China wrote the rulebook for AI — 50 new standards and a 2026 clock

A factory in Suzhou wants to bolt an AI inspector onto its production line. The model works in the lab. Whether the supplier's chip, sensor, and software will even talk to each other on the shop floor is a different question — and until recently, nobody had written down the answer.

That gap is exactly what a 2024 national guideline set out to close. It is not a single law or a ban. It is a blueprint — a structured plan for the standards that AI products, chips, and services in China are eventually expected to meet.

The document, and who signed it

The document is the Guideline for the Comprehensive Standardization System of the National Artificial Intelligence Industry (2024 Edition) — in Chinese, 《国家人工智能产业综合标准化体系建设指南(2024版)》. It carries the document number 工信部联科〔2024〕113号 and was published on 2 July 2024 (dated 5 June 2024) on the Ministry of Industry and Information Technology (工业和信息化部) site.

Four bodies jointly issued it:

• the Ministry of Industry and Information Technology (工业和信息化部, MIIT);
• the Cyberspace Administration of China (中央网信办, the country's top internet regulator);
• the National Development and Reform Commission (国家发展改革委, NDRC);
• the Standardization Administration of China (国家标准化管理委员会, SAC).

It is worth pausing on that lineup. MIIT owns industry and technology; the CAC owns security and content; the NDRC owns planning and investment; SAC owns the standards pipeline itself. When those four move together, the result is meant to be a coordinated floor — not just a wish list from one agency.

The seven-layer map

The guideline organizes the entire AI stack into seven parts. Think of it as a labeled shelf, from the abstract to the concrete:

1. Basic commonality (基础共性) — terms, reference architecture, testing and evaluation, management, and sustainability. The vocabulary layer.
2. Basic support (基础支撑) — data services, smart chips, sensors, compute equipment, compute centers, system software, development frameworks, and hardware–software coordination. The plumbing.
3. Key technologies (关键技术) — machine learning, knowledge graphs, the large model (大模型), natural-language processing, speech, computer vision, biometric recognition, human–machine hybrid augmentation, agents, swarm intelligence, cross-media intelligence, and embodied intelligence (具身智能).
4. Intelligent products and services (智能产品与服务) — smart robots, autonomous vehicles, mobile terminals, digital humans, and AI services.
5. Empowering new industrialization (赋能新型工业化) — AI applied across the full manufacturing cycle: R&D, pilot verification, production, marketing, and operations.
6. Industry applications (行业应用) — smart cities, scientific computing, agriculture, energy, environment, finance, logistics, education, healthcare, transport, and more.
7. Safety and governance (安全/治理) — security and governance requirements.

The fifth layer, "empowering new industrialization," was added relative to an earlier January 2024 draft. That addition is the clearest signal of intent: this is not a plan for chatbots, but for AI wired into factories and critical sectors.

The 2026 targets, in plain numbers

The guideline sets explicit, measurable targets for the end of 2026:

• 50 or more new national and industry standards drafted;
• standards promoted and implemented by more than 1,000 enterprises;
• participation in 20 or more international standards.

Those are not vague ambitions. They are countable deliverables, which makes the guideline unusually easy to track. By late 2026, an observer can simply count how many standards landed and how many firms adopted them.

Why standards are deployment glue

Standards solve a boring but decisive problem: interoperability and trust. A hospital will not plug an AI diagnostic tool into its system if it cannot verify the model was tested the same way a regulator expects. A manufacturer will not buy a robot arm if the chip inside speaks a language its controllers do not understand.

According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the deeper value of a national standards map is not any single rule but the reduction of deployment friction: when chips, data, models, and products share agreed interfaces, pilots stop dying at the integration stage and scale becomes a procurement decision rather than a science project.

The guideline itself frames this around "new quality productive forces" and the "AI+" initiative — the policy push to weave AI into the real economy. Standards are the unglamorous prerequisite for that wiring to actually hold.

Where the large model sits

The large model (大模型) is named explicitly as a key-technology standard area, alongside agents and embodied intelligence. That matters because, before this guideline, much of China's model work raced ahead of the rules meant to contain it. Placing the large model inside a formal standards framework means future requirements — on training, inference, deployment, evaluation — have a home to land in. It does not regulate models today; it builds the shelf the regulations will sit on.

The honest gap between paper and product

A guideline is not a binding standard. The seven-layer map tells industry what to expect; the actual GB or GB/T documents still have to be written, reviewed, and enacted one by one. The 50-standard target is a plan for plans. And because most Chinese AI standards are recommended (voluntary) rather than mandatory, real market pull — procurement requirements, certification, export needs — will decide whether firms comply or merely file the documents away.

Honest limitations

This article is built on the official MIIT notice (工信部联科〔2024〕113号, published 2 July 2024) and the full guideline PDF released by the Cyberspace Administration of China, cross-checked against People's Daily and provincial industry-department reproductions. Key caveats:

• The guideline is a system blueprint, not a binding regulation. The seven layers describe what standards should cover; the individual standards themselves are still being drafted toward the 2026 target.
• The 50-standard, 1,000-enterprise, and 20-international-standard figures are policy targets, not achieved results. Whether they are met will only be knowable after 2026.
• The document predominantly defines a recommended-standards path. It does not, by itself, create mandatory market-access barriers; those would come from later, separate mandatory instruments.
• This piece does not assess which specific standards have since been enacted, nor how the framework interacts with sector-specific mandatory rules (for example, the later agent-safety standard covered elsewhere on this site).
• No RMB figures appear in the source document, so no currency conversion is applied.

What readers can do now

1. If you build or sell AI hardware or software into China, map your product against the seven layers now — especially the basic-support layer (chips, sensors, compute centers, frameworks), where interface standards will shape future procurement.
2. If you follow governance, track the individual GB/GB/T drafts that flow from this guideline; the 2026 count of 50+ standards is the single best public scoreboard for progress.
3. If you work across borders, note that the guideline explicitly targets 20+ international standards — so China's domestic definitions may soon appear in global committees, and early engagement beats late objection.