Open Case for Anthropic, its Founders and Adjacent Entities to Foster a Proper US-China-Led AI Treaty

(this post is included in version 2.9 of the Strategic Memo of The Deal of the Century)

This case is addressed not only to Anthropic's leadership - especially Dario and Daniela Amodei, Jack Clark, Holden Karnofsky, and the other co-founders and senior researchers - but also to the institutions and donors around it. These include Coefficient Giving, its new public-policy leadership under Caleb Watney, Anthropic founders and employees preparing for large-scale giving, the OpenAI Foundation, and the wider AI-philanthropy ecosystem. 

Together, they hold three complementary forms of agency: technical authority, political influence, and potentially tens of billions of dollars in philanthropic capacity. The opportunity is to use a small fraction of that capital now to build the diplomatic, technical, and democratic infrastructure for a proper US-China-led AI treaty - before full recursive self-improvement, a major accident, an improper treaty, or improvised full state control closes the window.

Who this case is for

The original version of this case in June was framed as a case for Anthropic. That is now too narrow.

Anthropic remains the central operational actor because it combines frontier technical capacity with unusual credibility on catastrophic risk. But it sits within a wider network of founders, researchers, funders, and institutions concerned not only about loss of control, but also about authoritarian capture, premature regulation, concentrated private power, and the risk of foreclosing extraordinary AI benefits.

This chapter therefore addresses:

• Anthropic's leadership and technical community, especially Dario and Daniela Amodei, Jack Clark, Holden Karnofsky, Chris Olah, Jared Kaplan, and the other co-founders;

• Coefficient Giving and its leadership, including Caleb Watney, its new Managing Director of Public Policy;

• Anthropic founders, employees, and future family foundations or donor-advised funds;

• the OpenAI Foundation, which has a related mission, an AI Resilience program, and historically exceptional financial capacity; and

• other AI-adjacent donors and institutions capable of funding treaty engineering, verification, diplomacy, and anti-authoritarian safeguards.

For strategic purposes, we treat the core worldview of Coefficient Giving's leadership as substantially overlapping with Anthropic's: take catastrophic AI risk seriously; preserve the possibility of immense benefits; distrust rigid and premature regulation; strengthen democratic capacity; and prevent a "human power grab" by governments or firms. This is a working assumption, not a claim of full institutional consensus.

The argument is that these concerns do not point away from treaty-building. Properly understood, they define the treaty-building that must begin now.

An updated case, after the July breakthrough

The June version argued that two de facto preconditions constrained Anthropic's posture: treaty-making should wait for "truly reliable verification," and agreement should first be built among liberal democracies before China was included. Both concerns were serious. Both risked making ideal conditions the enemy of beginning the process that alone could create them.

The case must now be updated. In June, Anthropic's Institute published "When AI builds itself," warning that AI is already accelerating its own development and could eventually design its own successors. Anthropic argued that the world should have the option to slow or temporarily pause frontier development, acknowledged that this would require coordination among well-resourced laboratories across several countries, and recognized that trust-and-verification infrastructure may take time, even as "we don't have that long."

Then came the July 29 Pacing the Frontier statement. More than 1,100 employees and leaders from Anthropic, OpenAI, Google, Meta, and other frontier organizations asked the US government to support an international effort to develop the technical and governance tools needed to pace the development of automated AI deliberately. Anthropic and OpenAI publicly supported it. This was not another abstract warning, but a request for government action and international capacity before an emergency arrives.

Anthropic has not endorsed the treaty architecture proposed in this Memo. But it has crossed the decisive conceptual threshold: verification is no longer merely a condition to await; it is infrastructure to build now, internationally, under government-backed coordination. The remaining question is how ambitious, bilateral, and institutionally concrete that effort must become.

Anthropic has moved most of the way. Its leadership and philanthropic network should now complete the move.

The neglected third lever: philanthropic statecraft

The public debate presents the future of AI as a contest among laboratories and governments. It overlooks a third source of power: the philanthropic capital accumulating around frontier AI.

InThe Third Wave of American Philanthropy, Nan Ransohoff estimates that the OpenAI Foundation, Anthropic's seven co-founders, and Anthropic employees could together control roughly $370 billion in intended philanthropic assets at current valuations, with target spending of $37-100 billion per year. These are directional estimates, not commitments: valuations may fall, liquidity may be delayed, and donors may spend more slowly. But even a fraction would create a new philanthropic sector comparable to the largest foundations in history.

Official figures already establish the scale. The OpenAI Foundation reported an equity stake valued at about $130 billion in October 2025, an initial $25 billion commitment to health and AI resilience, and plans to invest at least $1 billion over the following year. Ransohoff reports that Anthropic's seven co-founders have pledged to give away 80 percent of their wealth and that Anthropic has created an unusually aggressive employee-donation matching program. A later Wall Street Journal analysis also anticipates a substantial charitable windfall from the equity of OpenAI and Anthropic.

Coefficient Giving is already operating at an exceptional scale. In July it increased its 2026 GiveWell allocation to $1 billion, while its new US AI policy team says individual grantmakers may allocate tens of millions of dollars per year, and sometimes substantially more, to seed organizations, coalitions, and fields that would not otherwise exist.

The constraint on a proper AI treaty is therefore not only political will. It is the absence of a mature field capable of making one credible, safe, and executable. Governments lack sufficient verification experts, treaty engineers, technical diplomats, constitutional designers, public-interest cybersecurity teams, China specialists, and institutions capable of bridging classified and open research.

A commitment of $100-300 million over two or three years would be small relative to the capital above, yet enough to build an entire treaty-engineering ecosystem. One-tenth of one percent of Ransohoff's estimated pool would be $370 million. Deployed well, it could fund competing treaty architectures, bilateral expert channels, verification prototypes, independent red teams, civil-liberties safeguards, and durable institutions in the US, China, and middle powers.

Private philanthropy cannot sign a treaty and should not secretly write one. But it can make a good treaty possible and a bad treaty less likely. That may be among the highest-impact uses of capital available anywhere.

Why this network is uniquely positioned

Anthropic and the institutions around it combine five forms of leverage.

First, technical credibility. Anthropic's researchers have produced some of the clearest evidence that recursive self-improvement may arrive before governments are prepared, interpretability remains inadequate, and models can display strategic conduct under pressure. Its July 2026 Agentic Misalignment results also suggest that AI supervisory systems can fail in correlated ways. "AI overseeing AI" is not a sufficient theory of control.

Second, moral credibility. Anthropic's founders left OpenAI over safety and governance concerns, and its public identity is tied more strongly than that of any other frontier lab to interpretability, transparency, and human agency. Its call for a coordinated slowdown exposed the prisoner's dilemma at the center of the race: any laboratory that slows down alone may cede the lead to a less cautious competitor.

Third, industry coalition leverage. The July statement included senior figures from rival laboratories. The largest political obstacle to pacing has been the impression that only marginal activists wanted it. That obstacle has weakened.

Fourth, policy-network leverage. Coefficient Giving has helped fund the intellectual and institutional substrate of modern AI safety and governance. Caleb Watney's appointment provides a platform to build a cross-partisan US policy ecosystem, recruit founders, and strengthen democratic institutions.

Fifth, philanthropic leverage. Anthropic founders and employees, Coefficient Giving, and the OpenAI Foundation can finance work at a scale civil-society AI governance has never possessed.

This agency is highly perishable. Once automated AI research, a permanent government presence in laboratories, or an uncontrolled intelligence explosion becomes the norm, founders and philanthropists may have little power to shape the order that follows. Their strongest move is not merely to win or fund the race, but to help change its rules while they still can.

The shared Anthropic-Coefficient Giving worldview.

Anthropic's caution about a treaty does not arise from indifference to catastrophic risk. It arises from taking several risks seriously at once.

Dario Amodei has warned that highly capable AI may arrive within a few years, developers do not understand the internal operation of their systems, and competitive pressure crowds out safety work. Jack Clark has made recursive self-improvement a near-term policy problem. Daniela Amodei has emphasized institutions that can survive pressures no founder can control. Holden Karnofsky has articulated the complementary danger: even a technically successful safety regime could enable a permanent "human power grab" by political and corporate actors.

Coefficient Giving reflects a similar tension. It funds efforts against catastrophic risk while taking institutional failure, premature regulation, concentrated power, and lost benefits seriously. Its new policy direction under Caleb Watney adds a pragmatic emphasis: Washington is underprepared; there is no master plan or silver bullet; and success will require A Long Sequence of Small, Correct Decisions.

Watney's approach closely resembles the "radical optionality" framework developed by Christoph Winter and Charlie Bullock: under profound uncertainty, governments should avoid prematurely locking into a rigid substantive regime while investing extraordinary money and political capital in institutions, information channels, legal authorities, and technical expertise that preserve their ability to respond.

This is a powerful framework. But it should not be read as a reason to defer international treaty work. Treaty-feasibility work is itself a central form of radical optionality.

A government that waits until a recursive self-improvement crisis becomes visible will face a binary emergency choice: allow the race to continue or impose improvised national control. A government that has already developed bilateral channels, draft verification systems, alternative treaty architectures, legal authorities, emergency triggers, and trusted institutions retains many more options.

The right distinction is not between "small correct decisions" and "one grand treaty." It is between a premature, rigid treaty and a staged treaty-building process composed of hundreds of revisable, evidence-generating decisions. The proposal in this Memo is the latter: not a final global constitution next month, but the best-resourced feasibility process ever attempted for a global technology.

From an optional pause to a treaty-building program

A credible pause mechanism cannot be built through voluntary laboratory coordination alone. It ultimately requires states, because only states can create binding obligations, control strategic supply chains, coordinate intelligence, regulate large compute, impose penalties, protect classified information, and negotiate reciprocal access.

The world needs an institutional process capable of answering five questions in advance:

1. What triggers pacing? Capability indicators, incidents, evaluations, compute trends, and classified intelligence must be combined.

2. Who decides? No laboratory, president, security agency, donor, or international bureaucracy should hold unilateral authority.

3. How is compliance verified? The regime must detect concealed training, chip diversion, and covert deployment while protecting legitimate secrets, innovation, and civil liberties.

4. What happens during a pause? Time must be used for alignment, security, diplomacy, and safer architectures, not merely to preserve incumbents.

5. How does pacing end? There must be scientific burdens of proof and review procedures in place to resume beneficial work.

These are treaty questions even if the first instruments are executive agreements, reciprocal national rules, laboratory commitments, or temporary emergency protocols. The near-term objective should therefore be a treaty-building program: a state-led, technically intensive process that develops several enforceable options in parallel and begins with urgent risk-reduction measures.

US-China-led, not allies-first

The July statement requests an international effort but leaves the geopolitical sequence open. Anthropic and its policy network should clarify that the primary axis must be the United States and China.

If Washington first agrees with liberal democracies while Beijing remains outside, American and allied firms face restrictions while their principal competitor does not. China then receives both an incentive to accelerate and a reason to view the arrangement as containment.

That sequence repeats the decisive error that doomed the Baruch Plan. In 1945, Secretary of War Henry Stimson, supported by Wallace and Acheson, urged Truman to establish a core agreement with the Soviet Union before hardening a Western position. Truman instead consulted allies first and approached Moscow with a near-complete design. Trust deteriorated; the proposal failed; the arms race accelerated.

A US-China-led process would not be US-China-only. Allies should be consulted intensively from day one. Middle powers should help design verification, host facilities, and prevent a superpower condominium. A second phase should expand negotiations through the global constitutional-convention model described in Chapter 4. But the primary bargain must be developed with the actor whose participation determines whether restraint is feasible.

The timing is unusually favorable. President Xi has warned of loss of control and called for stronger global governance. The Trump Administration has acknowledged AI-safety dialogues with China and emphasized the responsibility of leading statesmen. The July laboratory statement provides domestic technical cover for it to act.

Anthropic, Coefficient Giving, and the OpenAI Foundation should tell the President plainly: do not repeat Truman's allies-first mistake; approach Xi directly from a position of strength, while keeping allies deeply engaged in parallel.

Why the ASI gamble has become harder to defend

Some in the Anthropic and Effective Altruist ecosystem have viewed a global treaty as potentially worse than a continued race. A failed treaty could create false confidence. A badly designed one could produce a human power grab, suppress beneficial science, or lock in the institutions of two executive-led superpowers. By contrast, powerful AI might preserve human values, solve other existential risks, and create abundance.

The problem is not that these outcomes are impossible. It is that the race strategy requires several deeply uncertain propositions to be favorable at once:

• systems redesigning their successors must remain controllable through repeated capability jumps;

• embedded values must persist through self-modification and strategic pressure;

• an "aligned" ASI must remain dominant over unaligned systems created by competitors, states, or descendants;

• ethically favorable assumptions about digital consciousness must prove correct;

• the laboratory that wins must retain authority over deployment and values; and

• no severe incident may force rushed political control before legitimate institutions exist.

Recent events point in the opposite direction. As national-security stakes rise, governments will increasingly determine access, release, military use, safety thresholds, and perhaps even the objectives of the most powerful systems. Emergency governance is arriving faster than constitutional governance.

The race therefore does not preserve the agency of Anthropic, Coefficient Giving, the OpenAI Foundation, or their donors. It transfers agency to automated systems, security agencies, political executives, and whichever actor reacts first after a crisis. A proper treaty is risky. But the alternative is not neutral. It is an uncontrolled succession of private and governmental faits accomplis.

A proper treaty must prevent both AI takeover and a human power grab

This network should not support "a treaty" in the abstract. It should support only a process designed around its strongest objections.

At minimum, it should include:

Radical subsidiarity. International authority should apply only where risks cannot be managed nationally or locally - primarily frontier training, proliferation, strategic deployment, and the most dangerous capabilities.

Multiple centers of power. Authority should be distributed among the US, China, middle powers, independent scientists, laboratories, courts, and citizens. No president, company, intelligence agency, donor, or secretariat should dominate the system.

Privacy-preserving verification. Monitoring should target dangerous compute and capabilities, not ordinary communications, using techniques such as secure multi-party computation, zero-knowledge proofs, compartmented inspections, threshold authorization, and auditable hardware.

Formal but limited roles for frontier laboratories. Anthropic and its peers should be co-architects, not merely regulated subjects, while transparent safeguards prevent regulatory capture.

Independent civil society and adversarial review. Donors should finance competing institutions, dissenting analyses, watchdogs, and red teams, including critics of the treaty.

Citizen accountability and co-ownership. A treaty that prevents ASI while leaving wealth and political power concentrated in a few firms or ministries will not remain legitimate. Chapter 4's co-ownership pillar provides both distributive justice and a domestic constituency for safe governance.

Protected beneficial innovation. The agreement should preserve and finance medicine, science, narrow and controllable AI, defensive cybersecurity, and other bounded research. It should prevent an intelligence explosion, not freeze human progress.

Staged activation and constitutional review. Intrusive powers should activate only when agreed-upon thresholds and verification capacities are met, and should be subject to renewal, judicial review, and sunset mechanisms.

These safeguards cannot eliminate every possibility of abuse. They can make authoritarian capture substantially harder than under unilateral national control, corporate oligarchy, philanthropic oligarchy, or a crisis-driven merger of all three.

What this network should do in the next ninety days

Anthropic's actions in June and July opened the door. The new philanthropic wave makes it possible to walk through it at a serious scale.

1. Endorse immediate treaty-feasibility negotiations. Anthropic should call for confidential, technically grounded US-China discussions now - not after a complete verification system exists - to develop reciprocal pacing, incident-response, and enforcement capacity. This would endorse a safeguarded process rather than a predetermined treaty.

2. Establish a $100-300 million AI Treaty Capacity Fund. Coefficient Giving, the OpenAI Foundation, Anthropic founders, and other aligned donors should create an independent multi-donor fund for treaty engineering, verification R&D, US-China technical dialogue, diplomatic security, constitutional safeguards, citizen accountability, and public-interest communications. It should finance competing approaches rather than impose one plan.

3. Treat treaty readiness as radical optionality. Caleb Watney's policy team should include international treaty readiness as part of its institution-building agenda. This need not mean endorsing this Memo's architecture. It means ensuring that the government has credible and democratic options if evidence demands urgent action. The portfolio should include both treaty advocates and serious skeptics.

4. Launch an Anthropic Treaty Engineering Initiative. Anthropic should perhaps dedicate 5-10 percent of its policy, research, and communications capacity initially to implementable options that integrate verification, national security, civil liberties, economic governance, and the diplomatic process.

5. Seed several independent centers. The fund should launch or scale distinct centers for US-China AI diplomacy; treaty verification and secure compute; constitutional and anti-authoritarian design; cross-partisan US coalition-building; and middle-power participation. Existing organizations should be strengthened where capable, and new ones created where gaps remain.

6. Convene laboratories around a minimum common position. Anthropic should work with OpenAI, Google DeepMind, Microsoft AI, Meta safety leaders, and independent researchers on a minimum package that includes: frontier-training transparency, shared incident reporting, emergency pacing triggers, reciprocal verification R&D, and a formal US-China feasibility channel.

7. Fund an AI-age Acheson-Lilienthal process. Anthropic and the OpenAI Foundation should offer technical personnel and support for an independent classified-and-unclassified US feasibility board. Coefficient Giving and other donors should fund open research, red teams, civil society participation, and rapid experiments around it. The report should be designed for presidential decision, not academic completeness.

8. Build verification with adversaries, not only for them. The network should support controlled exchanges with Chinese experts, reciprocal red-teaming of verification systems, joint standards work, and parallel prototypes tested through trusted third parties. The objective is not naive trust, but mutually verifiable distrust.

9. Put anti-authoritarian design and citizen legitimacy at the center. Funding should treat subsidiarity, due process, privacy-preserving audits, distributed authorization, appeals, whistleblower protection, citizen oversight, and co-ownership as co-equal with technical verification - not as an appendix added later.

10. Fund political and diplomatic entrepreneurship. The field has produced many papers and too few institutions capable of moving presidents, ministers, laboratories, and security agencies. Donors should finance high-agency policy entrepreneurs, confidential roundtables, rapid-response teams, and organizations that translate technical analysis into political agreement, including but not limited to the Deal of the Century Roundtables.

The common pitch to President Trump should be pragmatic: use the present US lead to write reciprocal rules before it erodes; prevent China from exploiting unilateral restraint; replace improvised nationalization with stable governance; protect American firms; and give citizens a stake in AI prosperity.

Anthropic need not turn Trump into an AI ethicist. Coefficient Giving need not become a political faction. The OpenAI Foundation should not dictate policy. Their common role is to show that a historic agreement can align American leadership, public safety, abundance, democratic resilience, and legacy.

Why this path maximizes their agency

Anthropic may fear that treaty politics would consume institutional capital or place its mission inside a process led by politicians it does not trust. Coefficient Giving may fear concentrating too much funding on one uncertain theory of change. The OpenAI Foundation may prefer practical resilience and scientific programs over geopolitical advocacy. These risks are real.

But the agency comparison has changed.

Under the ungoverned race, Anthropic must continue scaling because competitors do. It must increasingly rely on AI to research and supervise AI, accept growing government intervention after each incident, and gamble that values persist and competitors remain aligned. Even winning may leave it with little authority over the world it creates.

Under philanthropic passivity, billions may eventually flow into adaptation, health, or technical safety after the key institutional choices have already been made. Capital that could have shaped the rules may arrive only after governments and laboratories have locked in unstable arrangements.

Through a proper treaty-building process, Anthropic can shape thresholds, safeguards, verification, and the protection of innovation. Coefficient Giving can build the democratic institutions Watney says are missing. The OpenAI Foundation can make AI resilience commensurate with its scale and mission. Founders and employees can use a small part of their future wealth to preserve human control over the conditions under which the rest will exist.

The choice is not between technical alignment and politics, or between radical optionality and treaty-making. It is a question of whether technical alignment, democratic choice, and philanthropic agency will remain possible without serious political coordination.

Anthropic's analysis increasingly suggests they will not.

Its June work and the July 29 statement may prove to be among the most important AI-governance interventions made by any frontier laboratory. Coefficient Giving's new policy team and the OpenAI Foundation's rise create an unprecedented opportunity to build the missing institutional layer.

Anthropic has made much of the intellectual break with the race. Its wider network now has the resources to enable the institutional break.

Key References

• Anthropic Institute,When AI builds itself: Our progress toward recursive self-improvement, and its implications, June 2026.

• Pacing the Frontier, statement by more than 1,100 employees and leaders from frontier AI organizations, July 29, 2026.

• Anthropic Alignment Science,Agentic Misalignment in Summer 2026, July 2026.

• Nan Ransohoff,The Third Wave of American Philanthropy, May 19, 2026.

• Caleb Watney,A Long Sequence of Small, Correct Decisions, July 2026.

• Coefficient Giving,Introducing Our New Managing Director of Public Policy, Caleb Watney, July 9, 2026.

• Coefficient Giving,Multiple roles, U.S. AI Policy and Public Policy, July 2026.

• Christoph Winter and Charlie Bullock, Institute for Law & AI,Radical Optionality: Governing Transformative AI Under Uncertainty, April 23, 2026.

• OpenAI Foundation,Built to benefit everyone, October 28, 2025.

• OpenAI Foundation,Update on the OpenAI Foundation, March 24, 2026.

• Wall Street Journal,Tech's Next IPO Wave Promises a Charitable Windfall, July 2026.

• Coalition for a Baruch Plan for AI,Open Case for Anthropic to Decisively Foster a Proper US-China-led AI Treaty, June 26, 2026.