Geopolitics & Sovereign Positioning

President Trump signed an executive order on AI oversight that requires companies to give the government 30 days to vet new frontier models before public release, alongside the creation of an AI cybersecurity clearinghouse. The order was a significant dilution of an earlier draft that proposed a 90-day window — itself shelved under industry pressure — and was signed without a public ceremony. According to The Diplomat, the framing is explicitly strategic: advanced AI models are now treated as national security assets, not commercial products, and the 90-day window was abandoned precisely because it was judged to create an unacceptable innovation gap relative to China.

The 30-day vetting requirement has enforcement mechanisms attached — it is enacted policy, not a diplomatic commitment — but its practical bite depends entirely on how 'frontier model' is defined and whether the cybersecurity clearinghouse develops genuine adjudicatory capacity. The order's geopolitical significance lies less in its specific provisions than in what it reveals: the US government now openly acknowledges that domestic AI regulation must be calibrated against competitor speed, a logic that will constrain how far future administrations can tighten oversight without industry pushback framed as a national security argument. South China Morning Post notes the order was watered down after concerns it would hinder competition with China — meaning Beijing's perceived pace is now an explicit input into Washington's governance decisions.

The US Department of Commerce issued guidance on Sunday to halt Nvidia Blackwell and Rubin processor shipments — along with AMD's MI350x — to overseas subsidiaries of Chinese AI firms, specifically targeting entities in Malaysia, Singapore, and similar jurisdictions. As South China Morning Post reports, the loophole had existed for approximately a year and had allowed Chinese entities to acquire frontier chips by routing purchases through non-Chinese incorporated subsidiaries. The guidance closes this vector but is described as 'unexpected,' suggesting reactive rather than anticipatory enforcement.

This action has two immediate second-order consequences. First, it substantially raises the compliance burden for Southeast Asian tech firms that host Chinese-affiliated operations — Malaysia's semiconductor and data centre ecosystem is directly implicated. Second, it confirms that Chinese AI firms had already identified and exploited the subsidiary routing vector, meaning some volume of the most advanced chips has already entered the Chinese AI supply chain. The closure of the loophole now, rather than at the point of original export rule design, illustrates a persistent structural problem in US export control architecture: rules are written for known evasion vectors while adversaries operate on novel ones. The question of whether this guidance constitutes a formal regulatory amendment with penalties or interim administrative direction with uncertain enforcement teeth is not yet resolved.

Under sustained Nvidia exclusion, China's AI chip industry is now engaged in a foundational design debate: whether to develop GPU-style general-purpose accelerators or commit to ASIC-based architectures optimised for specific AI workloads. As South China Morning Post reports, the competition is no longer about replicating Nvidia — it is about building a self-sufficient silicon ecosystem across the full stack. Domestic players including Huawei (Ascend series), Cambricon, and a range of ASIC-focused startups are operating in parallel tracks, with the market yet to converge on a dominant domestic architecture.

This divergence is strategically significant because it means the US-China AI hardware ecosystems are now on trajectories toward architectural incompatibility — not just a capability gap. If China's domestic ecosystem coalesces around ASIC-dominated infrastructure, it will create a distinct AI hardware standard that third countries adopting Chinese AI infrastructure will be locked into, mirroring the 5G infrastructure dynamic. The embodied AI dimension amplifies this: South China Morning Post reports that Hangzhou-based Spirit AI's foundation model for physical AI surpassed Nvidia's Cosmos 3 on a global benchmark just two days after the latter's launch — a direct signal that robotics and embodied intelligence are now active battlegrounds, not future ones.

DeepSeek is closing its first-ever external funding round at over 50 billion yuan ($7.4 billion), implying a valuation of just under $60 billion — a six-fold increase from its $10 billion valuation in April. Backers include Tencent and CATL, as reported by South China Morning Post. The round marks a strategic pivot: DeepSeek had previously operated without external capital, a posture that kept it outside the conventional Chinese tech investment ecosystem. Its inclusion of CATL — a battery and energy technology conglomerate — points toward embodied and industrial AI applications as a capital deployment thesis.

Simultaneously, Hong Kong is consolidating its role as the primary internationalisation vehicle for mainland Chinese AI firms. MiniMax and Zhipu (Knowledge Atlas Technology) have both reversed the conventional dual-listing sequence — listing in Hong Kong first to establish global valuations before seeking mainland A-share listings — as South China Morning Post reports. The Hong Kong Generative AI Research and Development Centre (HKGAI) has simultaneously launched HKGAI-V3, a DeepSeek-based model optimised to run on domestic chips, explicitly framed as a platform for exporting Chinese AI overseas, per South China Morning Post. Hong Kong is functioning as both financial launchpad and technology export corridor — a dual role that makes it a critical node in China's AI internationalisation strategy and a point of ongoing US regulatory attention.