On November 17, the website of China's Ministry of Industry and Information Technology published a signed article by Yang Huayong.
The byline lists four titles: academician of the Chinese Academy of Engineering, chairman of the International Intelligent Manufacturing Alliance, deputy director of the National Intelligent Manufacturing Expert Committee, and dean of Zhejiang University's engineering faculty.
Three blockers
The article breaks the problem of AI in manufacturing into three: knowledge and data do not connect, decisions and execution do not join, and research and application do not loop.
Each comes with a concrete scene. Industrial knowledge has stayed unstructured — process experience passes between master and apprentice, design rules sit scattered in drawings and documents, equipment behaviour hides inside expert judgement. Machines can reach the data but cannot read the engineering meaning.
The second is blunter. AI produces a judgement and the line cannot take it, because the decision system and the execution system are two separate things. Decisions live in software, execution lives in machines, and without a layer of translation in between the instruction never lands.
The third is a broken loop: a model validated in a lab degrades into a one-off project on the floor, and every change has to start over, so the cost of use rises.
Two ends, one chain
The article offers a pair: one end grips technology supply to industrialise intelligence; the other grips application to make industry intelligent.
The policy behind that pair is the joint implementation opinion on AI plus manufacturing, issued by the ministry and seven other departments.
Build the technology on one side, put it to work on the other, and the chain in the middle is the industrial ecosystem. The phrase the academician uses is two-way empowerment.
Why him
The hardest part of any industrial policy is not in the document but on the floor. However detailed the eight-department opinion is, someone has to tell an enterprise which section of its line can take AI first.
Yang sits in that position. The academy gives technical authority, the expert committee gives standards, the international alliance gives a cross-country reference point.
Three titles stacked together carry weight, and are harder to dismiss as a vendor pitch.
His formulation in the article: AI has to move from an external tool to an internal capability. In a factory that means two things — data must reach the model, and the model must reach the process.
Three blockers, three businesses
Translated into business, the list reads clearly.
The first maps to data governance and industrial knowledge bases. Whoever turns a veteran worker's experience into a machine-readable format gets the order first.
The second maps to edge control and real-time systems. However accurate the model, if the instruction cannot reach a PLC, the effect is zero.
The third maps to delivery and operations. Go-live is the start; continuous iteration is the reason a customer renews.
On the same day, the ministry's notice on AI service providers stressed delivery as well. Two documents, one conclusion: this round is not a contest over parameter counts.
For a manufacturing enterprise, the three blockers also work as a checklist. Is the data structured, are the machines open, and who maintains the model after go-live? The question you cannot answer is where the money should go.
What a policy article can and cannot do
An academician's signed article is not a regulation. It does not set a target or a deadline. What it does is legitimise a way of framing the problem, and framing decides where pilot money goes.
The three blockers, taken seriously, imply three kinds of pilot: knowledge bases for industries with heavy tacit experience, real-time control stacks for lines with legacy equipment, and maintenance contracts that outlast the model's first version.
That is a different portfolio from the one most local governments have funded, which tends to buy GPUs and demos. The article's contribution is to point at the unglamorous middle.
There is a measurement problem underneath. Factories can count robots installed; they cannot easily count decisions improved. Until the second number exists, budgets will keep flowing to visible hardware rather than to the connective layer the article describes.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*