WAICO’s stated purpose is to make AI a global public good, with open access and developing-country participation at the centre of the pitch.
Washington, D.C.
In July, 29 governments gathered in Shanghai to sign the charter of the World Artificial Intelligence Cooperation Organization (WAICO), a body the Chinese Communist Party (CCP) proposed a year earlier and will host permanently. Eight more governments signed within weeks, bringing the membership to 37. Its stated purpose is to make artificial intelligence a global public good, with open access and developing-country participation at the centre of the pitch.
It was then reported that China’s Ministry of Commerce was consulting Alibaba, ByteDance and Zhipu regarding export controls that would bar foreign users from downloading the weights of China’s most advanced models, although overseas customers would still be able to use the models as online services. The sponsor of a body whose legitimacy rests on open access was already weighing the power to close it.
This past summer, Chinese open models had overtaken their proprietary American competitors in global adoption. Chinese open-weight models, among them Alibaba’s Qwen, DeepSeek and Moonshot’s Kimi, became the default for developers across much of the world for the simplest of reasons: they were good and they were free. Weights can be downloaded, run on one’s own servers and modified without asking anyone’s permission. The leading American models, by contrast, are available only as a service. China did not win the “weight” race through diplomacy; it won on price.
The proposed controls would not reverse that. Weights do not behave like chips: a restriction cannot claw back what has already been downloaded, and models in the wild stay in the wild. What the controls would change is the terms on which the next model arrives. Organizations that became standardized on free Chinese weights assumed the successor would continue on the same basis. Under the proposal, it would come as a service hosted in one jurisdiction, subject to that jurisdiction’s filings, security reviews and licensing decisions. Those organizations did not acquire a free asset. They acquired a dependency and recorded it as a saving.
WAICO is where the rules governing that dependency will increasingly be discussed. Formally, its council of member states is the highest authority. In practice, however, bodies of this kind are governed less by votes than by whoever drafts the papers, staffs the secretariat and sets the agenda, and a permanent seat in Shanghai goes a long way toward settling that. The secretariat, to be led by a secretary-general and two deputies, has yet to be filled; who fills it will tell members more about the organization than its charter does.
This is not a Chinese innovation. The World Bank presidency has never gone to a non-American, nor the Asian Development Bank’s to a non-Japanese. What distinguishes WAICO is its price. Beijing’s headline contribution was 5,000 training places for developing countries over five years. For standing as a legitimate author of international AI norms, that is an extraordinary return on a rounding error.
The United States, Britain, European Union, Japan and South Korea are all absent. My presumption is that they were not invited. But the case for staying out has a familiar ring. In 2015, Washington declined to join the Asian Infrastructure Investment Bank, reasoning that membership would legitimize a Chinese-led institution bent on challenging the Western development bank model. It was legitimized anyway, and America’s absence bought no vote, leverage, or visibility into what projects the Bank supported.
Three conclusions follow. Governments and firms running Chinese open-weight models should treat provenance as jurisdictional exposure rather than a procurement detail, and secure fallback options, inspection rights and migration paths while switching is still cheap, not after an export catalogue has been amended. Anyone who has watched a borrower try to renegotiate a loan covenant after disbursement knows which side of that timeline to be on.
Middle powers attracted to sovereign AI should be candid about what is achievable. The realistic goal is not a national frontier laboratory but the capacity to run more than one stack, so that no single supplier’s policy change becomes a national emergency.
And Washington should recognize that export controls govern the frontier, not an installed base acquired at the bottom of the market. The only answer to a diffusion won on price is a competitive open alternative, together with a seat at the tables where the terms are being written.
Whether Beijing ultimately closes the tap matters less than who gets to decide. The only body now claiming global authority over AI has one host government, one set of priorities at its centre, and no member with the standing or the inclination to object. They were selected to become members for that reason. However, useful safeguards require that people who disagree to be in the same room. At present they are outside it, drafting communiqués.
* Daniel Wagner is Managing Director of Multilateral Accountability Associates and author of The AI Insurgency, forthcoming in October 2026 from De Gruyter Brill.
