Geopolitics & Sovereign Positioning

Nvidia's Jensen Huang and AMD's Lisa Su have accepted appointments to the advisory board of Tsinghua University's School of Economics and Management, one of China's most politically connected academic institutions, known for brokering relationships between global business leaders and senior Chinese government officials. The appointments, confirmed by South China Morning Post, come at a moment of acute tension: both companies are subject to US export controls that restrict the sale of their most advanced chips to China, and both are simultaneously central to American AI industrial policy.

The strategic logic is double-edged. For Huang and Su, maintaining access to Tsinghua's network preserves optionality in a market that — even restricted — remains commercially significant, and signals to Beijing that channels of communication remain open. For China, the appointments are a prestige signal and a potential source of informal intelligence on US industry direction. For Washington, they raise uncomfortable questions about whether the executives of America's most strategically sensitive chip firms should hold advisory positions at an institution with deep PLA research ties. This will attract congressional scrutiny and may accelerate calls for restrictions on executive-level institutional affiliations with designated Chinese universities.

The Center for Security and Emerging Technology has published a comparative cost-benefit analysis of two chip location verification methods — ping-based location verification (PLV) and physical inspections — concluding that PLV is more cost-effective and that a hybrid approach combining PLV with targeted physical inspections represents the optimal enforcement architecture. This matters because the central vulnerability of current AI chip export controls is not the restriction itself but verification: once chips are sold to nominally compliant jurisdictions, tracking their subsequent movement or use is operationally difficult. CSET frames PLV as technically feasible and economically justified at scale.

For foreign policy professionals, this shifts the debate on controls from 'are they working' to 'what would make them work.' The practical question is whether the US government will mandate PLV as a condition of sale for advanced AI chips — effectively embedding a surveillance mechanism into the hardware supply chain. This would represent a significant escalation in the extraterritorial reach of US export controls, and would face resistance from chip buyers in allied nations who object to US government visibility into their infrastructure. The geopolitical implication is that effective enforcement of AI chip controls may require a degree of monitoring infrastructure that allies find intrusive, creating friction within the very coalitions the US needs to maintain control effectiveness.

SemiAnalysis estimates that ByteDance accounts for approximately one-fifth of China's entire delivered data centre capacity, making it the country's single largest tenant and primary driver of AI infrastructure buildout. As reported by South China Morning Post, ByteDance rents nearly all of its footprint rather than owning it, meaning its AI compute capacity is intermediated through commercial landlords — a different risk profile than vertically integrated hyperscalers.

From a strategic standpoint, this concentration is significant in two directions. Domestically, it means that ByteDance's operational continuity — and its AI product suite, including the Doubao assistant — is structurally central to Chinese AI capability at scale. Any disruption to ByteDance (regulatory, financial, or geopolitical) would have outsized effects on China's deployed AI infrastructure. Externally, it reinforces Western concerns about ByteDance as a node through which the Chinese state could access or influence AI-generated data and inference at population scale. The rental model also means ByteDance's footprint is commercially visible and potentially subject to supply-side pressure — a lever that matters if China's domestic chip supply constraints tighten under export controls.

Samsung and SK Hynix together dominate the high-bandwidth memory market that is essential to frontier AI training and inference, and both are now competing aggressively to lock in supply relationships with Nvidia, OpenAI, and other US AI infrastructure players, according to Rest of World. This positions South Korea — not just its firms but its government — as a structural chokepoint in AI hardware supply chains. Seoul's policy choices on technology transfer, export controls alignment with Washington, and investment in domestic HBM capacity expansion carry direct consequences for US AI competitiveness.

The strategic implication is that the US-South Korea semiconductor relationship is no longer simply a commercial or alliance-management issue — it is a central pillar of US AI export control architecture. If Samsung or SK Hynix were to supply advanced HBM to Chinese AI firms in quantities that circumvent US controls, the entire chip control regime would be undermined. Conversely, South Korea's leverage over Washington increases as HBM becomes more critical: Seoul can extract concessions on trade, security commitments, or technology access in exchange for alignment on memory chip controls. This dynamic is likely to become more explicit in bilateral negotiations over the next 12 months.

Z.ai and Beijing-based safety consultancy Concordia AI have released what they describe as the first comprehensive, evidence-based framework for managing open-weight AI risk — a six-stage process designed to address the fundamental challenge that user-modifiable models cannot be controlled post-release. As reported by South China Morning Post, Chinese developers now dominate the open-weight ecosystem, making Beijing's position on safety governance for open-source models consequential far beyond China's borders. Separately, China's emergence as a Hugging Face competitor — with platforms ModelScope and MoArk vying to serve Chinese-speaking developers — signals an infrastructure-level effort to consolidate the open-source model distribution stack within Chinese-controlled platforms, as reported by Rest of World.

The strategic logic here is that whoever sets the de facto norms for open-weight model safety — and whoever controls the platforms through which open-source models are distributed globally — shapes the governance landscape more effectively than any regulatory body. China's move to publish safety frameworks positions Beijing as a responsible actor in international AI governance conversations while simultaneously building the domestic infrastructure to be the distribution layer for global open-source AI. This is standard-setting as soft power, and it mirrors China's playbook in 5G and other technology domains: publish standards, build infrastructure, then negotiate from a position of embedded influence.