Note · 2025
A Blueprint for Multinational Advanced AI Development
Note proposing a multinational cooperation architecture for the development of advanced AI systems, an alternative to competition between states and private labs.
A recent note: no retrospective is possible, and it would be dishonest to fake one. This entry therefore says three things only.
The problem. The development of the most advanced AI systems is concentrated in a small number of private companies under two jurisdictions. As of October 2025, the United States holds about 75% of global compute capacity, China 15%, the European Union 5%, that is, in H100 equivalents, just over a million chips on the US side, 231,000 on the Chinese side, and 79,000 across the rest of the G7. This asymmetry, compounded by a comparable concentration of talent, data and model ownership, leaves intermediate economies facing an equally bad choice: dependency, if they adopt foreign systems that a supplier can restrict, degrade or cut off, and whose design choices constrain everything built on top of them; or falling behind, if they abstain and let a gap widen in productivity, scientific discovery and military capability. Existing instruments of international cooperation (summit declarations, voluntary commitments, national safety institutes) create no interdependence capable of altering that dynamic.
The proposal. The note introduces the notion of a bridge power: a state with real AI capabilities (researchers, data, compute infrastructure) but not the scale to sustain a frontier effort alone. Its thesis is that these powers, pooled together, can produce competitive models. Three forms of pooling underpin it. Compute first: training is a massive fixed cost, independent of the number of users, which makes it the natural candidate for sharing: European capacity already committed or planned, from Jupiter to the five gigafactories expected by 2027, represents more than twenty billion euros. Talent next: of the hundred most-cited AI researchers worldwide, eighty-seven come from, or now work in, countries other than the United States and China. Data last, not for what is already public, but for what is expensive: expert annotation and cleaning. To this the note adds two methodological conditions: placing targeted research bets rather than chasing scale, and positioning on reliability, where industrial need is real and the quality of the legal framework is a structural advantage. The institutional form envisaged draws on CERN and Airbus: a semi-distributed structure built around a few sites, equitable sharing of costs and benefits, variable-geometry participation, tight governance.
What remains to be shown. Three reservations, which I raise all the more readily for having signed the text.
The first is that “bridge power” names a category of analysis, not a coalition. The note describes what these countries would gain by acting together; it does not say who convenes, who decides to join, or what happens at the first strategic disagreement. The step from economic case to political act is still entirely open.
The second concerns governance, and this is where my experience at the Campus Cyber makes me cautious. The note calls for an “agile” architecture, able to avoid the consensus-driven processes that have slowed other multinational projects. It is right to ask for this, but it does not say how to get it. Bringing together actors who remain competitors produces real commons, at a price rarely put into figures: the time spent maintaining consensus, the trade-offs avoided, the ambitions trimmed to preserve unanimity. At the scale of sovereign states sharing strategic assets, that cost does not fall; it rises. And the pace of the technological frontier will not wait for it.
The third is that the sovereignty being sought stops at the hardware layer. The partnership pools compute, but the chips themselves remain, for the most part, of American design. The note acknowledges this in a footnote. That is a residual dependency no pooling arrangement resolves, and it is a reminder not to confuse decision-making autonomy with technological independence.
What remains, to my mind, is the strongest argument, and one that requires no commitment to any particular scenario: option value. If capabilities plateau, the infrastructure funded serves national priorities and the cost of coordination stays modest. If they do not plateau, the window will have been seized. Few public decisions on technology present themselves in such a favourable form.