On November 18, Securities Daily carried a report whose first two paragraphs juxtaposed the two faces of Chinese manufacturing.
One is bright: In the World Economic Forum's Global Lighthouse Network list, as of June 2026, there were 238 benchmark manufacturing and supply chain sites, with China ranking first globally with 109. The Ministry of Industry and Information Technology recently also publicly announced 34 leading-level smart factory cultivation projects for 2026. The application penetration rate of key AI scenarios in manufacturing rose to 34.2%.
The other picture is dark: as of the first half of 2026, China had cumulatively built more than 56,000 basic-level smart factories and more than 9,000 advanced-level ones, with only just over 500 at the excellence level. Meanwhile, the Enterprise Digital Transformation Index Report (2025) shows that the digital transformation maturity index of small and medium-sized enterprises is only 28.6, compared with 35.4 nationwide, and 86.92% of them are still in the basic informatization application stage.
109
China's lighthouse factories, No. 1 in the world
More than 500
National Excellence-level smart factories account for less than 1% of Basic-level ones.
Put the two sets of numbers together, and the problem becomes obvious: lighthouse factories are being built one after another, so why is the workshop next door still the same as ever?
1. The four characters "buying intelligence" really hit home for many projects
Shi Youkang, chief expert at the China Academy of Information and Communications Technology, offered a judgment in the report that I think is the sentence in this article most worth copying down:
The biggest misconception in the intelligentization of manufacturing is "buying intelligence," that is, simply purchasing software and hardware to build a model workshop, divorced from on-site processes, data accumulation, and organizational workflows. Such demonstration projects, although visually impressive, are difficult to scale.
The term "bought-in intelligence" is very apt. Because it accurately describes the sequence of actions taken by most companies: first see what peers have deployed, then find suppliers and compare prices, then purchase, install, accept, take photos, and report. The only thing not addressed throughout the entire process is which judgment this equipment is actually meant to replace.
There is also a technical explanation in the report that many people gloss over when reading: the AI models of many benchmark factories are built on years or even more than a decade of data accumulation; small and medium-sized factories neither have data assets of comparable scale nor engineering teams capable of digesting complex algorithms.A production scheduling system that has been honed and matured on a hundred-billion-yuan-level production line often fails to adapt when transplanted to small and medium-sized workshops with drastic order fluctuations and a mix of old and new equipment.
The real reason transplantation fails is not that the algorithm isn't good enough, but that the inputs are different. What large factories feed the model is "this batch of material, this machine, this mold, these parameters, what result was produced last time"; on the same work order at a small or medium-sized factory, there may be only two reliable numbers: "quantity" and "delivery date". No matter how powerful the model, it cannot conjure causality that was never recorded.
II. Three Factories, Three Replication Methods
The three factories interviewed for the report are taking three different paths. Seeing their differences clearly is more useful than memorizing any one factory's equipment list.
Lianbao Technology: Replicating the "Solution"
Lianbao (Hefei) has 24 SMT production lines and 28 complete-unit assembly lines, with one laptop rolling off the line every 0.5 seconds on average; one out of every eight laptops worldwide comes from here. What it has done is not just make its own factory smarter, but export its entire smart manufacturing system—turning a single lighthouse into a contiguous cluster of smart factory buildings.
What is delivered externally is essentially a set of rules that others can follow: which process uses which parameters, under what conditions to stop, and to what extent inspection counts as passing. According to Song Tao, general manager of the enterprise industry vertical of Lenovo China's government and enterprise business, factory intelligence is not a product of equipment procurement, but a capability that takes root in the production site over the long term and grows out of frontline problems, production data, and employees' hands-on experience.
Sany Heavy Industry Beijing Pile Machinery Factory: Adapting to "Scenarios"
This factory, which began its transformation in 2019, is among the country's first batch of excellence-level smart factories. Heavy equipment manufacturing is a typical discrete manufacturing process: many varieties, small batch sizes, and large workpieces; the drill rod of a rotary drilling rig can be up to 27 meters long, and the power head can weigh up to 16 tons. Such a human-machine combination is difficult to directly copy from the automation solutions of electronics factories.
Their approach was to install 3D vision sensors on robotic arms in the welding process to achieve human-machine collaboration; at the same time, they introduced heavy-duty robots to take on high-intensity repetitive tasks such as grabbing, flipping, and tightening bolts. After the transformation, the number of frontline workers in the workshop decreased by more than half compared with before, labor productivity increased by 85%, and the production cycle was shortened by 77%.
Note the relationship among these three numbers: with fewer people, output actually increased, and the cycle became shorter. This only holds when "judgment is moved onto the equipment"—if whether a weld needs repair and to what extent still depends on human inspection, halving the workforce will only increase rework.
Gree Jinwan Factory: Forging the "tool for creating solutions"
On the 480-meter-long outdoor unit production line, 86 of 103 processes have achieved automation, with an overall automation rate exceeding 80%; the self-development rate of equipment is also 80%, and all more than 2,000 industrial robots in the factory are independently developed.
You Yicheng, head of Gree's intelligent manufacturing production line, said something very practical: the difficulty in building this factory does not lie in the automation of any single process, but in simultaneously scheduling and coordinating the massive number of smart devices and robots across hundreds of links on the production line.
This sentence points out where the value of the third path lies. Lianbao sells solutions, Sany changes scenarios, and Gree makes tools that can generate solutions. The advantage of tools is that they can be reused—switch to another product or another line, and the scheduling logic does not need to be rewritten.
III. What is copied is not the device, but the criterion
The three paths look different, but underneath they are the same thing: moving judgment from people into rules. The only difference is which layer it is moved to—moving it into deliverables is Lianbao, into equipment is Sany, and into tools is Gree.
I collectively call this type of rule "criteria". A criterion usually consists of three parts: what to do under what conditions, to what extent it counts as qualified, and what to do after it fails. Only when all three parts are clearly written can humans and machines share the same set of standards; if any part is missing, decisions can only be made by people on site ad hoc.
Take the most common example in manufacturing. The work instruction for a certain process states, 'welding current 180 to 220 A.' This is not a criterion; this is a range. A complete criterion must be traced down three levels: for plate thickness below 3 mm, should the lower limit or the upper limit be used; after how many seconds of continuous welding should the current be reduced; if porosity appears, should it be repaired by welding or scrapped. If these three levels have no answers, then no matter how many sensors are installed or how smart the model is, in the end you still have to call over an experienced master worker to take a look.
This is how "buying intelligence" fails: What you buy is actuators, not criteria. Actuators have no criteria and can only perform the simplest actions; slightly more complex judgments still fall back to people. So the more devices there are, the more people are needed to watch the devices, and headcount cannot be reduced.
IV. Where the criteria fall in Odoo
The good news is that these locations already exist in the standard ERP, so there is no need to develop a separate set of things just for AI. Taking Odoo 17 to 19 as an example (version 20 may be slightly adjusted; please verify before implementation), they are roughly the following five places:
| Criterion Type | Locations in Odoo | To what extent should it be written? |
|---|---|---|
| Working conditions | Descriptions and work hours of operations on the routing (mrp.routing.workcenter) | It’s not just "how long to do it"; you need to specify "under what conditions the parameters are changed and what they are changed to" |
| Acceptance determination | Quality inspection points and quality inspection records (quality.point / quality.check) | Upper and lower limits, sampling rules, and which action to take after a nonconforming determination |
| Capability Boundaries | Equipment ledger (maintenance.equipment) and work centers | Maximum tonnage, machinable dimensions, accuracy grade, current status |
| Anomaly Attribution | Work reporting loss reasons and scrap reasons | Constrain it to a finite enumeration; do not leave free-text boxes |
| Feasibility | Product attributes and variants (attributes and variant combinations of product.template) | Which parameter combinations can be done, which cannot, and if not, what conditions are required |
These five places share one thing in common: they are all filled in by people, not by devices. This is also why they are often left blank—equipment procurement has a budget, acceptance checks, and photos; filling in fields has no budget, no acceptance checks, and no one taking photos. But from a data perspective, it is precisely these five places that determine whether the collected data is meaningful.
Give a minimal example
On the work orders of an injection molding factory, the two fields "planned start time" and "planned end time" were always filled in, while actual start and actual end were never filled in. As a result, they could calculate the "planned cycle" but not the "actual cycle." After implementing MES, the equipment could capture startup and shutdown, but for the captured signals to be assigned to work orders, someone still had to match the work orders with the time periods—and the parts that couldn't be matched became an unexplainable black hole in the system.
The remedy was crude: have the person reporting work report the actual time once at the end—just one field. Three months later, they drew the cycle-time distribution of their own workshop for the first time and found that the median changeover time for a certain type of mold was twice the estimated value. This discovery did not cost a penny; it only cost the constraint of one field.
V. The first layer for SMEs is not sensors, but fields
The report mentions Morgan Stanley's forecast: from 2026 to 2035, China's Industry 5.0 transformation is expected to bring about RMB 80 trillion in cumulative incremental capital expenditure. This scale cannot be covered by fiscal spending alone; it requires industrial capital, financial capital, and government funds to work together. The report also mentions a change that is taking place: after the operational experience of lighthouse factories is packaged into subscription-based services, manufacturing enterprises' intelligentization investment is shifting from one-time capital expenditure to predictable operating expenditure, greatly lowering the decision-making threshold and trial-and-error costs for small and medium-sized enterprises.
The subscription model is indeed a good thing; it turns “can’t afford it” into “can afford it.” But it doesn’t solve another problem: being able to afford it doesn’t mean you can take it on. With an annual subscription for an intelligent production scheduling service, once it’s brought in, your process routes need to be accurate, your capacity data needs to be real, and your delivery commitments need to be binding. If all three of these things are estimates, the scheduling results produced by the system will only be more confidently wrong than manual scheduling.
So for most small and medium-sized enterprises, the first layer should not be sensors, nor models, but the fields on documents. The criterion is very simple:
- On this form, how many fields are "required to have a value in order to save"?
- Of these values, how many were selected, and how many were just filled in casually?
- Those things filled in casually—did anyone verify them afterward?
The answers to these three questions basically determine a company's data foundation. Values that are selected can be counted, while values filled in casually can only pad the numbers. What AI needs is something that can be counted.
A common sequencing mistake: first deploy device data collection, then add system fields. As a result, you collect a bunch of signals but cannot find the corresponding work orders, operations, or batches, and ultimately can only turn it into a dashboard hung on the shop floor. The dashboard isn't ugly, but it can't feed into costing, can't feed into traceability, and can't feed into any external report. The correct order is the reverse: first fill in the forms correctly, then decide what to collect.
VI. If You Only Do Three Things
1. Pick one of the most expensive processes and write all three sections of its criteria in full. What to do under what conditions, to what extent it counts as qualified, and what to do if it is not qualified. After writing it, show it to the workshop manager and ask him, "Is this the same as what you actually do?" Most of the time, you will find that what is written down differs from what is actually done in two or three places—those places are exactly what will later be handed over to AI, and they are also the places most prone to errors.
2. Change the reasons for anomalies from free text into a limited enumeration. Start with no more than five categories, run it for a month, and look at the distribution. If one category accounts for more than 70%, that means the classification is not granular enough; if all five categories are roughly even, that means there is a problem with the classification criteria. This costs nothing, but it determines whether you will be able to work out "where all the losses went" later on.
3. Before signing any smart factory or AI service contract, ask one question first: What data does this system require me to provide?If the other party answers, “We have a general model and don’t need your data,” be careful—a general model gives general advice, and when it reaches your shop floor, it still takes people to translate it. If the other party lists a data checklist, that checklist itself is a very good self-check list; filling in the gaps according to it is more cost-effective than buying according to it.
The 109 lighthouse factories show where the upper limit of Chinese manufacturing lies, while 86.92% shows where the starting point of most enterprises lies. The road in between has, over the past few years, been described as a road of equipment procurement, so many enterprises spent their money on the outermost layer.
But the answers given by LCFC, Sany, and Gree are actually the same: their real asset is not those robots, but the set of rules that tell the robots what to do. Rules can be copied, robots can be bought; if the order is reversed, the money is wasted.
AI grows on the production floor—that's true. But there's a prerequisite: the floor must have something for it to grow on.
Data source in this article: "Securities Daily", November 18, 2026, reporter Jia Li, report "Lighthouse Factories Move from Single Points to Contiguous Clusters" (World Economic Forum Global Lighthouse Network list: as of June 2026, 238 worldwide and 109 in China; Ministry of Industry and Information Technology: as of the first half of 2026, more than 56,000 basic-level smart factories, more than 9,000 advanced-level, and more than 500 excellence-level, with 34 leading-level smart factory cultivation projects publicly announced for 2026; the application penetration rate of key AI scenarios in manufacturing was 34.2%; "Enterprise Digital Transformation Index Report (2025)": SME maturity index 28.6, national 35.4, 86.92% at the basic informatization application stage; quotes from Shi Youkang, chief expert of CAICT, Song Tao of Lenovo, and You Yicheng of Gree; case data from LCFC, Sany Heavy Industry Beijing Pile Machinery Factory, Gree Jinwan Factory, and Sungrow; Morgan Stanley forecast of approximately RMB 80 trillion in cumulative incremental capital expenditure from 2026 to 2035). Odoo models and fields are based on versions 17 to 19 as examples; version 20 may change, so please check the official documentation before implementation.
Zhongdi Information · Enterprise Management Digitalization Observation · The enterprise cases in this article are cited from public reports and are not accompanied by third-party audits.
