Geopolitics & Sovereign Positioning

Meituan's release of LongCat-2.0, a 1.6 trillion-parameter model trained entirely on Chinese-made chips, is the most significant concrete challenge to date to the strategic logic underpinning US semiconductor export controls. The controls were premised on the assumption that denying access to NVIDIA H100/H800-class hardware would constrain Chinese frontier model development — particularly the computationally intensive training phase, as distinct from inference. LongCat-2.0, benchmarking on par with DeepSeek's latest flagship, directly contradicts that assumption. South China Morning Post reports the model has been open-sourced, which accelerates the propagation of these efficiency techniques across the Chinese AI ecosystem.

This development sits alongside a broader structural consolidation in China's semiconductor tooling sector. State-backed Piotech is acquiring Wuxi Shangji Semiconductor with backing from China's National Integrated Circuit Industry Investment Fund, as South China Morning Post reports, with explicit intent to build national champions capable of producing equipment that substitutes for restricted Western tooling. Simultaneously, ByteDance is targeting mass production of a proprietary CPU by late 2027 to power its own AI infrastructure, reducing dependence on foreign chip architectures at the data centre level. And Chinese chipmakers including Shenzhen-based Basic Semiconductor are moving aggressively into silicon carbide power semiconductors to address the energy efficiency bottleneck in AI data centres — a domain where US export controls do not yet apply with the same force. These are not isolated moves; they constitute a vertically integrated strategy to close every layer of the AI hardware stack.

China's State Administration for Market Regulation has released a national standard for AI agent interconnection, establishing a unified identity management framework — effectively 'digital ID cards' — for all autonomous AI agents operating within Chinese systems. As South China Morning Post reports, the standard creates a closed-loop system designed to make AI agents traceable, auditable, and controllable at a national level. This is enacted policy with administrative enforcement mechanisms, not a voluntary industry framework.

The geopolitical significance extends beyond domestic governance. China's pattern with technical standards — from 5G to facial recognition — has been to establish national frameworks, demonstrate their operability at scale, and then promote them through multilateral bodies and bilateral infrastructure deals as the default architecture for partner states. A Chinese-origin AI agent identity standard adopted across Belt and Road digital infrastructure would give Beijing persistent visibility and potential leverage over agentic AI systems operating in those partner states. Western countries lack a comparable sovereign-level framework for AI agent governance, which creates both a regulatory gap and a standards competition that is not yet adequately joined.

The Atlantic Council's framing of Middle Eastern AI strategy as 'new non-alignment' captures something structurally important: Gulf states are not choosing between US and Chinese AI ecosystems but are deliberately maintaining relationships with both as a source of leverage over each. Atlantic Council notes that these countries are moving beyond the role of wealthy customer toward actively shaping AI development through data centre co-investment, sovereign AI fund participation, and bilateral technology transfer agreements. The region's combination of sovereign capital, geographic position between Eastern and Western supply chains, and low regulatory friction makes it uniquely positioned to extract technology commitments from both blocs.

This dynamic has concrete implications for US export control policy. The Commerce Department's AI diffusion rules — which restrict the export of advanced AI chips to third countries that might retransfer them to China — create direct friction with Gulf states that want to maintain economic relationships with Beijing. Saudi Arabia's and UAE's simultaneous pursuit of US AI partnerships and Chinese infrastructure investment is not hypocrisy; it is deliberate hedging that forces Washington to choose between tightening controls and losing influence. The outcome of that tension will shape whether the Middle East becomes part of a US-aligned AI sphere or a genuinely multipolar AI infrastructure zone.

Foreign Policy's assessment that Europe will never be an AI superpower reflects a maturation in the debate — the question is no longer whether Europe can catch the US and China at the frontier, but what degree of technological independence is achievable and strategically sufficient. Foreign Policy points to the Eurostack concept — building European alternatives across the AI stack from cloud infrastructure to models — as a realistic path to partial independence, even if full frontier parity is out of reach. The practical implication is that European AI sovereignty will be selective: deep enough to ensure critical government and defence functions are not reliant on US or Chinese infrastructure, but insufficient to sustain a globally competitive commercial AI ecosystem.

This assessment intersects with the CFR's analysis of the US regulatory framework for AI, where CFR President Michael Froman examines what the de facto US restrictions on Anthropic's most powerful models reveal about American AI governance coherence. CFR notes the contradiction in US policy between promoting AI leadership and applying export-style restrictions that constrain even allied access to frontier systems — a dynamic that simultaneously pushes European partners toward domestic alternatives and signals that US AI partnership comes with significant strings. For European strategic autonomy, this incoherence is a net positive: it provides justification for Eurostack investment that would otherwise face political headwinds from pro-Atlantic factions.

Defense One reports that an agentic AI targeting tool is being developed to give US commanders new strike options within seconds, compressing the kill chain in ways that fundamentally alter the tempo of decision-making in combat. Defense One notes that governance concerns about the authority and accountability of software agents in targeting decisions remain unresolved — a gap that is not incidental but structural, reflecting the broader absence of an agreed US government framework for autonomous decision support in lethal operations. This is a confirmed development program, not a proposal, though the governance framework around it remains undefined.