Geopolitics & Sovereign Positioning
Top Line
China's strategic pivot to controlling AI inference infrastructure — the deployment layer where tokens are produced and monetised — represents a structural challenge to US chip-centric export control logic, as Beijing bets that owning the demand side of AI compute reshapes leverage regardless of who fabricates leading-edge chips.
ASML raised its 2026 sales forecast to €36–40 billion despite declining China revenues, confirming that US-led export controls are successfully compressing China's access to advanced lithography while failing to disrupt the broader Western semiconductor ecosystem.
A US congressional hearing formally framed China's AI technology acquisition as a dual-track strategy — licit purchasing and illicit theft — with witnesses warning that US immigration and research policy restrictions are simultaneously degrading America's own innovation capacity.
China's photonics computing sector is accelerating toward public markets, with Lightelligence clearing its Hong Kong IPO listing hearing, signalling Beijing's push to develop alternative semiconductor architectures that circumvent conventional silicon supply chain vulnerabilities created by US controls.
The Trump administration's 'Pax Silica' supply-chain coalition concept is being stress-tested against three structural gaps — minerals processing capacity, human capital pipelines, and standards-setting authority — without which it risks being a diplomatic framework without operational teeth.
Key Developments
China's Inference Strategy: Contesting the AI Value Chain Where Controls Are Weakest
A strategically significant analytical frame is crystallising around China's positioning in AI: while the US controls chip fabrication and advanced model training infrastructure, China is consolidating control over the inference layer — the mass deployment of AI that generates economic and strategic value at scale. As South China Morning Post argues, Beijing has concluded from the demand side what Nvidia's Jensen Huang articulated from the supply side — that inference tokens are the commodity of the AI economy, and whoever owns that compute revenue owns the scoreboard. This reframes the export control debate: restricting training chips constrains frontier model development but does not prevent China from deploying and monetising existing models at scale across domestic and Global South markets.
China's open-source model strategy reinforces this inference-layer play. As South China Morning Post reports, Alibaba and others have used open-sourcing to build global developer ecosystems and reduce per-token compute costs, lowering the barrier for widespread deployment. The strategic logic is explicit: free models seed dependency on Chinese AI infrastructure and standards globally, particularly in markets where US export controls create a vacuum. The question now — acknowledged in the article — is whether Chinese firms can sustain this model as competitive differentiation erodes and proprietary alternatives become necessary.
Export Controls: ASML Data Confirms Compression of China's Advanced Chip Access, but Adaptation Accelerates
ASML's Q1 2026 results provide the clearest quantitative signal yet that the US-Netherlands-Japan export control coalition is materially affecting China's semiconductor trajectory. South China Morning Post reports China's share of ASML shipments has declined significantly even as ASML raises its total revenue forecast, demonstrating that controls are redirecting demand — not destroying it — toward US, Korean, and Taiwanese fabs. ASML management explicitly flagged anticipation of a new US control round targeting China, suggesting the coalition's restrictive architecture is not yet at its final configuration.
The second-order adaptation dynamic is playing out in China's alternative hardware bets. Lightelligence's Hong Kong IPO clearance — reported by South China Morning Post — and the broader acceleration of silicon photonic computing reflects Beijing's strategic calculation that optical-electronic hybrid architectures could reduce dependence on EUV-produced conventional chips. This is not a near-term capability replacement but a hedge against a scenario where export controls tighten further. Chinese PCB suppliers like Victory Giant — which counts Nvidia among its customers — simultaneously pursuing Hong Kong listings reveals the dual-use nature of China's AI hardware ecosystem: integrated into the Western supply chain while building sovereign alternatives.
'Pax Silica' and the Structural Gaps in US AI Alliance Architecture
The Atlantic Council's analysis of the Trump administration's emerging AI supply-chain coalition — framed as 'Pax Silica' — identifies three structural deficits that separate a diplomatic concept from an operational power bloc: minerals processing capacity, human capital, and standards-setting authority. Atlantic Council argues that without resolving critical minerals processing — currently dominated by China across the rare earth and battery supply chain — any chip-centric coalition retains a structural vulnerability at the upstream end. On human capital, the analysis points to the contradiction in US immigration policy tightening precisely as the administration seeks to consolidate AI leadership through allied talent pools. On standards, the concern is that without active US engagement in international AI standards bodies, China and the EU will shape the technical and governance norms that determine interoperability and market access.
The Pax Silica framing reflects a genuine strategic instinct — that AI dominance requires a coalition approach across the full stack — but the gap between concept and implementation is wide. The administration's current posture combines aggressive export controls with diplomatic commitments to allied coordination, but the binding mechanisms and burden-sharing arrangements that would make such a coalition durable are not yet enacted. This is a proposal-stage framework, not an operational alliance.
China's Dual-Track Tech Acquisition and the US Self-Inflicted Capability Risk
Congressional testimony this week formalised what intelligence assessments have long indicated: China operates a deliberate two-track AI technology acquisition strategy — purchasing what is available through licit commercial channels and extracting through theft what is restricted. South China Morning Post reports witnesses characterised China as 'dependent on our tech stack' — a framing that both validates export control logic and underscores the stakes of enforcement gaps. The Anthropic identity verification story reinforces this: Chinese demand for Claude models is significant enough to sustain black-market workarounds within days of Anthropic tightening access controls, as South China Morning Post documents.
Critically, the same congressional witnesses flagged that US immigration and research funding restrictions are creating a self-inflicted capability constraint — reducing the inflow of AI talent and constraining the academic research base that feeds frontier model development. This is the central tension in current US AI strategy: the same political coalition driving export controls is also driving the immigration and university research policies that erode the domestic innovation base those controls are designed to protect.
Signals & Trends
Hong Kong is Emerging as China's AI Capital Markets Bridge Under Geopolitical Pressure
Three distinct Chinese AI hardware companies — Lightelligence in photonics, Victory Giant in PCBs, and others in the broader circuit-board space — are simultaneously pursuing Hong Kong IPOs as their capital raising venue of choice. This is not coincidental. As US capital markets become increasingly inaccessible to Chinese tech firms through delisting pressure and CFIUS scrutiny of US investor exposure, Hong Kong is being repositioned as the liquidity bridge between Chinese AI hardware ambitions and international capital. This matters geopolitically because it means the financing of China's AI semiconductor diversification strategy is proceeding with international capital through a jurisdiction the US does not directly control. Washington's leverage over this funding pipeline is limited, and the trend will accelerate if US-China financial decoupling continues.
World Models as the Next Frontier Competition: Physical-Reality AI Is Moving from Labs to Geopolitical Stakes
The simultaneous announcements from Alibaba and companies associated with Li Fei-Fei in the world models space — AI systems designed to understand and simulate physical reality, not just language — signal that the next phase of AI competition is moving beyond large language models into embodied and physical-world AI. This has direct military and industrial implications: world models are the foundational architecture for robotics, autonomous weapons, manufacturing automation, and logistics optimisation. The race to establish foundational capabilities and, critically, the training data and benchmarks that define progress in this domain is just beginning. Whichever bloc establishes early leadership in world model architectures will have a structural advantage in the physical-world AI applications that carry the highest strategic value — and current US export controls are not calibrated to this dimension of the competition.
The Inference Infrastructure Gap May Become the Next Major Export Control Battleground
The emerging analytical consensus — visible across multiple signals this week — is that US export controls have focused on training compute while inference infrastructure has remained relatively unaddressed. As AI economics shift toward deployment-scale inference as the primary value generator, the inference hardware layer becomes strategically significant in its own right. China's open-source model strategy, its photonics hardware development, and its positioning as an inference-at-scale provider to Global South markets all converge on this gap. US policymakers face a structurally difficult choice: extending controls to inference chips would be far more economically disruptive to allied relationships — given how widely inference hardware is deployed commercially — but failing to do so may allow China to build the deployment infrastructure dominance that renders training-level controls strategically insufficient.
Explore Other Categories
Read detailed analysis in other strategic domains