Geopolitics & Sovereign Positioning

Moonshot AI's decision to release Kimi K3 as a freely deployable model is qualitatively different from prior Chinese open-weight releases. According to Rest of World, governments can now run a top-tier model locally without paying for US hyperscaler cloud access, collapsing one of the key leverage points Washington has maintained over AI-dependent states. This matters most for middle-income and lower-income governments that have been de facto rule-takers in AI — dependent on AWS, Azure, or Google Cloud for inference capacity. A sovereign-deployable Chinese model at zero licensing cost changes their strategic options overnight.

The geopolitical framing here is precise: this is not primarily about consumer AI but about state AI infrastructure. Ministries running AI-assisted public administration, intelligence analysis, or military logistics no longer need to route queries through US-controlled cloud infrastructure or accept the data-residency compromises that entails. Washington's export control architecture — designed to restrict chip access and thereby constrain frontier AI capability — has limited effect on countries that can now receive capable models directly without building domestic compute. The Atlantic Council's concurrent argument that banning open-source AI is counterproductive gains urgency in this context: restricting US open models does not prevent Chinese model proliferation, it merely cedes that distribution channel.

CXMT's debut on Shanghai's Star Market — closing at 49 yuan against an 8.66 yuan offer price and reaching a 3.3 trillion yuan market capitalisation — is not just a stock market event. As reported by South China Morning Post, the Hefei-based memory chipmaker now exceeds ICBC in market cap, reflecting investor conviction that China's domestic memory sector will capture AI-driven demand for HBM and DRAM that US controls have sought to deny. The capital raised — 57.92 billion yuan — provides CXMT with resources to accelerate process node advancement and scale production.

The strategic read is that US restrictions on Samsung, SK Hynix, and Micron selling advanced memory chips to Chinese AI developers have created a protected domestic market for CXMT with guaranteed demand from Huawei, ByteDance, Baidu, and the PLA's AI procurement pipeline. Export controls intended to slow Chinese AI have instead subsidised Chinese memory chip investment by guaranteeing captive customers. This is a documented second-order consequence that US export control designers acknowledged as a risk but accepted — the question is whether CXMT's IPO represents an inflection point where Chinese memory quality is approaching a threshold sufficient for frontier AI training at scale.

The CFR survey of 350 foreign policy and AI experts — detailed in two CFR publications (CFR and CFR) — produces findings with direct implications for how AI functions as a tool of state power. Near-universal expert agreement that governments are failing to keep pace with AI, combined with the assessment that frontier labs are already gaining power relative to the public sector, points to a structural asymmetry: states are attempting to exercise AI as a geopolitical instrument while the actors who actually control frontier AI capability — a small number of private labs concentrated in the US and China — are not fully subordinate to state direction.

For US foreign policy, this creates a specific tension: the export control and alliance-building architecture assumes the US government has reliable access to and leverage over American AI frontier capability. But if OpenAI, Anthropic, and Google DeepMind are genuinely gaining autonomy relative to the public sector, Washington's ability to use AI capability as a diplomatic asset — promising access, threatening denial — becomes less reliable. The expert consensus on high uncertainty about frontier AI capabilities by 2035 also complicates allied AI agreements, since partners are being asked to align with a technology trajectory that even the most informed analysts cannot predict with confidence.

Vietnam's new AI law, analysed by The Diplomat, positions Hanoi as the leading AI regulatory actor in Southeast Asia — ahead of Indonesia, Thailand, and the Philippines in enacting binding domestic AI governance. The Diplomat's assessment notes important shortcomings in the approach, but the strategic significance is the direction of movement: Vietnam is establishing a domestic regulatory framework that will shape which foreign AI providers can operate in its market and on what terms. This is a sovereignty-affirmation move as much as a consumer protection exercise.

Vietnam's positioning is consistent with its broader foreign policy of strategic ambiguity — maintaining economic relationships with both the US and China while building domestic institutional capacity to manage both. An AI law with data localisation or algorithmic transparency requirements functions as leverage over both US cloud providers and Chinese AI platforms, giving Hanoi regulatory tools to play both sides. ASEAN's broader failure to harmonise AI governance means Vietnam's framework, if functional, becomes a reference point for neighbours — and the question is whether it tilts more toward the EU's risk-based model, China's state-centric control model, or a novel Southeast Asian variant.