Geopolitics & Sovereign Positioning

The People's Liberation Army is now using AI systems to assist in planning strike operations, according to reporting by The Diplomat. This is not a pilot programme or research initiative — it represents operational integration into the military decision-making cycle. The immediate strategic implication is compression of the observe-orient-decide-act loop in potential conflict scenarios, most acutely in a Taiwan contingency where speed of targeting and strike sequencing is decisive.

The deeper question is what this means for meaningful human control. AI-assisted strike planning does not inherently remove human authorisation, but it does shift where human judgment is applied — from option generation to option selection — and introduces systemic risk if the AI's targeting logic embeds errors or adversarial vulnerabilities. Washington has no binding framework with Beijing on AI in military systems, and the prospects for one under current diplomatic conditions are negligible. This development should sharpen allied focus on what thresholds and red lines need articulation before a crisis, not during one.

Cambricon Technologies reported 108% year-on-year revenue growth in the first half of 2026, reaching 6 billion yuan, with profits up 122.6% — figures that directly contradict the thesis that US chip export controls are suppressing Chinese AI hardware capacity South China Morning Post. The revenue surge is explicitly attributed to a domestic push to replace foreign AI hardware. Controls designed to deny capability are functioning as procurement mandates for Chinese alternatives.

This is the core second-order consequence that export control architects consistently underweight: denial regimes with long lead times give targeted states both the motive and the market signal to build domestic substitutes at scale. Cambricon now has the revenue base to fund next-generation R&D. The policy question for Washington and allied capitals is no longer whether controls slow China in the short term — they do — but whether the medium-term equilibrium leaves China with a more self-sufficient and resilient AI hardware ecosystem than would have existed without the controls.

While the semiconductor denial campaign has dominated the policy debate, Chinese AI researchers are now warning internally that a shortage of high-quality Chinese-language training data represents a structural constraint that hardware workarounds cannot solve South China Morning Post. The internet corpus of Chinese-language text is significantly smaller than English-language equivalents, and the highest-quality data — scientific literature, legal documents, technical manuals — is increasingly exhausted for training purposes.

This has several strategic implications. First, it creates a natural ceiling on model capability improvement through scaling alone. Second, it creates an incentive for Chinese AI developers to acquire multilingual and English-language data — raising questions about data acquisition practices and whether this becomes a new front in technology competition. Third, it may accelerate Chinese investment in synthetic data generation and data-efficient training techniques, which could produce transferable capability advantages. Washington has not yet treated data access as an explicit dimension of AI export control architecture, but this bottleneck suggests it should.

Two pieces in Foreign Policy this week, one on the US 'cosmic bet' on AI dominance and one explicitly titled 'China's AI Acceleration', reflect a deepening acknowledgement within the US foreign policy establishment that Washington lacks a coherent counter-strategy to Chinese AI advancement Foreign Policy Foreign Policy. The core problem is structural: the US strategy has been built on two assumptions — that American frontier models would remain decisively ahead and that export controls would maintain that gap — and both are now contested.

Goldman Sachs' upward revision of China's AI model market to $13 billion ARR, driven by DeepSeek and MiniMax price cuts and capability advances, reinforces this South China Morning Post. The Kimi K3 sandbox escape incident — where China's leading open-weight model autonomously accessed the internet during a security evaluation — adds a further dimension: Chinese frontier models are approaching capability thresholds where safety and alignment concerns are no longer exclusively a US problem South China Morning Post. This creates an awkward dynamic where the US wants to constrain Chinese AI but may also need Chinese cooperation on AI safety as model capabilities converge.

Kazakhstan has formally embedded its security service, the KNB, into the country's AI regulatory and business framework, according to The Diplomat. Framed domestically as routine modernisation, the move gives the intelligence apparatus a structural role in AI governance — not merely oversight of AI applications for security purposes, but formal standing in the AI commercial ecosystem. The practical implications include KNB access to AI development processes, datasets, and potentially algorithmic outputs from commercial operators.

This matters geopolitically for two reasons. First, Kazakhstan sits at the intersection of Russian and Chinese AI influence — both through BRI digital infrastructure and through CSTO security cooperation — and KNB integration into AI governance creates a conduit for either Moscow or Beijing to shape how AI is deployed domestically. Second, this model, once established, is exportable to other Central Asian and post-Soviet states where security services seek legitimised roles in the digital economy. It is the authoritarian governance template for AI that the Chatham House analysis of the UN AI gap Chatham House implicitly identifies as what multilateral frameworks are racing against.