AI & Algorithm Governance

China's AI policy is still built for speed, and the latest evidence pushes governance further toward operational experimentation rather than static rule-making. The most concrete signals are not new algorithm-content controls but named participants in Hong Kong's GenA.I. Sandbox++ ¹, reconfirmed compute targets from Guizhou ^2,3, and a Shanghai strategic-plan anchor that makes AI promotion a dated, numeric commitment ⁴. The promotion-control balance still leans decisively toward deployment; governance is being used to build trust and ecosystem cohesion, not to throttle the roll-out.

Guizhou reconfirms it is the flagship compute pitch: total compute 176 EFLOPS (Guiyang-Gui'an 174), smart-computing share over 90% (over 98% in Guiyang-Gui'an), 50 key data centers (40 operating, 26 large in Guiyang-Gui'an), 2024 digital-economy value added above 250 billion yuan equaling 53.3% of GDP, and 2026 targets of 190 EFLOPS, over 300 billion yuan digital-intelligence industry, and over 100 billion yuan compute industry ^2,3,5. Cost and latency claims hold: deployment is about 30% cheaper than the coast, 6 ms to Chengdu-Chongqing, 10 ms to the Greater Bay Area, over 90% of customers come from outside Guizhou, and utilization exceeds 96%; the Guian supercomputing center rendered over 40% of the movie "Nezha 2" ³. Shanghai's 15th Five-Year plan adds an upper-end anchor: 2.1 trillion yuan in value added by 2030, with AI directed at large models and agents, scientific intelligence, embodied intelligence, and the intelligent computing industry system ⁴.

The governance track is becoming concrete and venue-specific. Hong Kong's sandbox is no longer a concept: 36 use cases from 30 financial institutions and 27 technology partners were selected from nearly 100 proposals, focused on agentic AI in customer onboarding, payments, insurance claims, and client interaction, including AI systems that dynamically monitor other AI applications' autonomous operations ¹. This directly answers the earlier question of whether governance would move beyond paper review. On the mainland, MIIT's nearly 200 key AI standards and city-level ethics-review pilots remain vague on criteria and cities ⁶, and no new algorithm-registration or large-model content rule appears in this batch. The tightest sectoral control remains the staged medical-AI pathway. The Hong Kong model is explicitly designed to reinforce the SAR's financial-center status, so its migration to mainland practice is an open question rather than a done deal.

Internationally, China is selling itself as both an AI adopter and a governance standard-setter. Global Times frames Xi Jinping's 2026 WAIC keynote as turning global AI governance principles into "deeper, more concrete action" ⁷, and Shanghai's party committee study session repeats the "innovation hub for AI development and governance" formula ⁸. Macao's Third Five-Year Plan security-sector exchange adds network, data, and AI to the emerging-threat portfolio ⁹, indicating that AI governance is being embedded in strategic-security planning across the SARs. The international layer still supplies the diplomatic material for China's claim to a leading role in global AI governance.

The build-out is reaching factories and skills. PCB leader Hudian reported H1 revenue up 61.17% and net profit up 73.72%, with 32-layer-plus PCB revenue up 190.83%, as AI capex drives volume and price across the chain ¹⁰. Fujian deployed a grinding robot for 30-ton power-generation valve parts with a 50% efficiency gain ¹¹; Jiangsu's 264 craft academies run "AI+技能夜校" (AI skills evening school) at a Shanghai Zhenhua Heavy Industry subsidiary in Nantong ¹²; and Tianjin's Haihe Craftsman Cup tied 90 of its 96 events to modern industry, including AI training, data annotation, and robot welding ¹³. The deployment machine is transmitting through procurement, skills, and factory floors, not just data centers.

Until MIIT's standards harden into binding market-access requirements, or a mainland agentic-AI sandbox appears, the promotion-control balance will keep leaning toward deployment. The payoff is measured in compute orders, PCB exports, and skills pipelines; the risk is that governance becomes a patchwork of experiments without a single binding rule. Watch whether the sandbox's dynamic AI-monitoring-AI logic migrates north, and whether the near-200 MIIT standards become more than advisory.