China Has Founded the Institution That Could Divide the AI World

China did not unveil its most consequential artificial intelligence product in Shanghai this week.

On Thursday, representatives of 29 countries signed the agreement establishing the World Artificial Intelligence Cooperation Organisation, or WAICO. The agreement describes it as an independent intergovernmental organisation with its permanent headquarters in Shanghai.

That makes WAICO more than another conference, industrial association or declaration of good intentions. It gives China the beginnings of a permanent diplomatic machinery through which governments can negotiate standards, develop applications and co-ordinate their approach to artificial intelligence.

China calls it the world’s first intergovernmental organisation devoted specifically to AI. Its founding agreement says it will follow the principles of the United Nations Charter and promote the development of artificial intelligence that is beneficial, safe and equitable. António Guterres, the UN secretary-general, attended the signing ceremony, although his presence should not be interpreted as a UN endorsement of every Chinese proposal. China’s Foreign Ministry

The founding members include China, Russia, Brazil, South Africa, Indonesia, Pakistan, Kazakhstan and Malaysia, as well as states across Africa, Central Asia, South-East Asia and Latin America. Belarus, Serbia, Cuba and Venezuela also joined.

India did not attend. India has instead joined Pax Silica, the American led initiative intended to secure the minerals, energy, semiconductors and infrastructure upon which the AI economy depends.

Yet even the membership lists do not produce two tidy camps. Kazakhstan has joined both WAICO and Pax Silica. Other governments are likely to adopt the same strategy: take investment, technology and training from both sides while resisting demands for exclusive allegiance.

That is the first warning against exaggeration. WAICO is not an “AI NATO”, and the world has not divided into two sealed technological blocs.

But something important has begun.

Xi gives the organisation its doctrine

The day after the agreement was signed, Xi Jinping set out the political argument behind it.

China, he said, would provide 5,000 specialist AI training places to developing countries over the next five years. It would establish international AI application centres with ASEAN, the Arab League, the African Union, the Community of Latin American and Caribbean States, the Shanghai Cooperation Organisation and BRICS.

China would also deploy its AI-powered “Mazu” meteorological warning system in 30 countries and promote open-source development, co-operation over technical standards and international rules governing data, infrastructure and increasingly autonomous AI agents. Xi Jinping’s speech

Five thousand training places over five years Spread across the developing world, it is a modest undertaking. Nor has Beijing disclosed a substantial WAICO budget, an enforcement mechanism or binding commitments to transfer technology.

The application centres could prove considerably more important.

If they become places where governments test Chinese models, train civil servants, develop local-language systems and design applications for agriculture, education, medicine, weather forecasting and public administration, they could create lasting relationships between Chinese companies and foreign states.

The engineers who install a system . The officials trained to regulate it. The standards used to certify it matter. Once a government has built its data, procurement and public services around one technological ecosystem, changing to another becomes expensive.

That is how technical influence becomes institutional power.

The accompanying conference statement shows the breadth of Beijing’s ambition. It addresses open-source models, data ownership, energy consumption, employment, international trade, supply chains, safety standards, critical infrastructure and the limits that should be imposed on autonomous agents.

It calls for traceable agent behaviour, common crisis-response systems and international standards capable of being recognised across borders. It also says AI must remain under human control. Shanghai conference statement

But this distinction must be preserved: the conference statement is a political programme, not proof that WAICO already possesses authority over all those subjects. The institution exists. Its power does not yet.

The American system already exists

WAICO did not appear in a vacuum.

The United States launched Pax Silica in December and expanded it to 24 signatories in June. It brings together countries involved in critical minerals, semiconductor manufacturing, advanced computing, energy and other parts of the AI supply chain.

Washington’s strategy goes beyond Pax Silica. Under the American AI Exports Program, the US government is supporting complete technology packages for foreign markets. These may include American chips, servers, cloud services, data systems, models, cybersecurity, applications and technical standards.

The declared purpose is to extend American technological leadership and reduce global reliance on systems developed by American adversaries. The packages must comply with US export controls and security restrictions. White House executive order

The United States is offering a secured, high-performance system built around American technology and trusted partners. China is offering broader access, open-weight models, cheaper applications, domestic control and participation in a Shanghai-based multilateral organisation.

Those descriptions are partly competing sales pitches.

America also promises sovereign data, local companies, finance and technical training. China also speaks about security, controlled data flows and reliable supply chains. Neither system is purely open, and neither frees a participating country from dependence.

A government adopting the American stack may depend on US export licences, American cloud companies and continued access to advanced chips. A government adopting the Chinese stack may merely exchange that dependence for Chinese models, hardware, cloud infrastructure and systems integration.

The real choice is not between dependence and independence. It is between different forms of dependence and, where possible, a mixture of the two.

Why institutions may matter more than models

The most powerful AI model attracts the headlines. It does not necessarily determine the world’s technological order.

Most governments will never train a frontier model. The expense in chips, electricity, data and engineering expertise is too great. Their practical question is much simpler: who will help them use AI?

A country that needs an agricultural forecasting system, a national medical assistant or an automated public-service platform does not require the most powerful model on Earth. It requires one that is affordable, adaptable to local languages, deployable on available hardware and supported by engineers who can make it work.

China’s opportunity lies there.

Its models do not need to defeat every American system on every benchmark. They need to become good enough, cheap enough and accessible enough to be adopted across countries where the American frontier stack is unaffordable, restricted or politically uncomfortable.

Open weight models can be downloaded, adapted and operated inside national borders. That makes them attractive to governments concerned about data sovereignty or unwilling to place sensitive information in an American commercial cloud.

There are qualifications. “Open source” is often used loosely in AI. A company may release model weights without releasing its training data, complete source code or development process. A government may be able to operate a model locally without understanding how it was created.

Nevertheless, accessibility changes the economics of adoption. If Chinese models become the standard tools in universities, ministries and local technology companies across the developing world, the effects will accumulate.

Applications will be built around them. Engineers will learn them. Governments will regulate around them. Hardware will be purchased to run them. Chinese technical terminology and assumptions will enter national policy.

A model can be replaced. An ecosystem is harder to remove.

The struggle for the undecided world

The countries that matter most may not be those already firmly aligned with Washington or Beijing.

Europe has its own regulatory power and will not simply surrender its approach to either. India wants American technology but also strategic autonomy. The Gulf states will purchase the best systems available while financing competing companies. Brazil, Indonesia, South Africa and Kazakhstan will resist being reduced to members of somebody else’s camp.

These countries are not passive territory. They have minerals, energy, data, markets and political votes. They will bargain.

Competition may make AI cheaper and more widely available. It may force the United States to offer developing countries more than restrictions and lectures about security. It may force China to prove that its language of openness is supported by real technology transfer, local employment and meaningful participation.

Countries previously treated as markets may acquire bargaining power.

But a divided AI world also carries serious dangers.

Competing standards could make systems less interoperable. Research networks could fracture. Governments might be pressured to exclude foreign models or hardware. Security vulnerabilities discovered in one ecosystem might not be shared with the other. Rules governing autonomous agents, military applications and frontier-model safety could become instruments of geopolitical competition rather than common protections.

The deepest danger is that both sides begin exporting technology faster because neither wishes to lose a country to the other.

That would turn international AI development into an arms race for adoption. Safety requirements might then be treated as commercial disadvantages, while governments with weak institutions acquire systems they cannot adequately inspect or control.

The case against Beijing

Washington’s likely answer is straightforward.

It will argue that WAICO is Chinese industrial and geopolitical policy presented in the language of multilateralism. Training programmes and application centres could create markets for Chinese technology, advance standards favourable to Chinese companies and embed Beijing’s influence inside foreign ministries and regulators.

American officials will also question whether technological sovereignty is genuine when a country’s national systems depend upon Chinese suppliers. They will point to China’s domestic censorship and state control and ask whether its preferred governance principles will adequately protect privacy, free expression and independent scrutiny.

Those are not trivial objections.

But Beijing has an equally obvious reply. It can argue that the United States already uses its position in semiconductors and cloud computing to determine which countries and companies may receive advanced technology. “Trusted partners” can become another expression for politically approved customers. American sovereignty is offered within a system whose most valuable components remain subject to Washington’s permission.

WAICO is designed to exploit that contradiction.

China is presenting itself not as the country with the fastest chip, but as the country willing to help others acquire an AI capability of their own.

Whether it fulfils that promise is another matter.

Three possible futures

WAICO could develop in three very different directions.

It may become a diplomatic shell: an annual conference, a secretariat in Shanghai and a series of declarations without substantial funding or technological consequences.

It may become a useful development organisation, coordinating training, public-service applications, standards and technical assistance without requiring exclusive alignment with China.

Or it may gradually become the institutional centre of a Chinese-led technological sphere, with common procurement rules, recognised safety certifications, interoperable Chinese systems and preferential access to models, hardware and infrastructure.

There is not yet enough evidence to say which path it will take.

The tests will be practical. Will members ratify and implement the founding agreement? What budget will they provide? Who will control its secretariat? Will it fund real computing infrastructure? Will its standards enter national law and government contracts? Will application centres use several countries’ technologies or primarily Chinese systems?

Until those questions are answered, claims that a new AI order has already arrived are premature.

China spent the first stage of the AI contest constructing a technical ecosystem capable of operating despite American restrictions.

In Shanghai, it began the second stage: building an international constituency around that ecosystem.

The world has not yet split into two AI blocs. It has entered the struggle that may produce them.

The decisive contest will not be only over which country creates the most intelligent machine. It will be over who supplies the technology, who trains the governments and whose rules become ordinary across the countries that cannot build the technology themselves.

Shanghai is not the moment the AI world divided.

It is the moment that possibility acquired a permanent address.