On September 23, 2026, Made in China Journal published an analysis by RMIT University media communications professor Haiqing Yu on China's artificial intelligence development model, which the author reduces to the formula “the state builds the stage — enterprises perform” (政府搭台，企业唱戏): infrastructural statecraft plus managed pluralism. By this logic, Beijing is building AI not as a market of individual products, but as state infrastructure on five pillars — energy, computing, chips, talent, and regulation. For the reader, this is a ready-made framework through which news about Chinese models and chips stops being a set of scattered releases and becomes a system in which rules are set by government procurement, licenses, and standards.

image

What happened

In the article “China's Approach to AI: Infrastructural Statecraft, Managed Pluralism,” Haiqing Yu examines the structure of China's AI system through five pillars, naming specific programs, companies, and dates. Energy: China's energy capacity is twice that of the United States, and data centers are planned to be built jointly with nuclear power plants and “green” generation. Computing: the 2022 “East Data, West Computing” program is moving computing hubs to Guizhou, Gansu, Ningxia, and Inner Mongolia via ultra-high-voltage transmission lines. Chips: government procurement protects a niche for Huawei Ascend, Cambricon, Moore Threads, and Biren, while Huawei is building a full Ascend — ModelArts — CANN stack as an alternative to CUDA. Talent: in 2025, half of the world's top AI researchers are from China, and AI is taught in primary and secondary schools. Regulation: instead of a single comprehensive law, a series of incremental acts is in effect, in which licensing serves as a lever for selecting models.

Context

To understand the meaning of this construction, the author refers to the “war of a hundred models” (百模大战), which has been ongoing since 2023: dozens of Chinese LLMs compete on government subsidies, some releases are an artifact of subsidies, and the selection of champions occurs through government procurement and licenses, not only through public benchmarks. The stake is fixed in the 15th Five-Year Plan (2026–2030) with a course on “new quality productive forces.” A separate background is U.S. export restrictions on chips: they are precisely what turn the construction of a homegrown Huawei stack from a question of prestige into a question of industry survival. The resulting model, Yu describes as “managed pluralism”: enterprises compete with each other, but on a stage built and continued to be controlled by the state, and some Western claims about Chinese “frivolous innovation” ignore precisely this systemic bet on energy, computing, and talent.

Why this matters for the industry

For the industry, the main takeaway is the emergence of a second parallel computing stack. Huawei Ascend — ModelArts — CANN is being built as an alternative to CUDA with guaranteed demand through government procurement, so when calculating TCO and choosing a serving stack, it makes sense to account for the existence of this alternative, and when connecting Chinese models via API — to check licensing terms. Energy surplus is converted into cheap “AI factories” for training and inference: if Chinese models are trained and served on a subsidized computing layer, pressure on model API pricing will increase, and founders should recheck the unit economics of their products. Licensing and government procurement serve as a mechanism for selecting national champions after the consolidation of the “war of a hundred models,” so the moat shifts from the quality of an individual model to access to the stage. Indirectly, this affects the effect of U.S. export restrictions — the replacement of Nvidia with the rising Huawei/CANN stack — and global rare earth metal supply chains.

Why this matters for users

The article itself is a free open analysis with specifics: the names of programs, companies, and dates allow assembling a coherent map from the flow of news about Chinese models and chips instead of a set of headlines. The main skill it provides is understanding that individual releases of Chinese models are a “performance” on a stage prepared by the state, where rules are set by government procurement, licenses, and standards, and it is precisely for this reason that the hype around the next release should be evaluated more critically. A practical bonus for those who use Chinese models: after the consolidation of the “war of a hundred models,” the list of vendors and APIs will become shorter and more predictable, and within a six-month horizon, independent measurements of the Huawei Ascend/CANN stack should be expected, which will show how real the alternative to CUDA is in practice.

What is still unknown / limitations

Key quantitative claims in the article are given without methodology in the text itself. The claim that China's energy capacity is twice that of the United States is the author's estimate without a specified calculation method. The thesis that in 2025 half of the world's top AI researchers are from China needs to be checked against independent data on researcher mobility before being used as a fact. The central technical claim — that Huawei is building a full Ascend — ModelArts — CANN stack as an alternative to CUDA — is in principle verifiable, but not verified in the article: it contains no data on the maturity of the CANN software layer (kernels, compilers, framework compatibility), energy efficiency, or independent measurements on large model training, so formulations about migrating workloads from CUDA as a realistic scenario do not follow from the source. Finally, this is a political-analytical review, not a release: it contains no new benchmarks, architectures, or APIs, and scenarios for the consolidation of the “war of a hundred models” and possible bottlenecks in CANN maturity and data quality are interpretations that depend on how the 15th Five-Year Plan will be financed.

Sources

Author

Look at AI, editorial team