The U.S. decision last month to ban international access to Anthropic’s latest model exposed a growing gap between the pace of the technology and policy’s ability to keep up with it. Then, later in June, that G-7 leaders devoted a significant slice of their summit to it shows they see the stakes. But it also revealed the absence of an actionable agenda.

On this issue as on others, geopolitics curtails what is possible. The imperative to stay ahead constrains what countries can do domestically and what they can do together. But such steps as the Anthropic ban, which in its bluntness as an export control made no distinction between America’s allies and others, risks an inadvertent consequence: a technological fragmentation that weakens the West’s collective scale, which is critical to its strategic advantage.

A sovereign stack in Europe fully separate from the United States is neither achievable nor desirable for either Europe or the United States. Frontier capability at American scale is beyond Europe’s reach on any relevant timeline, and European capital is better deployed within the stack in the important areas of application and orchestration, where it can secure agency and amplify its strategic advantage. The United States, meanwhile, should have no strategic interest in pushing its allies toward such self-sufficiency. But the ban, which undermines allied certainty of consistent access to U.S. technology, risks exactly that. Building the international architecture to prevent this outcome is not just urgent but strategically essential.

What is needed is a pre-agreed playbook and pre-prepared toolkit that can be deployed deliberately without collateral damage to ensure safe release and stable access. AI’s open-ended nature, and open source especially, makes this genuinely complex. As a collective-action challenge rooted in national security and caught up in geopolitical competition, it is for states to determine a predictable basis for safe release. The good news is that the prescription for security—and for mitigating the risks the latest models have exposed—is clearer than the current confusion suggests.

The United Kingdom’s Bletchley Park summit in 2023, which brought together the United States, China, and nearly 30 other countries, established the basic framework for an international architecture on AI safety: global principles covering the most serious frontier risks, including bio- and cybersecurity; and collaboration between allies and companies on testing and trusted information sharing, backed by sufficient state capacity. In different ways, this is reflected in the recent proposals from the leaders of the frontier labs themselves. Had the framework not been rolled back last year in an executive order, the Trump administration would have had more than just a blunt ban to deploy when frontier models demonstrated the very risks that framework foresaw.

But while security challenges are inherently upstream, economic impacts are downstream, diffuse, and deeply political. That makes the economics harder—and arguably more consequential. The question for most people is not safety in the abstract but whether AI augments their work or automates it away. The politics of that are local. But because most countries are, to different degrees, takers rather than makers of the technology, the drivers that determine how AI develops and deploys into the economy—on everything from compute access to capital—are global. AI security governance has a framework, however incomplete. Economic governance does not.

There will be no Bretton Woods moment for AI. Governance of emerging technologies will evolve messily and incrementally. There are, however, parallels and precedents. The international nuclear regime, and the International Atomic Energy Agency in particular, is often cited. The parallels only go so far, but they show cooperation on shared existential risk is possible even at the height of strategic rivalry. The Financial Stability Board (FSB)—which began as a forum and has grown into something with genuine influence—may be the most underappreciated model. Its lesson is that durable architecture is built through successive layers of cooperation, not a single act of design. Waiting for that clarity will mean missing the moment. Instead, governments should start with practical steps, accepting that the agenda will evolve without certainty.

The first of those practical steps is establishing a shared understanding of what AI’s future means for policy. The International AI Safety Report that came out of the Bletchley summit provides a sound basis for assessing the development of the technology. That needs to be complemented with better insight into its deployment. That shouldn’t come from a single source (there is no monopoly on wisdom), but there needs to be a mechanism—this could be the Organization for Economic Cooperation and Development—for identifying which international policy prescriptions apply across scenarios and which are contingent on different pathways for the future.

The second step is building the apparatus for coordination, starting where interests converge and precedent exists. The latest frontier models have exposed cyber-related financial stability risks. But there is a much broader range of macro spillover risks—asset mispricing, productivity shocks, market concentration—that are potentially systemic and fit the macroprudential and crisis response toolkit developed since the financial crisis. This would be a practical place to begin. The FSB is ready-made for it.

The third is confronting allied interoperability—and its prerequisite. If fragmentation is to be avoided, the priority is putting in place guardrails that ensure certainty and stability of access to hardware and software. As the U.S. administration seeks to agree with the frontier labs on a more sustainable mechanism for the safe release of their models, this allied element should not be overlooked. Beyond that is the permanently unfinished agenda of ensuring that, as domestic AI regimes evolve, they do not diverge in ways that create barriers between allied markets, and even more consequentially, as AI diffuses through the economy, ensuring that rules written for a pre-AI world, on everything from cross-border services provisions to conformity standards, are fit for the AI one.

The fourth is combining this with strategic support for the deployment of AI globally into emerging markets and developing countries. Affordable access will be critical to countries’ prosperity and the contours of geopolitical competition. The United States and its allies need to learn lessons from technologies like 5G in this respect. The next generation of development policy should have this at its core, and it can only be achieved through collaboration between governments and the companies developing and deploying the technology.

The fifth goes to the heart of international politics: how countries share in the returns from AI as the gains inevitably concentrate. Ideas such as public-equity stakes in major AI companies on listing show that the domestic debate within the United States is beginning. Beyond the United States’ borders, the question is how countries that help create these companies’ value, not least through their data and customers, let alone any innovation generated there, can share in the returns. These revenues will be critical to managing the disruption of the transition. However politically difficult, the intellectual work inside and outside of governments on answers needs to start now, and elected leaders need to start the debate.

None of these steps can develop the necessary momentum without the machinery to carry them. What is needed first is a technical layer that can operate below politics and sustain cooperation, along the lines of the International Civil Aviation Organization, the International Maritime Organization, the International Telecommunication Union, and, to some extent, the World Trade Organization. On top of that must come a political architecture that will necessarily be stratified. Global norms will require the U.N.’s legitimacy despite its limitations.

Any regime for existential risk will have to emerge from dialogue between the U.S. and China. Allied collaboration will require something more focused. G-7 plus partners is the right starting point, especially as a forum for the United States and others to address the dynamics between them. But it will need to evolve into something more permanent and systematic—not a grand treaty, but something capable of sustaining action over time. That ultimately will require initiative from the United States, but a set of others could incubate. The United Kingdom has the presidency of the G-20 next year; it should build on what it started at Bletchley Park and make this its top priority.

The ban on Anthropic’s latest model may be lifting, but the gap it exposed will not close by itself. A dysfunctional default, with undesirable consequences for geopolitical competition and competitive markets, as well as security and safety, is already emerging to fill it. Forging the architecture to replace that is not a task for the future. It is for now.