AI & Algorithm Governance

China's AI policy remains built for speed, but the new policy batch shifts the governance track from advisory paper to operational infrastructure. The strongest evidence is MIIT's August 28 notice naming 17 provinces as national regional nodes for computing-power interconnection ¹ — a formal "1+M+N" architecture (one national node, M regional nodes, N industry nodes) with "unified identification, unified standards, unified rules" that makes cross-region compute exchange a governed market rather than a series of bilateral deals. For an AI strategy whose bottleneck is compute, this is a governance mechanism with teeth, and it changes the earlier picture of a patchwork of experiments.

Chongqing's inclusion is the most concrete case: the city had already written "national computing-power interconnection Chongqing system" into its 15th Five-Year AI development plan, and is building a "疆算入渝" (Xinjiang compute into Chongqing) data highway to share Chengdu-Chongqing data centers ¹. The node system effectively turns MIIT into the clearinghouse for how provincial compute pools interconnect, standardize, and price their services. This is control through interoperability rules — less visible than algorithm registration but more binding on the build-out.

Guizhou, the flagship compute province, now adds a new metric: 81.2% of its intelligent computing is domestically produced, framed as "owning your own compute power" ². Total capacity remains 176 EFLOPS with 169.94 EFLOPS in use (96% utilization), over 90% smart-computing share, 50 key data centers, and a 400G backbone reaching 44 cities directly ^2,3,4. Cost and latency claims hold — roughly 30% cheaper than coastal deployment, 6 ms to Chengdu-Chongqing and 10 ms to the Greater Bay Area, with over 90% of customers from outside the province ⁴. The domestic-rate number suggests the next promotional line is not just scale but sovereignty.

Meanwhile, "AI+" industrial promotion is spreading through provincial machinery. Liaoning will host the 2026 Global Industrial Internet Conference in Shenyang (September 8-10) under the theme "智改数转AI兴工" (intelligent transformation, digital transition, AI-powered industry), releasing an "AI+制造" development report and focusing on industrial large models, embodied intelligence, and data-factor marketization ⁵. Jiangsu's coastal economy push instructs ports to lay out compute infrastructure scientifically, deepen compute-electricity coordination, and seize AI, robotics, and intelligent-computing tracks ⁶. Shandong's party secretary inspected an AI machine-tool project and called for AI empowerment of the service sector ⁷. The deployment machine is now embedded in provincial industrial policy, not just data-center plans.

On the control side, the new documents show no new algorithm-registration or large-model content rules. The tightest sectoral restraint remains the staged medical-AI pathway; MIIT's roughly 200 key AI standards are still not binding market-access requirements. The governance energy is going into infrastructure interoperability and trust-building — the Hong Kong GenA.I. Sandbox++ logic remains confined to the SAR for now.

Internationally, China continues to market itself as both AI adopter and governance standard-setter. Vice Foreign Minister Ma Zhaoxu's APEC SOM3 remarks cite planning for an APEC digital and AI cooperation vision and AI-era tourism facilitation under China's host year ⁸, while Global Times frames Xi's 2026 WAIC keynote as moving global AI governance principles toward "deeper, more concrete action" ⁹. Shanghai's party committee study session repeats the formula of building an "innovation hub for AI development and governance" ¹⁰, pairing the diplomatic claim with a domestic anchor.

Implications: watch whether the compute-interconnection "unified rules" become de facto national exchange standards, and whether Guizhou's 81.2% domestic rate morphs into procurement guidelines for state-funded compute. The balance still leans toward deployment, but the control vector is now infrastructure — standards for how compute flows, rather than content rules for what models say.