Geopolitics & Sovereign Positioning

In a memo released Thursday, White House science and technology adviser Michael Kratsios warned that 'surreptitious, unauthorised distillation campaigns' were enabling Chinese entities to release models that appear to match US frontier capability at a fraction of the training cost. The warning signals the Trump administration is prepared to extend export control logic beyond chips to model weights and training data access, closing a loophole that has allowed Chinese labs to compress the capability gap by learning from American models rather than building from scratch. This is a policy proposal at this stage, not an enacted restriction with enforcement mechanisms, but the political signal alone carries weight given the administration's demonstrated willingness to act unilaterally.

Analysts cited by South China Morning Post suggest the practical effect could be a shakeout of weaker Chinese AI players within a year, as smaller labs that relied on distillation from GPT-4 class models lose their arbitrage advantage. The more durable Chinese players — those with genuine pre-training capability and domestic compute — would actually benefit from reduced competition. Paradoxically, aggressive enforcement could accelerate exactly the domestic model sovereignty Beijing is pursuing by forcing Chinese labs to build foundational capability rather than shortcut through distillation.

Huawei announced day-zero adaptation of its Ascend 950PR and 950DT chips to DeepSeek's V4 model in a livestream timed to coincide with V4's release, a deliberate demonstration of ecosystem coordination that mirrors the Nvidia CUDA-ecosystem lock-in model that has defined US AI infrastructure dominance. As reported by South China Morning Post, this is not merely a technical milestone — it is a geopolitical signal that China's domestic AI stack is reaching a level of integration where performance on domestic hardware is a feature rather than a compromise.

The significance for export control strategy is direct: every increment of performance parity between Ascend chips and Nvidia H100-class hardware reduces the leverage embedded in US chip restrictions. The Huawei-DeepSeek coordination also illustrates how Beijing is using state-adjacent champions to build a vertically integrated AI supply chain — from chip design through model training to deployment — that reduces the number of foreign chokepoints. DeepSeek V4 Pro's benchmark performance, ranked second globally among open-source models per Artificial Analysis, suggests this stack is not merely symbolic. However, the same benchmarks place V4 Pro behind both US frontier closed models and domestic rival Kimi K2.6, indicating the gap at the absolute frontier persists.

Research by Kenneth Payne at King's College London, covered by War on the Rocks, finds that in 95 percent of simulated crisis wargames across 21 matchups between three frontier models, at least one side engaged in nuclear signaling. The paper is a pre-print and has not yet undergone peer review — that methodological caveat matters for policy interpretation — but the directional finding is consistent with earlier work showing LLMs lack reliable escalation-management heuristics when operating under simulated time pressure and incomplete information.

The strategic stability implications are distinct depending on how AI is actually being integrated into military decision chains. No major power has publicly confirmed LLM use in live nuclear command and control, but AI-assisted intelligence fusion, targeting recommendation, and crisis communication systems are confirmed in development across the US, China, and Russia. The danger is not that an LLM launches a missile autonomously but that AI-assisted decision support in a genuine crisis systematically biases human operators toward escalatory options, particularly if adversary AI systems are doing the same. The absence of any bilateral or multilateral agreement on AI use in nuclear-adjacent decision systems — analogous to the 1963 hotline agreement — is an increasingly visible gap in the arms control architecture.

State Grid Corporation of China's 6.8 billion yuan commitment to deploy thousands of AI-powered robots across power grid infrastructure — covering substation inspection, high-voltage line maintenance, and operational roles — represents a qualitatively different category of AI deployment than consumer or enterprise software. As reported by South China Morning Post, this is state-directed sovereign infrastructure investment with a direct national resilience dimension: reducing human labor dependency in grid operations reduces vulnerability to workforce disruptions and, in a conflict scenario, reduces the attack surface associated with human operators in sensitive locations.

The initiative also reflects a broader Chinese strategic pattern of embedding AI into physical infrastructure — power, logistics, transport — as a form of technological sovereignty that is harder to sanction or disrupt than software dependencies. Western export controls target chips and model weights; they cannot easily reach an already-deployed robot fleet operating on domestically manufactured hardware. The optical communications sector boom documented by South China Morning Post, with firms like Zhongji Innolight seeing tenfold share appreciation, reflects the same dynamic: China is building out AI-enabling physical infrastructure at a pace and scale that creates structural long-term advantages independent of any single model's benchmark performance.