Geopolitics & Sovereign Positioning

The Trump-Xi summit has crystallised AI as a dual-use bargaining domain — simultaneously a source of competitive friction and a potential area for limited cooperation. According to Chatham House, both powers share an interest in AI safety discussions even as they compete on capability. Separately, Chatham House analysts argue that while nuclear arsenal size will not be formally discussed, the summit offers an opening to share threat perceptions about AI-induced nuclear escalation risks — a lower-threshold diplomatic ask that could produce substantive de-escalation value without requiring formal arms control commitments. These are diplomatic commentaries without binding force, but they reflect elite-level consensus that AI stability is now inseparable from nuclear stability.

On the commercial-strategic axis, Rest of World reports that chip exports and supply chain security are central agenda items. China enters these talks with structural leverage: its IC exports doubled year-on-year in April to $31 billion, driven by global AI hardware demand. US tech executives accompanying Trump signals Washington's awareness that decoupling has limits — American firms remain commercially enmeshed with the Chinese AI ecosystem even as the government pursues export controls.

Loongson Technology's confirmed shipment of over 1 million units of its 3A6000 desktop CPU — built on its proprietary LoongArch instruction set architecture with domestic IP cores — is an enacted, verifiable milestone rather than a projection. Reported by South China Morning Post, this represents commercial-scale adoption, not just laboratory demonstration. The strategic significance is not that Loongson is competing with leading-edge global CPUs — it is not — but that China has demonstrated the ability to sustain a domestic CPU ecosystem across the full stack: ISA, IP cores, and manufacturing.

For export control assessors, this is a second-order consequence worth tracking. US controls targeting advanced logic chips and EDA software were designed to impede China's ability to reach the technology frontier. What they have also done is accelerate investment in indigenous alternatives across the stack. At commercial volumes of 1 million units, Loongson is no longer a proof-of-concept; it is a supply chain. The question for Western strategists is whether controls are slowing frontier AI chip access — probably yes in the short term — while simultaneously reducing long-term dependency on Western IP — also yes, and at a faster pace than anticipated.

Three major Chinese tech platforms are simultaneously scaling AI infrastructure investment, converting early model capability into monetisable cloud services. Alibaba's CEO has signalled the company will overshoot its original capex targets, with AI products projected to generate over 50% of cloud revenue within a year and annualised AI recurring revenue targeting 30 billion yuan by year-end, per South China Morning Post. ByteDance's Volcano Engine is positioning its ArkClaw agent platform around cheaper inference and longer context windows — a cost-efficiency rather than capability-frontier strategy, per SCMP. Tencent's CEO acknowledged the company had been slow to adapt but signalled renewed strategic focus.

The geopolitical read here is not primarily about individual company valuations. It is that China's AI commercialisation is advancing on a separate stack — different models, different cloud infrastructure, different agent frameworks — which progressively reduces the interoperability and leverage that Western platform dominance once implied. As Chinese cloud AI revenues scale, the Belt and Road digital infrastructure layer gains AI services capability that can be exported to markets where US platforms face restrictions or where cost sensitivity favours Chinese offerings.

A Foreign Policy analysis identifies a foundational governance failure: states cannot agree on what AI actually is, creating definitional incoherence that makes binding international frameworks structurally impossible in the near term. This is not merely academic — export control regimes, treaty applicability, and liability frameworks all depend on jurisdictional definitions. When the EU, US, China, and multilateral bodies each operate with different taxonomies of AI capability and risk, mutual recognition agreements and compliance interoperability collapse at the definitional layer before they reach substantive rules.

For foreign policy professionals, this fragmentation is itself a geopolitical condition to manage rather than a problem awaiting a technical fix. The definitional vacuum benefits actors who prefer regulatory ambiguity — primarily those racing to deploy capability ahead of governance, and those seeking to avoid accountability for dual-use applications. It disadvantages countries building domestic governance frameworks in good faith, since those frameworks cannot be harmonised internationally without a common definitional baseline.