Xi calls for AI not controlled by a single country
Xi Jinping has called for international cooperation to prevent artificial intelligence from being dominated by a single country. His message places the race for models, chips and rules at the heart of the technological rivalry between China and the United States.
Chinese President Xi Jinping has called in Shanghai for artificial intelligence not to be dominated by a single country and urged international cooperation to develop it. Delivered at one of China’s major technology gatherings, the message places AI governance in the context of geopolitical balance.
His remarks come as the United States and China compete over the building blocks of advanced AI: computing power, semiconductors, data, researchers and companies capable of training large language models. This is not merely a trade dispute. Whoever controls that infrastructure will wield considerable influence over which systems are deployed and under what rules.
From the promise of openness to the battle over chips
Xi’s call echoes a position Beijing has defended in international forums: AI should benefit all countries, including those without major data centres or a domestic semiconductor industry.
But that appeal for cooperation sits alongside increasingly fierce technological competition. Washington has imposed export controls on advanced chips and the tools needed to manufacture them, aiming to restrict China’s access to the computing capacity required to train leading-edge models. China, meanwhile, has stepped up support for domestic processor makers, cloud companies and model developers.
The result is a paradox. Governments talk about shared standards and reducing global risks while building supply chains and technology ecosystems that are less dependent on their rival. AI has become an arena where cooperation and competition are advancing simultaneously.
The battle is also over the rules
Leading in AI does not simply mean having the most capable model. It also means taking part in decisions on safety, privacy, intellectual property, military use and access to digital infrastructure.
China already has specific rules for certain generative AI services. Providers must meet security and content obligations for services available to the public, and the rules require content to comply with requirements set by Chinese authorities. The European Union, meanwhile, has adopted its AI Act with a risk-based approach. So far, the United States has opted for a more fragmented mix of federal and state measures and corporate commitments.
These differences matter because national rules can ultimately have an impact beyond their borders. A company seeking to operate in several markets will need to adapt its products, content filters, auditing systems and data policies to different regulatory frameworks.
What it means for countries and companies
The push for less concentrated AI responds to a real concern: training frontier models requires very large investments and access to vast quantities of specialised chips. That leaves much of the world as a consumer of technology designed in the United States or China, rather than a participant in its development.
For countries with fewer resources, the promised cooperation will only have value if it translates into effective access to computing power, technical training, models that work in local languages and the ability to decide how their populations’ data is used. Publishing a model or offering a cloud platform alone does not resolve that dependency.
For European and Spanish companies, the split between technology blocs could raise the cost of accessing infrastructure and force them to choose providers on criteria that are no longer purely technical or economic. The origin of chips, where data is hosted and which jurisdiction regulates a model will become increasingly important business decisions.
The international cooperation Beijing is calling for will therefore have to be measured against the concrete interests of the major players. The next challenge will not be declaring that AI should be global, but agreeing on who can build it, audit it and benefit from it.
Turning a governance speech into checkable questions
The full text of Xi’s speech, delivered on July 17, 2026, proposes openness, coordinated development, inclusion and a central role for the United Nations. It is an official position, not a multilateral agreement. To know whether it changes anything, look for a mechanism: participants, decision body, funding, schedule, membership rules and a process for resolving conflicts.
“Global” can describe at least three things. It may mean that many countries use the technology, participate in writing its rules, or share its benefits and costs. A service available over the internet meets the first condition without meeting the others. Measuring inclusion requires observing who has a vote, computing capacity, their own data, personnel and the ability to audit.
Four layers that are often blended
Infrastructure determines who can train and serve models. Technical standards determine whether systems and records interoperate. Regulation creates duties within a jurisdiction. Diplomacy tries to coordinate risks that cross borders. A country may cooperate on a standard while competing for chips; it may release weights while restricting data; it may support common rules and apply distinct domestic controls. Those are layers with different incentives, not an automatic contradiction.
The European AI Act demonstrates the distance between a declaration and a rule: it defines categories, responsible actors, documentation and application dates. That does not make it a universal model, but it makes obligations inspectable. A global proposal gains comparable weight when it says who must do what, to whom they answer and which evidence they publish.
For countries with less infrastructure, “access” needs a denominator. Ask how many compute hours, at what price, for which institutions, in which languages and with which rights over data and outputs. A temporary grant or subsidised API is not durable capacity if the provider disappears or changes its terms.
Correction mechanisms matter as much as access. A framework needs routes for reporting incidents, reviewing cross-border harm, challenging an assessment and amending a rule as evidence changes. It must also decide which information can be published without exposing sensitive systems. If there is only a recurring meeting and a consensus statement, cooperation may support dialogue but still offers no concrete remedy to a person or country. The hardest governance test arrives when two participants disagree and nevertheless accept a common procedure.
Concentration can be measured across several layers: chip manufacturing, clouds, models, distribution and standards. Opening one layer does not automatically repair the others. A downloadable model may depend on concentrated hardware; an accessible cloud may impose a closed interface. Stating which layer a proposal intends to democratise prevents a local improvement from being sold as a complete redistribution of power.
A matrix for reading any major power
Readers can compare four columns: speech, instrument, resources and outcome. The first records the promise; the second records a treaty, law, standard or programme; the third records money, compute and staff; the fourth records access, incidents, audits and beneficiaries. Applying the same matrix to China, the United States, the European Union or a company avoids accepting or rejecting a proposal because of the speaker’s identity.
The transferable skill is to distinguish declared universality from effective participation. Global governance is not demonstrated by using those words, but by distributing voice, verification capacity and correction mechanisms among the people who will bear its consequences.
This article was produced with artificial intelligence under human editorial oversight.
