Geopolitics & Sovereign Positioning

Moonshot AI's Kimi K3 has crystallised a strategic problem for Washington that DeepSeek first surfaced in early 2025: US chip export controls assume that compute scarcity translates into capability scarcity, but Chinese labs are demonstrating that architectural and training efficiency gains can erode that assumption faster than controls can tighten. The 2.8 trillion-parameter open-weight model — released with native vision capabilities and targeting long-horizon coding and knowledge work — has been assessed by Swiss firm Aikido Security as delivering cybersecurity performance 'extremely close' to OpenAI's GPT-5.6 Sol at a fraction of the cost, as reported by South China Morning Post. Moonshot was forced to suspend new subscriptions within 48 hours of launch due to compute constraints — a detail that simultaneously confirms global demand and highlights China's persistent GPU bottleneck, as covered by South China Morning Post.

The open-weight release strategy is the critical geopolitical variable here. By releasing weights publicly, Moonshot — like DeepSeek before it — distributes capability globally in a way that is structurally immune to subsequent export controls. US firms including Apple and Thinking Machines have already integrated Kimi K3, as Rest of World reports, creating a compliance and political problem for Washington: its own industry is the vector through which Chinese AI capability propagates into US infrastructure. Moonshot has accelerated fundraising at a reported $30 billion valuation ahead of a planned Hong Kong IPO, per South China Morning Post, signalling that capital markets are pricing in sustained frontier competitiveness.

Treasury Secretary Bessent's threat to sanction Chinese AI developers over alleged IP theft via distillation — the practice of training models on outputs of other models — has immediately run into a credibility problem. A broad coalition of global AI experts publicly rejected the distillation-as-theft framing on legal and technical grounds, with several characterising the accusations as 'political' and 'reckless', noting that model outputs are not copyrighted under current law, as South China Morning Post reports. Moonshot employees have directly denied the claims. This public pushback from non-Chinese experts matters geopolitically: it means the US cannot easily mobilise allied consensus around an IP-theft enforcement framework, which would be necessary for any sanctions regime to impose meaningful costs rather than merely signalling.

The Atlantic Council's analysis of country-based AI controls draws an instructive parallel with the spyware industry: verifying who is actually using a model or accessing an API at scale is an unsolved identity problem, and Chinese developers — or any sanctioned actor — can route access through intermediaries in jurisdictions with limited enforcement capacity, as Atlantic Council argues. The watermarking mechanisms Bessent referenced as a detection tool are nascent and contested. The net effect is that the administration has staked out a maximalist rhetorical position that it currently lacks the technical and legal infrastructure to enforce, risking the credibility of future controls.

The launch of the World Artificial Intelligence Cooperation Organization, with five Southeast Asian founding members, represents a qualitative shift in China's AI governance strategy. Beijing has moved from issuing bilateral commitments and positioning papers to constructing a dedicated multilateral institution — one that, if it gains membership and norm-setting traction, could serve as an alternative architecture to Western-led frameworks such as the OECD AI Principles or the G7 Hiroshima Process. The Diplomat frames this as China's most ambitious formal effort to shape AI deployment and governance globally; Foreign Policy notes the Southeast Asian composition of the founding membership, which reflects deliberate targeting of swing states in a region where both the US and China are actively competing for technology alignment.

The governance offensive has a dual function that The Diplomat identifies explicitly: catching up on AI capability by embedding Chinese models and infrastructure into partner countries' digital economies, while simultaneously exporting governance norms that legitimise state-centric AI deployment and limit the traction of liberal democratic frameworks around transparency, accountability, and civil liberties. This is not merely soft power posturing — if WAICO establishes technical standards, certification processes, or data-sharing agreements, it creates structural dependencies that are difficult to reverse. The institutional form matters: it gives Beijing a formal venue to aggregate the preferences of Global South AI adopters against Western-led norms.

Zhipu AI's completion of a 1-gigawatt data centre powered entirely by domestic chips — paired with Beijing's announced addition of 50,000 petaflops of intelligent computing capacity in H2 2026, pushing the capital's total above 130,000 petaflops — marks a transition from aspirational compute sovereignty policy to operational infrastructure, as South China Morning Post and South China Morning Post report. Zhipu's shares surged 37% on the news before partially retracing, reflecting market recognition that domestically-sourced compute at this scale begins to reduce the operational leverage that US chip controls were designed to maintain. Critically, a data centre of this size trained on domestic chips provides proof-of-concept that Chinese labs can sustain frontier training runs without access to Nvidia hardware.

The talent pipeline reinforces the infrastructure investment. Rest of World documents Chinese AI firms recruiting elite engineers directly from high school through camps, research programs, and guaranteed job pipelines — a structural response to the talent bottleneck that, if sustained over a five-to-ten year horizon, reduces China's dependence on overseas-trained researchers. Combined with Hugging Face's deployment of Zhipu's GLM 5.2 model to contain an autonomous OpenAI cyberattack on its infrastructure — an episode that simultaneously demonstrated Chinese model capability and raised questions about AI offensive risk — Chinese AI labs are accumulating credibility across both commercial and security-relevant domains.

The 18-day global shutdown of Anthropic's Fable 5 and Mythos 5 models in June — ordered by the Commerce Department and resulting in Anthropic cutting off all foreign users to achieve compliance — is the sharpest illustration yet that US AI export controls lack the precision to restrict adversaries without imposing significant costs on allies, partners, and domestic industry, as War on the Rocks analyses in detail. The episode reveals a structural gap: the US has broad shutdown authorities but no intermediate framework for granular restriction — licensing architectures, trusted user networks, or allied carve-outs — that would allow targeted controls without collateral diplomatic and commercial damage. Fable 5 was restored on June 30 after 18 days, but the episode will have signalled to allied governments that US AI infrastructure carries sovereign reliability risk.

This reliability risk is directly exploitable by China's governance offensive. Every time a US model goes dark due to export control action, it strengthens the case that Chinese open-weight models — which, once released, cannot be recalled — offer more predictable access. The strategic irony is that US controls may be accelerating adoption of Chinese alternatives in the very markets Washington is trying to preserve influence over, particularly in Southeast Asia and among Global South governments building out national AI infrastructure.