Geopolitics & Sovereign Positioning

The launch of DeepSeek's V4, optimised specifically for Huawei's Ascend chips and accompanying software stack, represents the most concrete evidence yet that China's AI ecosystem is not merely diverging from US technology but building a parallel, self-sufficient architecture. As South China Morning Post reports, domestic chipmakers including Cambricon, Biren, and Moore Threads have rushed to validate V4 deployment on local hardware, creating a virtuous cycle in which frontier model developers and chip manufacturers co-evolve without reference to Nvidia or US software frameworks. This is not incidental — it is the deliberate materialisation of the Communist Party Politburo's longstanding self-reliance directive.

The infrastructure underpinning this is scaling rapidly. South China Morning Post documents an arms race among Chinese cities and technology companies to deploy 10,000-card compute clusters using domestic AI accelerators — facilities that enable frontier-scale model training without access to restricted US hardware. Huawei and Alibaba are among those operating or building such clusters. The strategic implication is that US export controls, while they have imposed real costs and delays, have also catalysed a domestic supply chain that is now approaching the scale required to train leading models independently. The controls may be succeeding in denying access to the absolute frontier, but they are simultaneously accelerating the conditions for that frontier to be reached domestically.

Beijing's decision to block a $2 billion deal involving AI agent startup Manus and Singapore-based investors — reported by Foreign Policy — marks a qualitative shift in Chinese AI governance. Previously, concern focused on foreign technology entering China; now Beijing is actively preventing Chinese AI assets and intellectual property from being internationalised through capital structures that could give foreign or US-adjacent entities leverage over Chinese-origin AI. Singapore, long positioned as a neutral conduit for Chinese tech capital seeking global access, is directly implicated as a jurisdiction Beijing no longer trusts as a clean intermediary.

The downstream effects are already visible. MiroMind, backed by tech billionaire Chen Tianqiao, has suspended services across mainland China, Hong Kong, and Macau, citing the post-Manus environment as too hostile for globally ambitious Chinese AI firms operating domestically, per South China Morning Post. The pattern suggests Beijing is drawing a sharper line between AI firms it considers strategic national assets — which will be kept onshore and under party discipline — and those with ambiguous international capital structures, which face regulatory pressure to choose. This bifurcation of the Chinese AI sector into domestic-controlled and internationally-facing entities has significant implications for how foreign governments and investors should assess Chinese AI partnerships.

First-quarter 2026 earnings filings reveal that Chinese chipmakers are deploying capital at R&D intensity ratios that dwarf US peers: Moore Threads at 50% of revenue, MetaX at 45%, compared to AMD and Intel operating at roughly 20-25%. South China Morning Post frames this as a state-backed attrition play — these firms are not optimising for near-term profitability but for closing the technology gap over a multi-year horizon, with implicit or explicit state backing absorbing losses. This is structurally different from how US chip companies, which are accountable to public markets, can operate.

The strategic calculus is clear: if Chinese chipmakers can reach 70-80% of Nvidia H100-class performance at scale, the marginal value of continued US export controls declines sharply. The current controls are most effective when the capability gap is large; they become less effective as domestic alternatives approach functional sufficiency for the majority of AI workloads. Senior foreign policy professionals should treat the R&D intensity ratio as a leading indicator of how quickly that sufficiency threshold will be reached.

An Atlantic Council issue brief positions the UAE as the leading example of a new category of AI power: states that are neither US-aligned consumers nor Chinese-aligned adopters, but active builders of sovereign AI capacity using a five-pillar framework encompassing compute infrastructure, talent, data, regulatory positioning, and diplomatic leverage. Atlantic Council identifies the UAE's Falcon model series, the TII research institute, and sovereign data centre investments as the core of this strategy. The UAE's approach is notable for deliberately maintaining technology partnerships with both US firms (Microsoft, Google) and, controversially, Chinese entities — a neutrality posture that maximises leverage.

The UAE model is instructive for how mid-sized, capital-rich states can exercise AI sovereignty without the scale of the US or China. The key variables are: access to sovereign capital to fund infrastructure, willingness to use geopolitical neutrality as a bargaining chip, and the ability to attract international AI talent through regulatory flexibility. Several Gulf states, Singapore, and potentially India are studying or replicating elements of this approach. For US foreign policy, the risk is that the UAE model normalises technology non-alignment — states building sovereign AI capacity while refusing to subordinate it to US security frameworks.

Foreign Policy's analysis that US frontier AI models are systematically too expensive for most of the world is not merely a market observation — it is a structural geopolitical vulnerability. Foreign Policy argues that China's open-weight models, including DeepSeek's series, are not just cheaper but are being actively distributed across the Global South as infrastructure — embedded in government services, education systems, and commercial platforms in markets where US API pricing is prohibitive. This creates path dependency: countries that build digital infrastructure on Chinese AI models will, over time, develop technical, economic, and political affinities with Chinese technology governance frameworks.

This dynamic is the AI equivalent of the infrastructure debt trap debate from the Belt and Road era — except the dependency is harder to see and harder to unwind because it operates at the software and data layer rather than physical infrastructure. The US has no current policy instrument that addresses the price differential as a geopolitical variable. Export controls restrict what China can access; they do nothing to make US alternatives affordable for lower-income markets.