AI & Algorithm Governance

The standing judgment holds and this batch does not move it: China's AI policy remains promotion-first. No new algorithm filing rule, large-model security review or content directive appears in these documents. The only item not already in the record is a US Federal Register notice partially repealing carbon pollution standards for fossil-fuel power plants ¹ — it says nothing about China's AI stack, and its only use here is as background to a divergence in how the two systems treat power for compute: Beijing couples compute to green power by design, while this filing loosens emissions rules on the generation side. Nothing in the batch changes the promotion-versus-control balance.

The anchor is still the Sept 11 State Council executive meeting chaired by Li Qiang, which made the computing-power network (算力网) a standalone agenda item and called it "the foundational support for AI development." Its orders are industrial and energy policy, not censorship: reasonable layout and orderly, regulated build-out; a multi-tier networked compute system; key technology and equipment R&D; the pairing of compute-power coordination and compute-network integration (算电协同、算网融合); compute-resource monitoring and dispatch; backbone optical-fibre construction; alignment of compute and power facility planning; and accelerated green-power direct connection and source-grid-load-storage projects. Market mechanisms are to do the matching work for supply and demand ². This is grid-and-cluster planning wearing an AI label.

Provincial evidence shows the template running. Guizhou is the only first-batch regional node whose consortium includes a grid company (Guizhou Power Grid, alongside the three telecom branches), producing a "three-network hub" tying the national compute network, the new-generation communications network and the power grid together ³. Heilongjiang, another of the 17 approved regional nodes, anchors on China Mobile's Harbin 10,000-card cluster and reports 430 provincial digital workshops and smart factories by end-2025 — up 212 since 2022, a 25.4% annual rise — with the share of enterprises digitising key business processes climbing from 34.6% to 51.8% ^4,5. Neither provincial write-up mentions regulatory compliance as a bottleneck.

The newest institutional layer is data annotation, and it is being staffed and counted like a manufacturing sector. The National Data Administration's second batch of national data-annotation innovation pioneer cities named Shenyang (industrial manufacturing) and Dalian (medical health), making Liaoning a "two cities, two domains" pilot province. Liaoning's 2026–2028 action plan reports 27,000 TB of annotated data, 334 high-quality datasets, more than 120 firms, over 20,000 workers and a 6.2-billion-yuan industry; Dalian's data-annotation park has drawn 19 firms including 金慧科技 (Jinhui Technology), 医通康合 (Yitong Kanghe) and 东软云科技 (Neusoft Cloud), sitting on a software cluster of 3,000-plus firms and 200,000-plus software workers, plus nine tertiary hospitals and over 53 million annual patient visits ⁶.

Internationally, Beijing keeps converting chairmanships into named deliverables. At the BRICS summit in New Delhi (Sept 12–13), Xi announced China will lead construction of a BRICS AI open-source zone, a digital-industry cloud platform and a youth sci-tech exchange plan aimed at helping Global South countries "benefit equally," and the MFA confirmed China takes the BRICS chair next year, hosting the 19th summit ⁷. The APEC energy ministers' meeting in Beijing put "AI and energy mutual empowerment" among three priority directions under China's chairmanship ⁸, and Assistant Foreign Minister Liu Bin opened the fourth AI capacity-building seminar for more than 30 countries at Peking University, framing capability-building as governance-building too ⁹. The Sept 15 MFA readout adds the services-trade channel — CIFTIS drew 1,830 offline and 5,300 online exhibitors and 1,200 outcomes across digital-medical, smart-education and related fields ¹⁰.

Diffusion keeps widening beneath the diplomacy. MIIT's software plan still targets high-standard coding tools reaching more than 20,000 major enterprises and 100 benchmark agent-based applications by 2028; Unisound's "token export" served Ethiopian cross-border e-commerce with over 10 billion tokens in week one ¹¹. Tokens are becoming ordinary retail: Jingu Yuan, a dumpling restaurant outside Beijing University of Posts and Telecommunications' south gate whose owner is a 2010 BUPT telecoms graduate, put a 10-yuan compute card on every table in July, after launching "Jingu Yuan Dumpling Restaurant · SKILL" in April — billed as the first restaurant to open its queue-numbering interface to customers' personal AI agents; its 2025 winter-solstice online queue passed 15,000 tables ¹². Skills are wired in through Macao's Huawei-backed HCIA-AI training and a roughly 500-person AI-and-brain-science symposium, and Liaoning's vocational competition added AI, low-altitude economy and industrial-internet tracks across 88 events with 2,100-plus contestants ^13,14,15.

The read for outside observers is unchanged: China is institutionalising AI's full stack — compute, power, data and skills — as planned infrastructure and turning multilateral platforms into delivery vehicles, while leaving domestic model rules untouched. Three things to watch: whether the State Council's compute-network instruction hardens into a named plan, budget line or agency; whether the BRICS AI open-source zone gets a launch date now that China holds the chair; and whether the data-annotation pilot model spreads nationally as a named program.