Geopolitics & Sovereign Positioning
Top Line
Anthropic's suspension of its most advanced AI model access for South Korean users has exposed a structural flaw in U.S. export control architecture: Silicon Valley firms operating as de facto policy instruments are penalising treaty allies without diplomatic recourse, creating leverage gaps that China is positioned to exploit.
Chinese domestic AI chip suppliers — led by Huawei and Cambricon — are projected to capture nearly 80 percent of China's AI server market in 2026, a structural shift driven directly by U.S. export controls that have accelerated Chinese substitution rather than constrained it.
China's state-backed Big Fund has become SMIC's third-largest shareholder following a $6 billion consolidation deal, signalling Beijing's transition from subsidy-based industrial policy to direct equity control over its most strategically critical semiconductor manufacturer.
Zhipu AI's GLM-5.2 open-weight coding model is drawing comparisons to DeepSeek's January 2025 disruption, with American researchers praising its cost-performance ratio — reinforcing the pattern that U.S. export controls are not preventing frontier-competitive Chinese AI model development.
China's push into photonic chips and 'air-space-ground-sea' integrated network infrastructure represents a dual-track strategy: circumventing silicon-based export control chokepoints while building sovereign AI infrastructure resilient to future sanctions.
Key Developments
Anthropic's South Korea Suspension: Export Controls as Allied Collateral Damage
Anthropic's decision to suspend access to its most advanced AI model for South Korean users has generated significant friction with a core U.S. security partner, exposing a governance gap in how Washington's AI export control regime interfaces with allied economies. As The Diplomat reports, the episode illustrates how Silicon Valley firms — increasingly operating as instruments of U.S. technology policy — can impose compliance costs on allied markets without the diplomatic safeguards that normally accompany government-to-government restrictions. South Korea receives no formal carve-out or consultation mechanism comparable to those embedded in AUKUS or the Chip 4 alliance framework.
The strategic consequence is twofold. First, it creates a market opening: South Korean enterprises denied access to Anthropic's frontier models will evaluate alternatives, including Chinese offerings like Zhipu AI's GLM-5.2 and Alibaba's Qwen series, both of which are being aggressively priced for international uptake. Second, it signals to other U.S.-aligned but non-Five Eyes partners — Japan, the Netherlands, India — that proximity to Washington's AI ecosystem does not guarantee stable access, weakening the credibility of U.S. technology alliance commitments. Seoul's response and whether it formally protests through diplomatic channels or quietly redirects procurement will be a leading indicator of allied tolerance for this pattern.
China's Semiconductor Consolidation: State Capital Takes Direct Equity Control
The completion of SMIC's $6 billion acquisition of its manufacturing subsidiary — with China's National Integrated Circuit Industry Investment Fund (the 'Big Fund') emerging as the third-largest shareholder — marks a qualitative shift in Beijing's semiconductor strategy. Previous phases of the Big Fund focused on debt financing and minority equity stakes in a broad portfolio of firms. Direct, concentrated ownership of SMIC, China's only firm with meaningful advanced node capability, signals a move toward command-economy logic for the most strategically sensitive layer of the AI supply chain. South China Morning Post reports this is the largest M&A transaction ever recorded on Shanghai's Star Market.
Simultaneously, JCET's 7.8 billion yuan investment in a new advanced packaging plant in Shanghai's Lin-gang Special Area, and Kuaishou's chip spin-off TranStreams completing a Series A+ round to develop proprietary AI silicon, illustrate that consolidation at the foundry level is being matched by vertical integration across the stack — from packaging to custom inference chips. The Lin-gang location is notable: it operates under distinct regulatory conditions from mainland China, including provisions designed to attract foreign-invested entities, and sits within a zone Beijing has designated for advanced manufacturing scale-up.
U.S. Export Controls Accelerating Chinese Chip Substitution, Not Containment
TrendForce data, cited by South China Morning Post, projects Chinese suppliers will hold roughly 79 percent of China's domestic AI server chip market in 2026, with Nvidia, AMD, and other foreign suppliers reduced to 21 percent — down from majority positions held before the 2022–2023 export control escalations. This is the clearest available market-level evidence that the primary effect of chip controls has been demand substitution rather than capability suppression. Huawei's Ascend series and Cambricon's MLU line are the primary beneficiaries, and both are receiving preferential procurement signals from state-owned enterprises and government cloud contracts.
Micron's acknowledgment during its Q3 2026 earnings call that CXMT and YMTC have 'grown in capability and market share' adds independent commercial validation to Chinese progress in memory — a segment critical to AI inference workloads. China's photonic chip program adds a longer-term vector: Shanghai's new optical computing laboratory, opened in June 2026, is pursuing architectures that bypass the silicon process node constraints that U.S. controls target. If photonic inference chips achieve commercial viability at scale, the entire export control framework — premised on controlling advanced EUV lithography and silicon node density — becomes strategically obsolete for the AI application layer.
AI in Nuclear Command and Military Logistics: Doctrine Hardening, Risks Unresolved
A War on the Rocks reassessment by Paul Scharre and Michael Depp of their 2024 nuclear stability analysis finds that AI integration into military command-and-control has accelerated substantially, but without commensurate development of safeguards, arms control frameworks, or bilateral risk-reduction mechanisms between the U.S. and China. War on the Rocks situates this within a broader pattern: AI is being integrated into the operational and administrative layers of military organizations faster than doctrinal or legal frameworks can absorb. A separate War on the Rocks discussion with a Salesforce national security executive highlights that AI agents are being deployed in military logistics — coordination, supply chain optimization, administrative judgment — in ways that blur accountability chains when errors cascade.
Ukraine's wartime AI experimentation, highlighted by the Atlantic Council, is generating battlefield-validated use cases that NATO members are actively studying — particularly in ISR fusion, drone coordination, and damage assessment. The Atlantic Council piece frames Ukraine as a 'living laboratory' whose lessons Western governments should incorporate into their AI strategy development. This is not speculative: multiple European defence ministries have active exchange programs with Ukrainian defence tech units, and the U.S. Defense Innovation Unit has maintained formal engagement channels with Ukrainian startups throughout the conflict.
Chinese AI Model Competitiveness: DeepSeek Expansion and Zhipu's GLM-5.2 Signal Sustained Frontier Pressure
DeepSeek's simultaneous moves — aggressive hiring across 33 positions to 'at least double every department,' rapid expansion of its Harness AI agent team under former Jane Street quantitative trader Cui Tianyi, and the commercial reception of its models — confirm that the firm is transitioning from a research-oriented organization to a full-stack AI product company pursuing autonomous agent deployment. The Harness team's focus on AI agents is strategically significant: agents represent the commercial application layer where AI generates sustained enterprise revenue, and DeepSeek's entry into this market directly challenges both U.S. incumbents and Chinese rivals like Alibaba and Baidu.
Zhipu AI's GLM-5.2, described by South China Morning Post as an open-weight model attracting praise from American researchers for its coding performance and cost profile, adds a second Chinese firm to the list of those producing frontier-competitive outputs despite chip access constraints. Alibaba's parallel move — cutting Qwen model prices by up to 80 percent on its international Qoder platform during U.S. business hours — reflects deliberate market capture strategy targeting American developer workflows. Taken together, these developments indicate that Chinese AI firms are competing simultaneously on model capability, cost, and distribution, not retreating to a domestic-only market.
Signals & Trends
China Is Building Architectural Alternatives to Silicon-Based AI, Not Just Catching Up Within the Existing Paradigm
The opening of Shanghai's optical computing laboratory, Fysics AI's physics-embedded world model that explicitly departs from the data-driven architectures favoured by OpenAI and Meta, and the broader photonic chip investment program collectively signal that China is pursuing architectural divergence — not just incremental silicon progress — as a hedge against permanent disadvantage at advanced nodes. If any of these alternative architectures achieves commercial viability for AI inference workloads, the entire U.S. export control framework, which is calibrated to silicon lithography thresholds, loses its primary leverage point. This is a long-duration risk, not a near-term capability shift, but the institutional investment suggests Beijing is treating it as a core strategic option rather than a research curiosity. Western intelligence and policy communities should be tracking not just node-level progress at SMIC but investment flows and publication rates in photonic and neuromorphic computing.
The 'Swing State' Problem Is Sharpening: U.S. Controls Are Pushing Non-Aligned Economies Toward Chinese AI Providers
The South Korea episode is unlikely to be isolated. The pattern — where U.S. export control enforcement or compliance decisions by American firms disrupt access for non-Five Eyes partners without diplomatic compensation — will recur as AI model controls tighten. Countries in Southeast Asia, the Gulf, and Latin America that lack formal technology alliance status with Washington are watching how Seoul handles this situation. Simultaneously, Chinese firms are pricing their models and infrastructure competitively for exactly these markets: Alibaba's Qwen pricing cuts target international users explicitly, and Chinese telecoms firms are framing their 'air-space-ground-sea' network infrastructure as a global offering. The geopolitical contest for the Global South's AI alignment is being fought less through diplomatic frameworks and more through API pricing, cloud infrastructure terms, and the reliability of access — a domain where U.S. export control overreach is currently generating self-inflicted disadvantage.
China's Vertical Integration of AI Infrastructure Is Approaching Strategic Self-Sufficiency at the Application Layer
The convergence of SMIC consolidation under state equity, JCET's advanced packaging expansion, Kuaishou's custom inference chip development, Huawei and Cambricon's server chip market dominance, and CXMT and YMTC's acknowledged memory gains means China is assembling a complete, domestically controlled AI hardware stack. This does not mean parity with U.S. frontier training clusters — SMIC's node limitations remain real — but it increasingly means China can build and operate competitive AI inference infrastructure without foreign components. For applications below the absolute frontier — enterprise AI, autonomous systems, surveillance infrastructure, military logistics — this level of self-sufficiency is strategically sufficient. The implication for U.S. policy is that the window for using export controls to constrain Chinese AI deployment at scale is narrowing, and the leverage that remains is concentrated at the training frontier rather than across the application landscape.
Explore Other Categories
Read detailed analysis in other strategic domains