high-risk systems, follow different transitions after the 2026 amendments. The general label “AI Act” therefore covers several deadlines. R06
Effective oversight then depends on resources. Receiving technical documentation requires the ability to interpret it, question its assumptions, arrange a review that tests competing views and protect sensitive information. An access right gains value when a team has the time, skills and computing resources to use it. Public expertise thus becomes an infrastructure for sovereignty.
EuroHPC already provides routes to computing access for eligible companies and scientific projects, with several allocation modes and conditions for access free of charge. These arrangements provide a material foundation for alternatives and expertise. Their existence should be distinguished from evidence of equivalence with the resources of the largest laboratories. R07
For French and European authorities, our analysis suggests three priorities. The first is to make obligations verifiable, with lasting budgets and expertise. The second is to negotiate public contracts that provide for continuity, reversibility, access to evidence and responsibilities in the event of an incident. The third is to support a diversity of research and supply capabilities, so that public decisions can draw on multiple interlocutors and a credible possibility of substitution.
Public investment in a European company can contribute to this effort. Its usefulness should be judged by the rights secured and the capabilities created: expertise, access, control over sensitive changes, continuity and accountability. A supplier’s nationality, a company’s headquarters and the location of a data centre each reveal a different dependency. Contracts must connect them to the decisions the authority wants to be able to make.
Write the rules before the crisis
Anthropic’s documents reveal a real effort to make a mission resilient to certain market pressures. OpenAI’s commitments show how a right to halt activity can be accompanied by resources and public oversight. Proposals for citizen funds seek an answer to the distribution of value. Industrial precedents and the military litigation show, finally, that the state has several routes for influencing a company. Analysis of these arrangements benefits from testing them against concrete conflicts.
Consider an explicitly hypothetical scenario: a laboratory receives concerning evaluation results a few weeks before a major launch. A public customer is waiting for the model, a partner has reserved computing capacity and a shareholder fund anticipates income. For the decision to remain governable, the tests must reach the competent body, that body must be able to assess them, its decision must be executable and someone must bear its cost. A route for review must allow a disagreement to be examined while preserving the effectiveness of precautionary measures.
This test leads to six requirements we propose for assessing a governance structure.
Provide independent access to evidence. Unfavourable results, incidents, evaluation limits and dissenting views must be able to reach safety officials and the competent authorities. Independence also requires knowing who funds the expertise, who defines its scope and who may publish conclusions within legitimate confidentiality limits.
Attach resources to decision-making power. An identifiable budget, expertise of its own and a continuity plan allow a body to sustain a halt or a demand for correction. Resource agreements are particularly important when the oversight body depends on the company it oversees for its operations.
Define reserved decisions and routes for review. Suspending a launch, changing a military use, transferring model weights, amending a mission and replacing an overseer may require distinct procedures. Timetables, criteria and the authority responsible for appeals should be understandable before conflict arises.
Make changes to governance visible. The actors able to remove a safeguard, the quorums and the transitional periods deserve the same publicity as its creation. A safeguard presented as durable should remain understandable when a coalition seeks to amend it.
Separate public mandates that can diverge. Returns on a holding, a buyer’s needs, industrial policy and protection of rights call for identifiable responsibilities. Rules on appointment, recusal, publication of votes and independent oversight should allow their interaction to be examined.
Represent affected people beyond security holders. Consultation, protected reporting, expert scrutiny that tests competing views and legal remedies provide avenues beyond share ownership. Their combination should encompass consequences outside the circle of founders, employees and financial beneficiaries.
These requirements form an analytical framework, without an artificial safety score. They call for verifiable evidence: a clause, an accountable person, a budget, a decision, a report, a remedy. They also help specify the questions that remain open on 28 October: final IPO terms, actual issuance of E shares, implementation of oversight commitments, the content of any public equity agreements and the next stages of the litigation.
Public ownership can become one tool in this arrangement. Philanthropy, special mission rights, regulation and public capabilities can provide others. Their value will be measured by a concrete possibility: having an important decision examined, securing a change when the evidence demands it and holding those with the power to decide accountable for their choice.
Further reading
Meta’s financing and Class C shares illustrate the distinction between economic participation and voting control. AI financing partnerships examine dependencies between customers, suppliers and capital providers.
Sources
The 40 cited documents retain their dates and status. Each entry gives source locators, scope and limitations.
- Anthropic, PBC : Form S-1
Passages and scope
Cover; pp. 13–17, 74–75, 144–145, 243–253, 257, 293–304.
Initial prospectus: capital, voting powers, board, trust and public benefit fund.
Proposed offering. Several arrangements are designed for IPO completion; price and offering size remain blank.
- Exhibit 3.3 : Form of Amended and Restated Certificate of Incorporation
Passages and scope
Articles III and VII; Article V, especially § 5.1(jj) and §§ 5.2–5.7.
Proposed charter: Classes A, E, F, N and T, voting rights and amendment mechanisms.
Filed form for the proposed IPO; completion remains prospective.
- Exhibit 99.2 : Form of Limited Liability Company Agreement of Anthropic Founders, LLC
Passages and scope
§§ 15, 16(a) and (d), 22; definitions in § 34(j), (k), (aa), (qq), (ww).
Founders’ LLC agreement: instructions for Class F and the required majorities.
Form with its effective date left blank; enhanced thresholds differ from ordinary decisions.
- Exhibit 99.1 : Form of Amended and Restated Long-Term Benefit Trust Agreement
Passages and scope
§§ 1.3, 2.1–2.7, 4.3–4.5 and 5.1–5.2.
LTBT agreement: director selection, amendments, duration and enforcement.
Proposed IPO restatement; the restatement date is blank in the filed document.
- Exhibit 10.14 : Form of Stockholders’ Agreement
Passages and scope
§§ 1.6, 2.1–2.5 and 3.5.
Stockholders’ agreement: nomination for a specified Class T seat and approval procedure.
Signing planned before IPO completion; effective on closing; conditional procedure.
- Anthropic files S-1 with the SEC for proposed initial public offering
Passages and scope
Filing announcement and proposed status of the offering.
Official announcement of the S-1 filing on 26 October 2026.
Company announcement to be read with the prospectus and exhibits.
- Our structure
Passages and scope
“Updated structure”, “Our Board”, “Equity Holders”.
Overview of OpenAI’s structure and Foundation control over the board.
Live page. Equity percentages describe the closing of the 2025 recapitalisation.
- Delaware General Corporation Law : Subchapter XV, Public Benefit Corporations
Passages and scope
Title 8, Chapter 1, Subchapter XV, §§ 362, 365, 366 and 367.
Delaware public benefit corporation law: purpose, duties and enforcement.
General statutory regime, supplemented by company-specific charters and contracts.
- Executed Memorandum of Understanding : California Attorney General and OpenAI
Passages and scope
§§ 2, 7–11, 13–20, 24, 27 and 28; PDF pp. 2–6.
Signed California commitments: control, safety, resources, information and notification.
Signed on 27 October 2025. Notice of specified structural changes is not general model approval.
- Statement of No Objection : OpenAI corporate restructuring
Passages and scope
§§ 1, 4–9 and 11–18; PDF pp. 2–6.
Delaware’s conditions for non-objection to the restructuring, including safety powers.
Commitments and representations governing the restructuring; future implementation remains to be observed.
- The next phase of the Microsoft–OpenAI partnership
Passages and scope
“Amended Agreement Provides Long-Term Clarity”, five bullet points.
Announced amendments to the Microsoft–OpenAI partnership in April 2026.
Public contract summary; the revenue-sharing rate and total cap are undisclosed.
- OpenAI announces leadership transition
Passages and scope
Opening announcement and the board’s stated rationale.
Announcement of Sam Altman’s departure, dated 17 November 2023.
The board’s position at that date, distinct from a judicial determination of the facts.
- A statement from Microsoft Chairman and CEO Satya Nadella
Passages and scope
19 November statement and 21 November update, on a page dated 21 November.
Satya Nadella’s statements during the November 2023 crisis, including Microsoft’s offer to host a team.
The page collects statements from several dates; the displayed time’s zone is unspecified.
- Sam Altman returns as CEO, OpenAI has a new initial board
Passages and scope
Messages from Sam Altman and Bret Taylor; 8 December review update.
Detailed announcement of Sam Altman’s return and the new initial board.
Published on 29 November 2023; the return agreement dates to 21 November.
- Review completed & Altman, Brockman to continue to lead OpenAI
Passages and scope
“Summary of WilmerHale review & findings”.
OpenAI’s published summary of the WilmerHale review of the governance crisis.
Board-commissioned review summarised by the company; its full report is outside this corpus.
- Partnerships Between Cloud Service Providers and AI Developers : FTC Staff Report on AI Partnerships & Investments 6(b) Study
Passages and scope
Executive summary; §§ 4.2–4.4 and 5.1–5.3.
FTC study of financial and operational relationships between cloud providers and AI developers.
Staff report: contracts available in September 2024 and public information through January 2025; no illegality finding.
- Moore's Law for Everything
Passages and scope
Part 3, “Capitalism for Everyone”.
Sam Altman’s personal essay proposing, among other measures, an annual equity contribution by large companies.
A 2021 policy proposal and its author’s forecasts.
- Industrial Policy for the Intelligence Age: Ideas to Keep People First
Passages and scope
Cover; p. 7, “Public Wealth Fund”.
OpenAI’s economic policy proposals, including a public fund holding diversified assets.
PDF dated April 2026; financing and governance arrangements remain exploratory.
- Anthropic’s Economic Policy Framework
Passages and scope
pp. 8 and 13–14.
Anthropic’s economic framework: options for sharing equity and wealth.
Policy document dated June 2026; the text establishes no firm equity transfer to the state.
- Investing in the public benefit
Passages and scope
Paragraphs of the 18 October 2026 announcement.
Announcement of Anthropic’s public benefit fund and its intended funding.
The 26 October S-1 updates the completed contribution, denominator and further-contribution conditions. A private Anthropic-owned fund.
- OpenAI proposes handing Trump administration 5% stake
Passages and scope
Opening paragraphs.
News report.
Paywalled FT article, also consulted through its authorised Yahoo Finance syndication. No completed agreement established as of 28 October.
- Trump administration and Anthropic have not discussed the government taking a stake in it, source says
Passages and scope
Opening paragraphs.
Reuters report that Anthropic and the administration had held no such talks as of 2 July 2026, according to a source.
Attributed, time-bounded reporting; Anthropic declined comment.
- Read the Full Transcript of Donald Trump’s 2026 Interview With TIME
Passages and scope
Interview section.
Transcript.
Remarks attributed to their author; interview and publication dates are distinguished in the registry.
- S.4825 : American A.I. Sovereign Wealth Fund Act, introduced text
Passages and scope
§ 3: proposed IRC §§ 5000E and 9512, Commission duties; § 4.
Introduced S. 4825 bill: share issuance, public fund, governance and structural separation.
Bill introduced on 18 June 2026; described effects are conditional on enactment.
- S.4825(IS) : Content Details
Passages and scope
“IS” version; “Last Action Date Listed”.
Official record of the introduced version of S. 4825 and referral to the Finance Committee.
Record for this version; no enactment established in the collection closed on 28 October.
- Warrant and Common Stock Agreement
Passages and scope
Recitals, § 1.1, § 4.6 and Annex A.
Intel–Commerce agreement: voting alignment, exceptions and the foundry-related warrant.
Agreement dated 22 August 2025; historical terms, distinct from a current ownership statement.
- Intel and Trump Administration Reach Historic Agreement to Accelerate American Technology and Manufacturing Leadership
Passages and scope
Paragraphs on funding, passive ownership and the warrant.
Intel’s 2025 transaction announcement: intended passive ownership and escrowed shares.
The announced 9.9% includes escrow; the figure describes the August 2025 terms.
- Intel Form 10-Q for quarter ended 27 September 2025
Passages and scope
“US Government Agreements”; risk factors; quarter ended 27 September 2025.
Intel’s Form 10-Q confirming closing timing and ownership reported at 27 September 2025.
Historical data: closing on 27 August, 5.8% ownership at 27 September 2025.
- Condensed Consolidated Financial Statements, three months ended 31 March 2026 and 2025
Passages and scope
Note on p. 5, the seventh PDF page: NSA and “Golden Share”; quarter ended 31 March 2026.
U. S. Steel financial statements confirming the non-economic Class G share.
The registry event date refers to the 13 June 2025 NSA; the PDF’s exact publication day is unverified.
- Nippon Steel Corporation and U.S. Steel Finalize Historic Partnership
Passages and scope
pp. 2–3: NSA, director nomination and consent rights.
Nippon Steel–U. S. Steel closing announcement describing golden-share rights.
Document dated 18 June 2025 despite its filename; targeted powers over enumerated decisions.
- OECD Guidelines on Corporate Governance of State-Owned Enterprises 2024 : ChapterIII
Passages and scope
Chapter III, recommendations III.A–C and III.G.
OECD principles on separating state ownership from regulatory functions.
Normative framework used to develop the article’s recommendations.
- Sovereign Wealth Funds: Generally Accepted Principles and Practices : Santiago Principles
Passages and scope
GAPP 6–10 and 16–21.
Santiago Principles: mandate, governance, transparency and management of sovereign wealth funds.
Voluntary 2008 principles, subject to applicable law; their existence does not establish implementation.
- Anthropic PBC v. U.S. Department of War : Order on cross motions for summary judgment
Passages and scope
Case 3:26-cv-01996-RFL, Document 250; pp. 1–2, 21–23 and 57–59.
California judgment on directives and the designation affecting Anthropic, including under 10 USC § 3252.
A separate proceeding from the D.C. Circuit FASCSA case; the two decisions concern different legal tracks.
- Anthropic PBC v. United States Department of War : Opinion and dissent
Passages and scope
Nos. 26-1049 and 26-1162; majority pp. 4–5, 20–21, 42–43; dissent, PDF pp. 44–51.
D.C. Circuit decision upholding exclusion under 41 USC § 4713 by two votes to one, with dissent.
Interpretation contested by the dissent and a rehearing petition. The decision establishes no remote kill switch.
- Anthropic seeks rehearing of case over Pentagon's security risk branding
Passages and scope
Opening report; petition filed on Tuesday 13 October.
Report on Anthropic’s rehearing petition filed on 13 October 2026.
Published on 14 October; the filing is also noted in the 26 October S-1. No later outcome established in the corpus.
- AI Act, Article 55: Obligations of providers of general-purpose AI models with systemic risk
Passages and scope
Article 55(1)–(3); page identifies the consolidated text as of 27 July 2026.
AI Act Article 55: duties of providers of general-purpose AI models with systemic risk.
Official Commission reproduction; the authentic instruments remain those in the Official Journal.
- AI Act, Article 88: Enforcement of the obligations of providers of general-purpose AI models
Passages and scope
Article 88(1)–(2); consolidated text as of 27 July 2026.
AI Act Article 88: Commission competence for Chapter V obligations.
Competence defined for this chapter, distinct from other national or European powers.
- AI Act : Application timeline
Passages and scope
“Application timeline”; reference to the AI Omnibus entering into force on 27 July 2026.
Official AI Act application timeline, including GPAI obligations.
Live explanatory page; precise obligations must be read in consolidated law.
- AI Factories Access Modes
Passages and scope
“AI for Industrial Innovation”; “AI for Science and Collaborative projects”.
EuroHPC AI Factories access conditions for industrial innovation and research.
Access depends on eligibility and allocation; no capacity equivalence to a particular laboratory follows.
- AI Act, Article 111: transitional provisions
Passages and scope
Article 111(3).
Transition for providers of general-purpose models already placed on the market before 2 August 2025.
Compliance by 2 August 2027 for this scope. Other systems and obligations follow their own provisions.
This analysis is not investment advice.
// cite this analysis
l0g, “Who will govern AI? Capital, mission and the state”, l0g.fr, published October 28, 2026, updated October 28, 2026, https://l0g.fr/en/analysis/ai-governance-anthropic-state-ownership/
$ cd ../analysis