China closed the World AI Conference in Shanghai by launching a 29-country AI alliance. A few weeks earlier, at the G7, Anthropic and Google DeepMind called for an American-led coalition. None of it was written with the buying governments in mind, which is not really surprising once you notice they were not in the room when it was drafted. (As a side note, this mess is also a full platter for international relations scholars – hegemonic transition, contested multilateralism, forum shopping, … – but that particular seminar can happen without me, ideally with Dawisson Belém Lopes organizing it.)
The word “sovereignty” itself is sold by people who pretend not to know what it means and bought by people who do not. Nvidia’s chief executive has been evangelizing “sovereign AI” for about two years, arguing that every country needs its own AI infrastructure, the way it once needed a telecoms network. He’s also the one selling it, which is worth remembering before any minister signs. “Sovereign AI,” if any analytical rigor were applied, should be called compute mercantilism.
But dependency theory aside, let’s get to where the rubber meets the road: what binds a government to a vendor is not the chip. It is what happens after the chip arrives, who trains the local engineers, whether the government’s own staff ever really learn to run the system, or whether that knowledge stays with the vendor’s support desk. Nineteenth-century railway loans worked on a similar logic, and the bill was often repaid long before anyone noticed that the track gauge, chosen in a foreign capital, was still dictating which rolling stock the country could buy decades later. By then, it had stopped looking like a foreign decision at all. It had just become so notoriously dumb that most of us heard a teacher talking about it sometime in high school.
None of this is inevitable, and the hopeful truth is that part of the remedy sits in procurement, not only in geopolitics. Knowledge transfer clauses, training quotas for government engineers, portability of models and data, code in escrow: dependencies of this type can be negotiated before the ink dries.
The catch is that no small state gets those terms bargaining on its own, which brings us to a more positive note: two blocs bidding for the same government is real leverage. Cold War financing rewarded governments that held out, but the advantage shrinks the moment one commits, and faster when each negotiates alone. The non-compute world has what both blocs want: data, minerals, energy, deployment markets. That leverage is lost one signature at a time. Pooled, it is how developing countries got affordable vaccines through Gavi, the Vaccine Alliance. So, governments of the non-compute world: unite, and don’t sign alone!