AI & Algorithm Governance

China's AI policy remains built for speed, but governance is now becoming a concrete operational layer rather than a flat control track. In the same week, MIIT Vice Minister Xin Guobin (辛国斌) said China has formulated nearly 200 key AI standards and is organizing multiple cities to carry out AI ethics review and service practices ¹, while Hong Kong's financial regulators unveiled the first batch of their GenA.I. Sandbox++ with 36 use cases for agentic AI from 30 financial institutions and 27 technology partners ². Both are framed as enabling rather than constraining: standards for supply-chain coherence, sandboxes for responsible experimentation.

The strongest promotional signal is now a formal, quantified plan. Shanghai's Strategic Emerging Industries "15th Five-Year" Plan sets a 2030 target of 2.1 trillion yuan in value added, with strategic emerging industry output above 50% of total industrial output and the three leading industries — integrated circuits, biomedicine, and AI — growing at over 10% annually. AI is explicitly directed toward large models and agents, scientific intelligence, embodied intelligence, and the intelligent computing industry system ³. That gives the promotion machine a dated, numeric anchor for upcoming investment and procurement.

Guizhou's compute pitch is reconfirmed with identical figures: total computing power of 176 EFLOPS (Guiyang-Gui'an 174 EFLOPS), smart-computing share above 90% (over 98% in Guiyang-Gui'an), 50 key data centers (40 operating), and 26 large data centers in Guiyang-Gui'an including Huawei, Tencent, and the three telecom operators. The province reports 2024 digital-economy value added above 250 billion yuan, 53.3% of regional GDP, and 2026 targets of 190 EFLOPS and a digital-intelligence industry above 300 billion yuan ^4,5,6. The sales pitch holds: deployment costs about 30% less than the coast, latency to Chengdu-Chongqing is 6 ms and to the Greater Bay Area 10 ms, more than 90% of customers come from outside Guizhou, and utilization exceeds 96% ⁵. The Guian supercomputing center rendered over 40% of the movie "Nezha 2" ⁴.

Governance is shifting from static rule-making to experimentation. MIIT's ethics-review pilots and the standards count remain vague — no cities, no published criteria ¹ — but Hong Kong's GenA.I. Sandbox++ is concrete: participating firms will test full customer onboarding, payments, insurance claims, and client-interaction flows, and will examine how AI can dynamically monitor other AI applications' autonomous operations ². This directly answers a prior open question about whether governance would move beyond paper review. The venue matters: the sandbox is outside mainland regulatory space, and is explicitly designed to reinforce Hong Kong's global financial-center position, not to throttle deployment.

The international layer continues to sell China as both an AI adopter and a governance standard-setter. An English-language Global Times piece frames Xi Jinping's 2026 WAIC keynote as turning global AI governance principles into "deeper, more concrete action" ⁷. The domestic standards push supplies the material for that diplomacy. Shanghai's party committee study session repeats the "innovation hub for AI development and governance" formula, with autonomy, application traction, ecosystem collaboration, open win-win, and safe-and-controllable development as the pillars ⁸.

The real-economy transmission is uninterrupted. PCB leader Hudian (沪电股份) reported H1 revenue of RMB 13.689 billion, up 61.17%, and net profit of RMB 2.923 billion, up 73.72%, with 32-layer-plus PCB revenue surging 190.83% ⁹. Jiangsu's 264 craft academies are running "AI+技能夜校" at a Shanghai Zhenhua Heavy Industry subsidiary in Nantong ¹⁰, while Tianjin's Haihe Craftsman Cup added AI training, data annotation, and robot welding among 90 of its 96 events tied to modern industry ¹¹. Fujian has deployed grinding robots for 30-ton power-generation valve parts, lifting efficiency by 50% ¹². The build-out is reaching skills and factory floors, not just data centers.

The key limitation is that no new algorithm-registration or large-model content rule appears in this batch; the tightest sectoral control remains the staged medical-AI pathway noted previously. The new question is whether the Hong Kong agentic-AI sandbox model — dynamic monitoring of AI by AI — migrates to the mainland, and whether the nearly 200 MIIT standards harden into binding market-access requirements. Until that happens, the promotion-control balance still leans decisively toward deployment, with governance as a tool for cohesion and trust rather than a brake.