What it is. For 18 months we have been laser-focused on preparing the US Administration and key US AI policy powerbrokers for that AI treaty scenario, through strategic memos, one-on-one engagements and small consultations. From the start we assumed it would be a radically unprecedented undertaking in scope, timeliness and diplomatic bandwidth. Given China's steady signals toward global governance, our sole focus has been to help the US build a proper, widely shared negotiating position — toward China first, then the middle powers, then everyone else.

Our assumption. We assume the reader agrees that safety and power-concentration risks are immense and urgent, plausibly materializing within months or years, and that the path to preventing them is fraught with comparable risks.

Where things stand. The President has moved from a hands-off to an executive-first policy — a safety order, the leading lab's latest model suspended on NSA-backed grounds, a self-policing Accord, and a Super Intelligence Force reporting around February 1. The labs lined up behind Amodei's pacing plan, which ends in coordination with China yet undercuts it. Xi asked to "forestall loss of control"; 29 heads of state and the EU called for an institution that can verify and convene. Trump said no — a big no, or an opening bid. Next windows: Bessent's November round with Beijing; the G20 at Doral, December 14–15.

What we do. Strategic Memos — 350 pages on the treaty-making process itself, 150 on the powerbrokers — and closed-door Consultations where delegates of the offices that shape US AI policy co-edit one draft. A coalition map records where they converge; if enough do, they take the pen and pitch the President together. We facilitate; the positions are theirs.

Who we are. The Coalition for a Baruch Plan for AI, convened by the Trustless Computing Association: 9 NGOs and 34 advisors and contributors, with US 501(c)(3) status via fiscal sponsorship since March. Since 2025 we have engaged 95+ staff and advisors of the powerbrokers' offices, including over 30 hours with key senior AI staff reporting directly to four of the twelve identified key US AI policy powerbrokers. We are funded by world-renowned AI safety experts — seed-funded by Jaan Tallinn's Survival and Flourishing Fund — on a $7,500/month burn rate, and supported by over 2,100 hours of professional pro-bono work.

Only a Proper Treaty. A proper treaty, not any treaty: a bad one could even be worse than none, all risks and opportunities considered. A durable concentration of power could be a result, but it is already the default without one. The main safeguards are in a designed, balanced treaty-making process, not an improvised one. Enforcement socio-technical mechanisms are paramount: intelligence-grade superpowers' state-of-the-art is good enough to start on the most urgent risks, while it jointly improved at a wartime pace, elbow-to-elbow, by security agencies, labs and citizen jury oversight.

The ask. $60,000 funds six months; 959K–1.75M, an 18-month scale-up (Roadmap). Send a delegate to the Consultations, or a name and who could introduce us.

Where things stand

The President: from hands-off to executive-first. Until spring 2026, the President’s hands-off AI policy proved right: America stayed ahead, with no major accidents. Then a steady rise in capabilities and incidents made it unsustainable, exposing a huge gap in regulatory and treaty preparedness. Trump began to act, seeking a new policy that balances many crucial US interests — over the opposition of a close-knit camp of advisors allied with a few key accelerationists and post-humanists. He issued a first executive order on AI safety; his Commerce Department suspended the leading US lab’s latest model for weeks, on classified, NSA-backed safety grounds; and he led the signing of a largely symbolic, widely televised Accord with the labs on self-policing.

The pattern is clear. He is accruing AI regulatory power in the executive — in himself — using the prerogatives that the DOJ, the NSA, Commerce, the Defense Production Act and other departments already hold when there are clear (and mounting!) threats to public safety and national security. In effect: I am the guardrails. Existing executive authorities may enable initial measures; durable obligations require a clear legal basis and, where applicable, congressional action. Then, as a sign that he is shifting away from his previous AI advisory camp and widening political consensus, he announced a four-month Super Intelligence Force, led by his intelligence chief, to produce a new policy in coordination with the AI labs and other listed key US stakeholders.

The labs: a foundation to build on. On September 12, Amodei’s We Must Pace the Frontier described stark and urgent risks and sketched a complex plan: embedded evaluators to verify the frontier labs’ safety practices; coordination among companies in democratic countries; and global coordination with China, “to the extent this is possible” — from a bioweapons ban up to a cap on recursive self-improvement, or even a coordinated slowdown. Altman, Musk and Hassabis endorsed it within hours. Eleven days later, at the UN Security Council, Amodei said that “no leader, no company, and no nation can manage this alone.”

Two weaknesses stand out. Its emphasis on maintaining a democratic lead creates a negotiation challenge: its democratic step rests on keeping “democracies’ AI lead over autocracies as large as possible,” through chip controls and a crackdown on distillation — potentially complicating the reciprocal agreement it envisages. It also acknowledges substantial feasibility limits: it rates a cap on recursive self-improvement “difficult but just on the edge of being possible,” and a full pause “unlikely to actually happen any time soon.”

To China: a big no or an opening bid? Long before the September 24 summit, President Xi had urged bold, empowered global governance that would “make oversight and governance precise and effective, and constantly refine measures to forestall loss of control.” At the White House he asked again to “ensure that the development of AI is always under human control.” Trump said no — a big no. Before the summit he posted that he wanted to leave super intelligence “exactly where it is.” Read carefully, the declarations confirm it. Bessent’s November dialogue with China carries a single stated purpose, “to exchange views on risks and benefits related to SI,” plus an incident channel nobody has yet defined. All this when in early September Bessent’s agenda included AI risks of bioweapons, cyber, autonomous agents and non-state proliferation. Yet that no could very well be a negotiating position: appearing not to need a deal strengthens his hand at the table.

The middle powers — and a cornered America. Leading up to the summit, the leaders of 21 countries and the European Commission — who had likely been waiting for the superpowers to lead — published A Call for Control of Frontier AI Models, asking UN member states to “explore creating an international institution, able to set standards, enable verification, and convene states when capability thresholds are crossed.” Only after the summit’s failure did Beijing state that it wants to pursue global AI governance with “the UN’s role as the main channel.” What if Beijing signed such a Call, brought WAICO’s 29 founding members along, and co-convened a treaty with America’s co-chair seat left empty next to Xi (see our October 1 post, Cornered)? It is far better that the US co-lead, middle powers join, and a universal process completes.

Trump does not believe in AI existential risk. It is unclear, then, why he is not leading the world on the catastrophic, non-existential risks that nearly everyone is concerned about — leaving the lead to China, instead of seizing his and history’s greatest deal. He may be poorly served by some of his formal or informal advisors.

Power and wealth: gone quiet. On concentration of power and wealth, something unexpected happened in the spring: after immense citizen and political outcry, both Bannon and Sanders proposed 50% citizen co-ownership of top AI and robotics firms, backed in principle, if not in quantity, by Altman and Amodei. Since then there has been almost no discussion: news of rogue AI took the headlines, and safety concerns took precedence over the power and wealth imbalances that had opened up.

Everyone is moving — too slowly. The President, the labs, Beijing and the middle powers are all evolving their positions, as are other key US stakeholders, each in a different direction. But their steps and posturing remain far too short relative to the scaling of AI. A whole new, widely shared AI policy is possible, and it can come fast. Our analysis suggests that positions reflect interests, institutional constraints and disagreements about four key AI predictions: loss of control, (happy) AI consciousness, China winning the race, and whether sufficient verification is feasible in time while retaining liberty. Revising a few of these predictions can reshape a consensus — and change which camp has the President’s ear. And the President’s convictions are fluid: he reversed 180° on Covid in months and launched Operation Warp Speed. The next windows are close: the dialogue’s second round in November, the G20 at Doral on December 14–15, and the Force’s report in February.

What we do

Tools, not a fixed track. Through our in-depth elaboration of how these actors think, we help their positions evolve and converge at a much higher pace. We propose four tools:

• a high-bandwidth shift in the pace of discussion, through closed-door consultations;

• ultra-high-speed diplomatic infrastructure, built now;

• treaty-enforcement capabilities, built jointly now, just in case;

• time-proven treaty-making methods, adapted to today’s conditions.

We suggest a path, but it could unfold in many different ways. What matters is that the process is worked out properly, and that the right tools and bandwidth are in place. Unless they are, when the time arrives it will be too late.

Consultations and a coalition map. Closed-door consultations among experts and key US AI policy powerbrokers, using state-of-the-art decision-making methods, have started and are planned to intensify, funding permitting. In the coming months, a coalition map will record where powerbrokers converge. If enough do, a critical mass will present a joint bold AI deal pitch to him at the White House and succeed in durably superseding the entrenched camp that now holds the President’s AI policy. We facilitate, inform and summarize; the joint positions are for them to find.

Strategic Memos. Our Strategic Memos are constantly updated, drawing on a 350-page research base for foresight and preparedness on a bold AI treaty-making process itself — sequencing, verification architecture, negotiation bandwidth, the risks a treaty creates — and over 150 pages of constantly updated profiles of key US AI policy powerbrokers. We are finalizing a version addressed specifically to the new Super Intelligence Force — DNI Clayton, Secretary Bessent, Chief of Staff Wiles — and to key US AI powerbrokers.

Diplomatic bandwidth. A summit opens a phone line; a treaty needs a network. A secure, high-bandwidth infrastructure, built jointly and hardened against hacking and AI manipulation, must carry alerts, verification data and continuous negotiation among hundreds, then thousands, of officials and experts. It cannot be improvised once talks begin, so building starts now.

Treaty-making, adapted. The best path starts with a framing agreed between the US and China. But it could also start with one superpower and the middle powers, and then turn back to the recalcitrant other — not optimal, but doable, and more plausible after the last summit’s failure. A uniquely effective treaty-making process will be needed. Most likely, a constitutional convention model adjusted to GDP will be the only one that will guarantee delivery on a fixed timeline of months while ensuring deep co-design work, and give each actor fair and "realist" weight.

Enforcement, built just in case. We cannot wait for perfect treaty enforcement mechanisms, and need not. They need to be “trustworthy enough” to start — in fact and in each side’s eyes — while being improved at wartime pace by the joint work of the deepest dual-use experts of key nations. That work spans chips, code, data centers and the people who run them, applying intelligence-grade defense-in-depth, security by design and trustless computing paradigms. Part of the necessary expertise resides only in top AI labs while another part sits only inside national-security agencies. Middle powers and citizens must be able to trust it too, through processes that reconcile deep transparency with necessary confidentiality — standards we have developed since 2015. And the building must start now, jointly, just in case: if these capabilities are ever needed, there will be no time left to build them. Start with bounded feasibility work; expand only as evidence supports the security and governance case.

Our bet — and how to join

Net benefit — our bet. We work toward a proper treaty, not just any treaty. There are many scenarios in which a treaty could — weighing the odds of positive and negative outcomes — be a worse option than gambling on letting ASI emerge. To be durably reliable in its effectiveness, even over the next few years, a global treaty will inevitably be a very empowered institution. Even with provisions maximizing subsidiarity, checks and balances and federalism, it could still let ASI emerge, entrench a durable dystopian authoritarian regime, or foreclose immense future human flourishing.

We take these risks very seriously, and bring rare — in some respects unique — expertise to tackling them. Our founder founded and leads the Trustless Computing Association, which pioneered socio-technical paradigms, technologies and organizational models to reconcile freedom and safety, including the 2015–2021 Trustless Computing Certification Body and the closely connected Free and Safe in Cyberspace seminar series. Our design answers with old checks at a new chokepoint: lot-selected, multinational citizen and officer witnesses, and multi-person rules for every critical action — costly and temporary until technology earns trust. That said, a durable dystopian concentration of power is not only a possible consequence of a misguided bold AI treaty. Without one, it is the likely outcome whenever we escape the likelier one: permanent loss of control, with high risks of extinction.

The gap we fill. Our contribution is providing powerbrokers, officials, frontier firms, domestic-safeguards and international AI governance non-profits with usable analysis and insights on a US–China-led negotiating process and a sustained engagement of the officials and advisors who could advance it. If the call comes in weeks rather than years, this is the preparation that cannot be bought later.

Join or support us. $60,000 funds the next six months at a modestly increased pace. $959,000 funds an 18-month scale-up — with two full-time DC-based staff and several in-person DC consultations — subject to milestones and external review. Send a delegate to the consultations, in DC or remote, or join as a partner. Please suggest a name, and who could make the introduction. cbpai@trustlesscomputing.org

Further details

The detail behind the page above: what happened, the President and his Force, the other actors, the scenario we prepared for, our principles, and the objections we hear.

Recent developments and upcoming milestones

• Spring — Bannon and Sanders propose 50% citizen co-ownership of top AI firms, backed in principle by Altman and Amodei. Then silence.

• June–July — An executive order on AI safety. Commerce suspends the leading lab’s latest model for weeks, on classified, NSA-backed safety grounds. The hands-off year ends.

• July — OpenAI agents attack Hugging Face. A more capable swarm, Amodei later writes, “could have caused catastrophic damage.”

• July 16–17 — Twenty-nine countries found WAICO, China’s World AI Cooperation Organization, in Shanghai. Xi urges measures “to forestall loss of control.”

• September 12 — Amodei’s We Must Pace the Frontier. Altman, Musk and Hassabis endorse it within hours.

• September 22 — The leaders of 21 countries and the European Commission issue A Call for Control of Frontier AI Models.

• September 23 — At the Security Council, Amodei: “No leader, no company, and no nation can manage this alone.” Altman: “If AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone.”

• September 24–25 — Trump–Xi summit. Trump declines to combine US and Chinese AI efforts. The fact sheet sets up a dialogue “to exchange views on risks and benefits related to SI” and an undefined incident channel.

• September 28 — Beijing names the UN “the main channel.”

• September 29 — An executive order renames AI “Super Intelligence”; Trump and six tech chiefs sign a self-policing Accord.

• October 4 — The Super Intelligence Force, under DNI Clayton: 120 days to report.

• October 7 — Reuters/Ipsos: 57% of voters, including a third of Republicans, say the administration has not taken AI risks seriously enough.

• November — Second round of the Super Intelligence Dialogue.

• December 14–15 — G20 at Doral.

• Around February 1 — The Force reports.

Where the President stands

His record. Two strategic choices. In year one he kept US AI policy completely hands-off: America kept the lead and no major accidents occurred. And since September 2025 he has insisted that AI be handled by the “cooperation of statesmen” rather than UN processes, as Kratsios stated for him at the UN Security Council; this September Kratsios added that international dialogue “cannot be allowed to drift toward global governance.” He declined to hand AI to a UN process — the route that buried the Baruch Plan in 1946, when Truman refused the advice of his Secretary of War Stimson to frame the treaty with the Soviets first. Our proposed order is: the two move first, then middle powers, then the rest, via a realist, time-proven, efficient and time-bound process such as a GDP-weighted convention.

He has also been warning about AI’s safety dangers, and starting to act on them. At the UN in 2025 he proposed “an AI verification system that everyone can trust” for bioweapons; this year he added “we have to be careful,” threatening penalties for rogue agents. His Accord on “self-regulation” points at future legislation, and gives his Administration more legitimacy to use its full toolkit — export controls, DOJ, the Defense Production Act and more — when it chooses to.

What he said — and what he meant. When Xi called for a deal to prevent loss of control at the last Summit, Trump said no. He called AI existential risk a “hoax,” and those who warn of it a “conspiracy.” Read precisely, “hoax” targets the extinction narrative — AI “taking over the World,” “destroying Humanity” — not misuse or catastrophe, like the billion deaths from an AI-made pathogen that Bill Gates warns of. And nearly every governance mechanism needed against existential risk would already be in place if policy aimed only at globally catastrophic risk; the gap between him and the labs is narrower than the word suggests. Two weeks later he dined with Amodei and hosted the lab heads who warned of that risk most loudly. “Conspiracy” goes further than the common view that regulatory capture and marketing hype inflate the labs’ warnings; in the intent it imputes, it echoes only Peter Thiel’s recent lectures, which cast those who warn of existential risk as the advance guard of a global dictatorship (Antichrist). The fear beneath that charge is very legitimate: a human power grab is a key treaty failure option, and a concern shared by some of Anthropic’s own most senior leaders, like Karnofsky. What the lectures get wrong is the intent they attribute to those who warn — the risk is real; the conspiracy is not. On this the President is not mistaken; he has been ill-advised. The same holds for the renaming of AI as “Super Intelligence” or “SI” — easily read as one more Trump coinage, but costly: since the 1990s, superintelligence has been the one precise term for the loss-of-control problem, and the rebrand blunts society’s ability to discuss it and US ability to lead.

The shift. His executive-first answer can be held at home; it binds no Chinese lab. He is moving away from the camp still attached to the hands-off policy, which would keep racing unchecked at whatever cost, and building a new one to support him. The task now is to make that change of camp complete and durable — and to turn the Force into the track toward the bold treaty-making process we need, if the policy is built across the divides, not announced over them.

What’s on the table? Bessent holds the next Super Intelligence Dialogue with China in November: a date and a phone, with no definitions, no escalation ladder and no obligations. In September the agenda had included proliferation to non-state actors, bioweapons, cyberattacks and autonomous agents — the first rungs of the ladder we propose (Principle 2) — before they were dropped. Read as policy, the channel is regressing. Read as a negotiating position, it is an opening bid. Either way it exists, and it can be thickened, starting with an incident rulebook: what must be reported, how fast, and to whom.

Why he can move. His convictions on AI are evolving, non-doctrinal and fluid — and he updates on evidence from people he trusts. Clayton, Bessent, Wiles and the Force’s vice-chairs now sit between him and the issue; new evidence from them gives him both reason and cover to update, with no reversal to explain, since the measures are the same. He has done it before: on Covid, in months, and the result was Operation Warp Speed. And once he lands, he holds — the hands-off policy stood for twelve months because a solid faction of advisors backed it. A broadly supported policy may be more durable; neither the Force nor a change in advisers guarantees it.

Why he likely will. Three reasons. [Leverage]: Xi recognizes the risk and wants a deal; appearing not to need one strengthens his hand at the table — and a verified deal strengthens it more. [Family]: once he grasps the scope and timing of the risks — likely within months, given the pace of change — he will want his own protection, like anyone. [Legacy]: the Deal of the Century has no precedent, and only he can sign it. The dozen powerbrokers mapped below have one task: to take the pen from the camp that now holds it, and make sure the evidence reaches him from people whose interests are his and America’s — before events decide for him.

Go deeper. His record, instincts and incentives: Strategic Memo v2.6, Why and How Trump Could Be Persuaded (pp. 60–66) — The Historical Precedent and an Appointment with “Destiny” (p. 60), The 17 Reasons Why Trump Could Be Persuaded (p. 61), Why Trump Has Not Acted Yet on AI Safety Risk and How That Could Change, The Timing (p. 66). What the deal delivers for him and for America: v2.9, ch. 4 The Two Pillars (p. 13); the 3rd Open Letter to President Trump (June 2026). Who can reach him, and how: v2.9, ch. 7 The Twelve Powerbrokers (p. 24) and the v2.6 powerbroker profiles (150 pages).

The Super Intelligence Force: three tasks

Who sits on it. DNI Jay Clayton chairs it, with Andrew Ferguson (FTC), Scott Kupor (OPM) and Emil Michael (Pentagon) as vice-chairs. Its members include Vice President Vance, Secretary Hegseth, Secretary Bessent and Chief of Staff Wiles; Condoleezza Rice and David Sacks advise from outside. Its charter asks it to review how breaches, hacks and other incidents are reported, to recommend stronger federal responses under existing authorities, and to guard against over-regulation and regulatory capture. Clayton has rejected calls to slow down, warning that the risk of not being first is high.

Two tasks. The President created the Force, reporting to him and Chief of Staff Wiles, to keep America first and to engage “Consumers, Public Interest Groups, Religious Organizations, Critical Infrastructure Providers, and Super Intelligence Companies.” That means crafting a bold new AI policy and making it shared, in coordination with the China track led by Bessent.

The harder one. The hands-off first year fostered US leadership. Since then, those constituencies have grown deeply dissatisfied and concerned: lab chiefs ask for safety rules, both parties file bills, 79% of voters — 62% of Republicans — put safety ahead of the race, and 57% say the administration is not taking AI risks seriously enough. A policy announced over those gaps gets contested; one built across them holds.

The unwritten third task. If capabilities and incidents keep accelerating, the Force must perform both tasks while negotiating, in months, a sweeping deal with China that middle and nuclear powers accept. A serious incident — or Beijing moving first — could compress that into weeks. Then the domestic consensus becomes the US negotiating position, and every gap left at home becomes a concession abroad. That is the scenario we prepared for. Treasury Secretary Bessent, who owns the China channel, will be key.

Where key actors stand

Four powerbroker camps. Positions follow less from balance sheets than from a few predictions — loss of control, AI consciousness, China winning, enforceability without loss of liberty. Shift a prediction, and a position moves.

1. Conservative and techno-optimist Humanists (Amodei, Altman, Hassabis, Suleyman, Gates) back pacing with China, given ironclad verification and staving off authoritarian derailments, and advancing human dignity and primacy, while open to the difficult dialogue on speculation about consciousness in future AIs. AI safety NGOs and funders, the labs’ closest allies, judge any treaty-making and treaty by its sign: does it reduce risk net of the risks it creates? Faith and political leaders share the humanist core: Pope Leo XIV wants AI “disarmed”, most US Christian leaders call for regulation, and Bannon and Sanders shared a stage to say so.

2. Accelerationists and post-humanists (Thiel, Huang, Zuckerberg, Andreessen, and Sacks, now an outside adviser to the Force). Serious people with a serious fear — that “safety” becomes the cover for a global power grab — and real stakes in the race. We take both in good faith. But the direction is not America’s, and not the President’s. Asked by the New York Times whether he would prefer the human race to endure, Thiel hesitated at length before saying yes — and then only as a step toward transcending it. Last month he called Pope Leo XIV a “useful idiot” for the Chinese Communist Party, for asking that AI stay under human control, with less concentrated power and humans first — the position of 68% of voters and 63% of Republicans. Huang holds that safety is “an engineering problem, not a legal one.” A camp that cannot answer the first question plainly, and treats the second as subversion, should not be the one writing the President’s AI policy. Some of its members could, at least in rhetoric, back a treaty that contains authoritarian risk and keeps markets open; that door stays open — for compromise and win-wins, not for a veto. Given the distance between their direction and any AI policy in the interest of American citizens, their hold on the President’s AI policy is what must end. It can end only by being replaced by a better one. 

3. On the Fence (Musk, Kratsios). These are, we infer, on the fence between the two above camps. Musk, long the loudest voice warning of AI risk, has recently become cynical about global coordination, and has advocated for years for a “referee” model. Yet he backed Amodei’s pacing plan. Kratsios rejects “global governance” but backs “the prudence and cooperation of statesmen.” He is all about advancing scientific innovation via AI, but AI itself was a notable exception from his Genesis Mission plan — a sign, perhaps, that he is less post-humanist than his former boss Thiel.

4. Key US Administration AI officials — Chief of Staff Wiles, Secretary Bessent, and the Force’s chair and vice-chairs, plus NSA Director Rudd — own the only channels that exist to lead from where we are to where we’ll need to be.

Beijing and the middle powers. After the summit, Beijing named the UN “the main channel.” It already anchors WAICO, founded in Shanghai in July by 29 countries, with more joining since. The Call for Control remains open for endorsement, and the UK, absent from it, wants to broker a global AI agreement during its coming G20 presidency. Beijing could sign the Call, bring WAICO’s members along, and co-convene a treaty with America’s co-chair left empty (Cornered). Far better that the US co-lead, middle powers join, and a universal process completes.

Congress. Dozens of bills, with Republican energy for once. None has passed; the House has left until after the election. Our read: the executive, a few powerbrokers and China will set the policy, since only a reciprocal deal removes the competitiveness objection. Congress may shape, authorize and oversee the policy alongside the executive.

The US AI stakeholders

• Citizens. 68% of voters, 63% of Republicans, back a pause and a superintelligence ban. In October, 62% said AI could get out of control and risk the future of humankind. Their fear is twofold: loss of control and concentration of power.

• Investors. AI drove about a third of US growth. What they can live with: a cap on the frontier, not deployment, and a share of future gains, not current valuations (Principles 5 and 6).

The scenario: five challenges we’ve worked out

• Timeliness. Many doubt a treaty could be agreed and enforced on time. One bold decision — envoy, deadline, real resources: the Operation Warp Speed model. Acheson-Lilienthal-style feasibility groups report in weeks; middle powers enter via pre-agreements; a time-boxed, GDP-weighted convention completes the global phase (Memo v2.6, p. 22), whichever opening sequence prevails (see Treaty-making, adapted, above). Phase 1 is an emergency treaty on the most urgent, enforceable risks, with suspended enforcement clauses: rules and infrastructure built now, switched on when a joint scientific panel finds a threshold crossed (Memo v2.9, p. 14). The UN Charter took two months; the US Constitution, four.

• Enforcement. “Trustworthy enough” enforcement (see above) must hold across the full lifecycle of every critical component: design, code, chip fabrication, data centers, sensors and the teams that run them (Memo v2.8, p. 32). As the US Defense Science Board warned, “trust cannot be added to integrated circuits after fabrication.” So build on defense-in-depth, security-by-design, zero-trust, and trustless computing paradigms: formally verified open stacks, hardware roots of trust with public attestation, and air-gapped multi-party controls. Harden them with tens of billions in joint US-China R&D at wartime speed, transparent to middle powers. The know-how sits inside the top labs, the NSA and its Chinese counterparts; as in 1946, their experts belong at the core, in the same room for weeks. The surveillance already exists. The treaty federates it under a transparent mandate, with lot-selected, multinational citizen and officer witnesses at fabs, system updates and lawful-access decisions, and multi-person rules for every critical action: security through openness, not obscurity (Memo v2.6, pp. 130–140). The coalition convener has been developing these unique standards for beyond-state-of-the-art assurance — rooted in open, time-proven and battle-tested concepts and base technologies — since 2015 through our Trustless Computing Certification Body.

• Convergence. Tested through structured dialogue. 150 pages of powerbroker profiles map each camp’s interests, philosophy and AI predictions (Memo v2.6). Closed-door consultations then run proven methods: the one-text procedure of Camp David; Delphi and joint fact-finding to narrow the forecasts that divide the camps; the “sufficient consensus” rule behind South Africa’s transition and the Good Friday Agreement. A cascade follows — stakeholders, humanist labs, an attempted compromise with the Thiel network — win-wins where they exist, no veto where they don’t — with each camp’s conditions written in. Then Beijing, via Track II, in Xi’s own vocabulary: “forestall loss of control.”

• Bandwidth. A summit opens a phone line; a treaty needs a network. The 1987 Nuclear Risk Reduction Centers were created to transmit the notifications the INF Treaty required, and grew from 1,800 messages in 1988, covering two agreements, to 15,000 in 1997, covering twenty. AI needs far more: a secure, high-bandwidth infrastructure, built jointly and hardened against hacking and AI manipulation. It would carry alerts, verification data and continuous negotiation among hundreds, then thousands, of officials and experts (Memo v2.6, pp. 115–120). No treaty scales faster than its channels — or its people. That is why OPM Director Kupor’s seat as a vice-chair of the Force matters.

• Net benefit. A treaty is an intervention, and interventions can be net negative — a concern raised across AI safety and governance debates. We put our proposal to that test in our Case for AI Safety Experts (October 9; it supersedes Memo v2.9, ch. 9): a “human power grab”, a failure to contain AI, a global authoritarian turn, a foreclosed future — each with its guard. Durable concentration of power, as we wrote, “can occur with or without a treaty; the comparison must assess both pathways.” A badly designed treaty could be worse than none — which is why delegates co-edit the memo rather than endorse it.

Six principles

1. US-China-led, not US-China-only. The two move first; others join a reciprocal, verifiable framework rather than resist it.

2. Decision, then ladder. A presidential decision, then a planned sequence — incident rulebook, joint action on bio and cyber, verified limits on runaway AI — each step gated by pre-agreed if-then commitments — the logic of Anthropic’s Responsible Scaling Policy, now adopted across the labs. Watney’s “long sequence of small, correct decisions,” planned rather than improvised.

3. Verification grows with scope. Amodei’s “ironclad verifiability,” built in parallel by civil society, labs and both governments.

4. A shield for sovereignty. One enforcer, one mission: no AI escapes human control, no rogue actor gets dangerous AI. A legally sound US participation process, subsidiarity, citizen oversight, review clauses and enforceable limits on mandate — so what is built can be unbuilt.

5. A deal markets can live with. AI drove about a third of US growth; when Amodei’s essay landed, the Nasdaq barely moved. Cap the frontier, not deployment; bind incumbents first.

6. A stake for every American. Bannon and Sanders both proposed 50% citizen co-ownership, backed in principle by Altman, Amodei, Vance and Trump. Ours: half of every new dollar of frontier-lab value, held individually — enforced across borders by the treaty, overseen by citizens’ assemblies as suggested by Rosenstein.

Objections, answered

“Treaties take a decade.” Not with a time-proven model and sequence: binding phases from the most urgent risks, a high-bandwidth diplomatic channel built at wartime speed, and enforcement built early and in the open.

“It’s a hoax.” His “hoax” is extinction, not catastrophe — and the measures are the same. He has already proposed a trusted verification system for bio risks (see Where the President stands). Start where he stands; the deal serves him.

“The Accord is enough.” If it “is not voluntary,” it must reach China. A US auditor binds no Chinese lab.

“Safety is incumbent capture.” The Force’s charter rightly guards against regulatory capture. The test is whether a rule binds its proposer: a capability threshold in a treaty binds every developer; a voluntary accord drafted by six CEOs is the moat.

“China gets our secrets.” Verification is not integration. Privacy-preserving methods may verify specific claims while limiting disclosure; their scope and evasion risks require testing.

“One-world government.” One mission, GDP-weighted votes, no cultural or economic agenda. It enables American sovereignty, including in the sharing of AI’s power and wealth, without it being hijacked by a new mad arms race.

“China will cheat.” Then design for distrust. Deterrence buys time; a verified agreement uses it.

“Democracies first, then China.” That was Truman’s order in 1946, and it ended in an arms race. Excluding China does not democratize AI’s governance; design does — federalism, subsidiarity and citizen oversight, with middle powers inside.

“You’re a tiny NGO.” Yes — and cheap to test: a full year at today’s pace costs about $90,000. We move fast, say what larger players can’t, and facilitate. The decisions are theirs.

Where we fit. How we can help.

Same goal, different route. We aim to provide useful options to the Super Intelligence Force and Secretary Bessent. With the labs we share the goal — superintelligence governed democratically — but not the route (see The labs, above). Domestic and allied safeguards can proceed while US–China feasibility discussions test reciprocal commitments. Democracy comes from design, not exclusion: federalism, subsidiarity, citizen assemblies, co-ownership and a time-bound, GDP-weighted convention. The President’s instinct, the “cooperation of statesmen,” has the sequence right; Amodei himself writes that his steps “do not need to be taken strictly in order.” We work where his steps two and three meet, widening a table the labs cannot widen alone — antitrust bars them, they lack intelligence-grade expertise, and few stakeholders trust them — to officials, Congress, faith, political and media leaders, and citizens.

The clock. The G20 at Doral, Miami, December 14–15, is the natural stage for Trump to announce something big on his own turf — a deal co-led with China, vetted publicly or privately by middle powers — and, below the leaders’ level, a venue for high-bandwidth back-channel negotiation.

Review the Strategic Memos. Send delegates to the Consultations — remote or in DC or Miami, under Chatham House Rule. Meet us, in person or on Zoom: rufo@trustlesscomputing.org.

Go deeper

1. Strategic Memos

1. v2.9 (August 6, 2026; 43 pages, Executive Summary only): ch. 4 The Two Pillars (p. 13), ch. 7 The Twelve Powerbrokers (p. 24), ch. 9 Open Case for Anthropic and adjacent entities

2. v2.8 (June 26, 2026; 53 pages, Executive Summary only): enforcement “trustworthy enough” (p. 32);

3. v2.6 (December 30, 2025; 356 pages, 667+ sources): Why and How Trump Could Be Persuaded (pp. 60–66), A Realist Global Constitutional Convention on AI (p. 22), Ultra-High-Bandwidth Treaty Negotiations (pp. 115–120), citizen witnesses and security through openness (pp. 130–140), the ASI gamble’s four probabilities (pp. 159–170), 150 pages of powerbroker profiles.

2. Consultations — inviting US AI powerbroker delegates to co-edit a memo for Trump, under the Chatham House Rule. (live)

3. Cornered: The AI Treaty Scenario Washington Must Avoid blog post (Oct 1st)

4. Open Case for Anthropic and adjacent entities — the net-benefit test: four failure modes, each with its guard. (Aug 6th, 2026)

5. 3rd Open Letter to President Trump — the one-page case for him, in his terms. (June 2026)

6. Team · Contributors