Dialogue can reduce the risk of miscalculation, but Washington and Beijing are building rival AI ecosystems that will increasingly force other countries to make practical choices.
President Donald Trump recently distilled the strategic pressure shaping the artificial-intelligence contest into four words: “Whoever wins AI, wins.” His dismissal of warnings about potentially catastrophic AI risks exposes the central difficulty confronting international efforts to govern the technology: neither Washington nor Beijing wants safety measures that might allow the other to gain a decisive advantage. Against that background, the first official US-China artificial-intelligence safety dialogue under Trump may be diplomatically important, but it should not be mistaken for the beginning of technological reconciliation. The talks are intended to address risks from frontier systems, including AI-enabled cyberattacks. They are separate from the already scheduled Trump-Xi Jinping meeting on 24 September, at which AI is expected to be discussed alongside broader bilateral issues. The more consequential development remains outside the meeting room: the two powers are organising increasingly separate technology systems.
This is because Washington no longer treats AI as software alone. The contest now spans the entire physical and digital stack: data, telecommunications networks, data centres, advanced chips, manufacturing equipment, energy and critical minerals. The US-led Pax Silica initiative explicitly links AI leadership to trusted supply chains across connectivity, compute, semiconductors, advanced manufacturing and mineral processing. Its companion Joint Statement on AI Opportunity has attracted close to three dozen participating economies.
China has taken the initiative in establishing another institutional centre, but WAICO should not simply be characterised as a Chinese bloc. In July, 29 founding members signed the agreement establishing the Shanghai-based World Artificial Intelligence Cooperation Organization, or WAICO, as an intergovernmental body devoted to beneficial, safe and fair AI development. Its membership remains open to other countries, giving it the potential to develop into a genuinely global organisation. The presence of the United Nations Secretary-General in Shanghai when WAICO was established reinforced that wider multilateral ambition. Conceivably, it could eventually follow earlier precedents and become associated with the United Nations system. The emerging initiatives therefore do not yet constitute two sealed blocs, although their future development could still produce increasingly distinct technological and governance ecosystems.
The New Delhi Declaration adopted at the BRICS Summit adds an important qualification to this emerging competition. BRICS leaders committed themselves to implementing the BRICS Leaders’ Statement on the Global Governance of Artificial Intelligence and called for wider access to AI resources, trustworthy technologies, greater energy efficiency and stronger international cooperation. Their emphasis on inclusion and the needs of developing countries indicates that the Global South does not necessarily accept a choice between two closed technological camps. It seeks meaningful access to AI capabilities and a voice in shaping the rules governing them.
The harder question is whether countries can operate American AI models over Chinese-built networks and remain trusted by Washington. US concern is rooted partly in China’s legal order. Article 7 of the 2017 National Intelligence Law requires organisations and citizens to support, assist and cooperate with state intelligence work. China argues that its laws also protect legitimate rights, but the obligation has reinforced Western fears that Chinese companies cannot reliably refuse state demands for information.
That concern turns interoperability into a security issue. If Chinese-origin infrastructure is judged capable of exposing data, replacing it with American or politically neutral alternatives becomes the logical policy response, notwithstanding the cost. This is not simply a contest between brands. It is a contest over which state, legal system and supply chain can be trusted with technological dependence.
Export controls reveal the same hierarchy. The US Commerce Department continues to license advanced semiconductor exports to China only under security conditions, including customer screening and independent testing. Even where policy has become more permissive, the stated purpose remains protection of national security and the American technology ecosystem. Dialogue may alter the calibration of controls, but it is unlikely to displace their strategic logic.
Indeed, Reuters reported in August that US officials were preparing to tell partners that they could not ‘have it both ways’ in the AI competition with China. That independently confirms the emerging distinction between diplomatic flexibility and practical alignment. A government may endorse both American and Chinese declarations, but choices over networks, chips, cloud services, standards and mineral supply chains are becoming more exclusive. Security is likely to trump the commercial advantages of mixed systems.
What, then, can the bilateral dialogue achieve? A useful, but limited purpose. It can clarify red lines, establish communication channels and develop narrow arrangements for reporting or containing AI-enabled cyber incidents. It may become a scene-setter for more technical negotiations. What it is unlikely to produce is a comprehensive governance settlement between powers that define the same infrastructure as opportunity on one side and vulnerability on the other.
Nor can governments manage these risks alone. Frontier laboratories and chipmakers hold much of the technical knowledge needed to identify model vulnerabilities, trace cyber misuse and implement safeguards. Reuters reports that one proposal under consideration would ask American and Chinese AI laboratories to share information and help intercept AI-driven cyberattacks. The diplomacy may remain formally government-to-government, but any workable safety arrangement will depend heavily on private-sector execution.
This is why the multilateral route matters. The United Nations’ Global Dialogue on AI Governance, established by the General Assembly, held its first session in Geneva in July with delegations from 163 countries and participation from industry, academia and civil society. It gives every state, including those outside the two principal technological centres, a place in discussions on safety, human oversight, interoperability and equitable access. The dialogue does not itself enforce global rules: it is not a treaty body, international inspectorate or global technology regulator. Its value will depend on whether participating states can move beyond broad ethical declarations and develop practical, reciprocal and independently verifiable assurance standards. If they can, the UN offers the most credible route towards minimum global rules that are neither dictated by two competing superpowers nor fragmented between rival technological blocs.
For middle powers and developing countries, the policy challenge is immediate. They should not assume that today’s ability to participate in both camps will survive tomorrow’s security demands. They need to map dependencies across the full AI stack, assess the financial cost of substitution and use multilateral forums to resist rules written exclusively by either superpower. The US-China dialogue may help keep competition manageable. It will not stop the underlying separation, and may ultimately make the coming choices clearer.
Daryl Swanepoel is the Chief Executive Officer of the Inclusive Society Institute and a Research Fellow at the School of Public Leadership, Stellenbosch University. He is a leading voice on global governance reform. With a career spanning politics, business, and academia, he has served as a parliamentarian, strategist, and educator. His current focus includes work on reimagining the UN system, challenging entrenched power dynamics, and driving inclusive solutions.