The term was popularised by the writer Moisés Naím, who defined the idea as bringing to the table 'the smallest possible number of countries needed to have the largest possible impact on solving a particular problem'. Instead of the 190-plus members of a universal negotiation, each with a potential veto, a minilateral effort convenes the dozen or so actors whose participation is genuinely decisive. The G7, the G20, the Financial Action Task Force, and the Proliferation Security Initiative are all examples of governance built this way.
Why AI is a strong candidate for minilateralism
Frontier AI has an unusual property that makes it especially suited to the minilateral approach: the capability is extraordinarily concentrated. Every leading frontier lab is located in a small number of jurisdictions — chiefly the United States and China, with meaningful roles for the United Kingdom, the European Union, and a handful of others. The specialised hardware that trains advanced systems is designed and manufactured within an even narrower set of countries. A rule agreed among that small group would cover the overwhelming majority of actual frontier development, without requiring the assent of the many states that have no frontier labs at all.
This concentration flips the usual objection to exclusive clubs. Normally, leaving most countries out undermines legitimacy and effectiveness. But when a problem is generated by a handful of actors, a small coalition of exactly those actors is not a workaround — it is the efficient design. Agreement among the states that host the labs and make the chips is where the leverage actually is.
The advantages of going small first
- Speed. A small group of aligned states can negotiate in months what a universal body takes decades to approach, and can act while the technology is still governable.
- Depth. Fewer parties allow for more ambitious, more specific commitments, because agreement does not require diluting the text to the lowest common denominator.
- No universal veto. A single reluctant state cannot block the whole effort, as it can in a consensus body like the Conference on Disarmament.
- A nucleus to expand from. A functioning core agreement becomes a standard others converge on and eventually join — the way many regimes grow outward from a small beginning.
The hard part: including China
Minilateralism for AI runs into one decisive complication that landmine or financial coalitions did not face. The concentration of frontier capability includes China, and a minilateral effort that gathers only Western democracies would exclude one of the two states that matter most. A club of the willing can set standards, as the G7 Hiroshima Process did, but it cannot deliver the core function of AI governance — a mutual halt to the most dangerous racing between the leading powers — if it leaves out a leading power.
This is the central design tension. A minilateral effort small enough to be fast and deep may be too narrow to include both AI superpowers; one broad enough to include China loses some of the cohesion that makes minilateralism work. The realistic answer is probably a layered approach: a like-minded coalition to set standards and build institutions quickly, running alongside a separate, narrower bilateral or trilateral track focused specifically on the US-China risk-reduction that only those parties can deliver.
You do not need the whole world to start governing AI. You need the handful of countries where the labs and the chips actually are. That is a small enough room to reach agreement in — if the right players are willing to sit in it.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
Where minilateralism fits
Minilateralism is best understood not as a rival to a universal treaty but as the fast first stage that makes one reachable. A small coalition can establish the standards, build the verification institutions, and demonstrate that governance is workable, creating a core that a broader agreement can later formalise and universalise — the pattern by which the Financial Action Task Force's small-group standards became near-global. For AI, where the capability is concentrated and the clock is short, starting with the few who matter is not a retreat from ambition. It is the most plausible route to acting in time. The design challenge is to make sure the 'few who matter' includes both of the states whose racing the world most needs to slow.
Common questions.
An approach to international cooperation that brings together the smallest number of states needed to have the largest impact on a problem, rather than seeking universal agreement. The term was popularised by writer Moisés Naím. Examples include the G7, the G20, the Financial Action Task Force, and the Proliferation Security Initiative — small groups that act without waiting for the full membership of the international system.
Because frontier AI capability is extraordinarily concentrated: nearly all leading labs sit in a handful of jurisdictions, and the specialised chips are designed and made in an even narrower set of countries. A rule agreed among that small group would cover the overwhelming majority of actual frontier development. When a problem is generated by a few actors, a coalition of exactly those actors is the efficient design, not a workaround.
Including China. A minilateral effort small and cohesive enough to be fast and ambitious may end up as a club of Western democracies that excludes one of the two states that matter most — and AI governance cannot deliver a mutual halt to dangerous racing if it leaves out a leading power. The likely answer is layered: a like-minded coalition for standards plus a separate US-China track for risk reduction.
No. It is best understood as the fast first stage that makes a broader agreement reachable. A small coalition can set standards, build verification institutions, and prove that governance works, creating a core that a wider treaty can later formalise and universalise — the way the Financial Action Task Force's small-group standards became near-global. For AI, starting with the few who matter is a route to acting in time, not an alternative to eventual universality.