AI & Algorithm Governance

China's AI policy remains firmly promotion-first: the state is now injecting capital directly into generative-model companies and commercial metrics are improving, while the governance track continues to move through courts, procurement and green standards rather than new content rules. The strongest new evidence is the National AI Fund (国家人工智能基金) taking a stake in Beijing Kling, Kuaishou's video-generation model subsidiary, with a 1.4 billion yuan cash injection alongside Chia Tai Robotics ¹, and Zhipu's first-half results showing revenue of 954 million yuan, up 399.7% year-on-year, with API revenue now 86.5% of the mix ².

The compute-interconnection architecture set out on August 28 remains the backbone. MIIT formally named 17 provinces as national regional nodes under a '1+M+N' system with unified identifiers, standards and rules, and Chongqing's selection is tied to its 15th Five-Year AI plan and the '疆算入渝' data corridor ³. Guizhou's numbers demonstrate the scale: 176 EFLOPS, 169.94 in use, 38万 racks, 40万+ servers, 50 key data centers, and a 400G backbone reaching 44 cities ⁴.

The new batch shows the promotion track moving beyond infrastructure to the model layer. Kling's accession to the National AI Fund is notable because it applies state capital to a consumer-facing generative video model, not just a national champion ¹. iFlytek will open-source two edge-side models on September 1 and release Spark X2.5 (293B) on September 7, with a new all-domestic-compute flagship at its 1024 developer conference ⁵. Hubei's meeting chaired by executive vice governor Zhang Wenbing continues to push for AI companies to go public and build an IPO reserve pipeline ⁶.

Zhipu's earnings call adds a commercial chapter to the story. Revenue growth was 399.7%, but the company stresses mix shift: API platform revenue reached 825 million yuan, up more than 27 times year-on-year, accounting for 86.5% of revenue, while localization deployment fell from 84.8% to 13.5% ². API gross margin turned positive at 24.6%. Founder Tang Jie defended the 'only post-training' critique by arguing scaling is multi-dimensional, and confirmed training data of 30-50 trillion tokens and 100,000-card domestic chip clusters for low-cost inference. This is a deliberate counter to the narrative that Chinese models trail on scale.

No new central algorithm-registration or large-model content rules appeared in this batch. The governance signal is instead indirect: Guizhou's 'green renewal' trajectory, under the 2024 State Council equipment-renewal plan and the Data Center Green Low-Carbon Development Action Plan, adds recycling and circular use as a second filter on state-linked compute ⁴; Guizhou High Court's 16 measures embed AI-model and algorithm disputes into judicial rules ⁷. The Macao government's photography contest explicitly rejecting AI-generated or computer-generated images is a small reminder that cultural-heritage gatekeeping remains a control lever ⁸.

Internationally, China continues to default to consensus framing. Xi's visit to Kyrgyzstan produced a joint declaration and permanent treaty that 'strengthen cooperation' in AI among other fields ⁹; the SCO summit speech frames regional governance around 'higher-quality development' without unveiling AI-specific institutions ¹⁰. Hong Kong and Macao are being used as auxiliary promotion nodes: China Chengxin Green Finance and Miaoying Technology merged in Hong Kong to export AI-integrated ESG fintech ¹¹, and Macao's labor bureau is sending 30 youths to visit AI and semiconductor firms in the GBA ¹².

The balance of evidence still says promotion dominates, but the control track is hardening underneath: green standards, court rules and procurement language will shape which compute and models actually scale. Watch whether the National AI Fund follows Kling with other model-layer investments, whether Zhipu's API mix becomes industry standard, and whether the SPC nationalizes the Guizhou judicial approach.