Geopolitics & Sovereign Positioning
Top Line
DeepSeek's open-sourcing of Huawei-optimised AI software tools marks a concrete step toward China's operational independence from Nvidia, directly targeting the gap that US export controls were designed to preserve.
OpenAI's provision of advanced GPT 5.6 Sol cyber defence capabilities to Ukraine represents a US-aligned commercial actor becoming a direct participant in active conflict — a significant precedent for AI's military integration.
Anthropic's public warning that Chinese firm Z.ai's open-weight GLM-5.3 nearly matches its best model on cyber capabilities but carries far weaker safety constraints signals that export controls are not preventing dangerous capability diffusion via open-weight releases.
The Trump-Xi summit produced a two-month trade truce extension and nascent AI dialogue channels but no binding commitments, leaving the structural US-China AI competition unchanged while Trump explicitly ruled out AI cooperation.
China's generative AI user base crossing 700 million — over half the population — gives Beijing a domestic deployment scale advantage that is increasingly difficult for Western rivals to match, hardening China's position as a mass-market AI power regardless of chip access constraints.
Key Developments
DeepSeek-Huawei Axis: China's Concrete Move to Escape Nvidia Dependency
DeepSeek on Wednesday open-sourced six software modules optimised for Huawei's Ascend AI chips, explicitly designed to replicate the functionality of its prior Nvidia-facing tools and construct what it terms an 'independent and controllable' GPU software ecosystem. This is not a research announcement — it is a practical infrastructure layer that lowers the barrier for Chinese developers and enterprises to migrate AI workloads from Nvidia hardware to domestic alternatives. South China Morning Post
The strategic logic is direct: US export controls have denied China access to Nvidia's most advanced chips, but software ecosystem lock-in has been a secondary constraint — developers and firms accustomed to Nvidia's CUDA environment face friction migrating to Ascend. DeepSeek's move begins to dissolve that friction. Combined with Alibaba's concurrent announcement of a full-stack AI infrastructure roadmap emphasising domestic efficiency and monetisation at its Apsara Conference South China Morning Post, the picture is of coordinated Chinese tech sector investment in supply-side AI sovereignty — not just at the model layer but at the infrastructure and tooling layer where US controls have had most leverage.
Z.ai's GLM-5.3 and the Open-Weight Proliferation Problem
Anthropic's published assessment that Z.ai's open-weight GLM-5.3 nearly matches its own frontier model on cyber offensive capabilities — while exhibiting substantially weaker safety guardrails — represents a direct challenge to the assumption that export controls can manage dangerous AI capability diffusion. Open-weight models, once released, cannot be recalled or access-controlled. GLM-5.3's combination of high cyber capability and low safety constraint means any actor globally can download and deploy a near-frontier offensive cyber tool with minimal friction. South China Morning Post
This development directly undermines the policy architecture underpinning US AI export controls, which focus on restricting chip access to slow Chinese model development. GLM-5.3 demonstrates that Chinese firms can reach near-frontier cyber capability without necessarily depending on the restricted hardware — or that sufficient compute workarounds exist. It also raises a secondary question about Ukraine: OpenAI's simultaneous provision of GPT 5.6 Sol for Ukrainian cyber defence BBC occurs in a threat environment where adversary AI cyber tools are proliferating through open-weight channels that no bilateral agreement can close.
Trump-Xi Summit: AI Dialogue Channels Open, But No Binding Architecture
The Trump-Xi summit in Washington — Xi's first US state visit in eleven years — produced a two-month extension of the existing trade truce and agreement to establish 'potential channels' on AI and military crisis communication, according to Lawfare's analysis with Brookings' Jonathan Czin Lawfare. Expert opinion is divided on the summit's significance: some analysts characterise it as substantive diplomatic progress, others as procedural delay without structural resolution on Taiwan, technology transfer, or military posture War on the Rocks.
Trump's explicit and public rejection of AI cooperation with China — stating collaboration would give Beijing 'immediate access to things we know that they don't know' — signals that the administration's posture remains competitive rather than coordinative, even as dialogue channels are nominally opened. The reliance on industry self-policing domestically, with Trump declining to impose additional regulatory constraints on US AI giants South China Morning Post, creates a structural tension: the US is neither cooperating with China nor robustly regulating its own sector, leaving AI safety and proliferation risks managed by neither bilateral agreement nor domestic mandate.
OpenAI's Ukraine Cyber Defence Deal: Commercial AI Actors Enter Active Conflict
OpenAI's provision of GPT 5.6 Sol — described as a rival to Anthropic's Mythos and Fable model tier — to Ukraine for cyber defence purposes marks a qualitative shift in how frontier AI is being integrated into live conflict environments. This is a confirmed enacted arrangement, not a diplomatic proposal. The deal positions a US commercial AI firm as a direct capability provider to a belligerent in an active war, with implications for how adversaries perceive AI-enabled cyber operations and for the legal and normative frameworks governing AI in armed conflict. BBC
The timing against the backdrop of Anthropic's GLM-5.3 warning is notable: Ukraine is receiving US-aligned frontier AI cyber defence tools at precisely the moment that open-weight Chinese models with near-frontier offensive cyber capability are freely available for download. This asymmetry — Ukraine getting a controlled, commercially provided defensive AI capability while its adversaries can access comparable open-weight offensive tools — illustrates the fundamental instability of the current AI-in-conflict regime. It also sets a precedent that other US-aligned governments will reference when seeking preferential AI capability transfers for defence purposes.
Global South AI Governance: Countries Seeking Independent Safety Evaluation Capacity
A Rest of World event in New York surfaced a structural concern among countries outside the US-China axis: as they adopt AI models from OpenAI, Anthropic, and other frontier developers, they are becoming dependent on the safety evaluations those companies themselves conduct — with no independent national capacity to verify capability claims or safety properties. Speakers explored frameworks for countries to develop their own evaluation infrastructure, rather than accepting AI companies' embedded safety assessors as a substitute for sovereign oversight. Rest of World
This dynamic intersects directly with India's semiconductor positioning: India already possesses significant AI design talent but lacks the fabrication base and supplier ecosystem to translate that into sovereign AI infrastructure The Diplomat. Countries in this position — substantial human capital, limited sovereign infrastructure — are structurally positioned as rule-takers on AI safety standards unless they invest in evaluation capacity independent of the frontier developers whose models they deploy. The Asia-Pacific more broadly faces this alignment pressure, with nations needing to choose whether to rely on US or Chinese AI infrastructure and on whose safety standards to anchor their governance frameworks The Diplomat.
Signals & Trends
China's 700 Million User Base Creates an Evaluation and Fine-Tuning Advantage That Chip Controls Cannot Address
China's generative AI penetration rate exceeding 50% of the population — 700 million users as of June 2026, up 16% in six months — gives Chinese AI developers access to a feedback, evaluation, and fine-tuning dataset of a scale no Western competitor can replicate domestically. Export controls target compute and chip access; they do not address the deployment-scale advantage that comes from having half a billion active users generating preference data, edge cases, and real-world usage signals. As model performance increasingly depends on post-training alignment and fine-tuning at scale, this user base represents a structural advantage that compounds over time and is invisible to the chip export control framework.
The Pentagon-Anthropic Exclusion and the Emerging US AI Defence Supply Chain
The D.C. Circuit's rejection of Anthropic's challenge to its exclusion from Pentagon AI procurement — noted in Lawfare's coverage of the Trump administration legal developments — signals that US defence procurement is actively shaping which AI firms gain strategic relevance through government contracts, and that courts will not intervene on commercial grounds. This creates a bifurcating dynamic: firms that secure defence procurement relationships gain both revenue and the institutional validation that makes them strategically significant to allies seeking US-aligned AI capability. Firms excluded from defence supply chains — even if technically capable — are disadvantaged in the geopolitical AI partnership market. Watch for other allied governments to use Pentagon procurement status as a proxy signal for which US AI firms they engage for sensitive applications.
Open-Weight Chinese Models Are Becoming the De Facto AI Export Control Evasion Mechanism
The pattern across multiple developments this week — GLM-5.3's near-frontier cyber capability released as open-weight, DeepSeek's open-sourced Huawei tooling — points to open-weight model releases as a deliberate or emergent strategy for circumventing US-led technology controls. Unlike proprietary models or physical chips, open-weight models cannot be embargoed post-release. The US export control framework, built around hardware and supply chain interdiction, has no effective instrument against software once published. If Chinese frontier labs continue releasing capable open-weight models — whether for commercial, reputational, or strategic reasons — the practical effect is a sustained capability transfer to any actor globally, undermining the geopolitical rationale of the entire export control architecture. This is the sharpest unsolved problem in AI geopolitics and no allied government has articulated a credible policy response.
Explore Other Categories
Read detailed analysis in other strategic domains