《中小工廠裝 AI：意願同掏錢之間，隔住 70 個百分點》Small Factories and AI: Seventy Points Between Wanting It and Paying for It

[Zheng Ge's Take]

On 20 October the Ministry of Industry and Information Technology spoke at a press conference about a new generation of intelligent manufacturing, and said one thing that matters: the policy measures concerned are being studied and drafted. Drafted, not issued. I am putting that sentence first because plenty of the reposting this week has read it as a policy already in force.

The press conference also said that the main direction for a new generation of intelligent manufacturing, set at September's national advanced manufacturing conference, is now entering the policy-implementation stage, and that the ministry will speed up the deep integration of artificial intelligence and manufacturing.

Put the two together and the meaning is this: the direction is settled, the measures are still being written. No rule is in force, there is no deadline, and there is no threshold at which a factory that does not install AI is out. I would rather no reader finishes this piece thinking an order has been handed down.

What the press conference said, and what it did not say

The same press conference put three tiers of numbers on the table. I copy them as given:

And one more set: domestic compute and domestic large models have already produced more than a thousand application scenarios across twenty industries, steel, automobiles and pharmaceuticals among them.

Taken together those three tiers are a pyramid. Basic level, more than 56,000: the body is thick. Advanced level and above, under 10,000: the waist narrows at once. Leading level, 15: a needle point. This is a description of where things stand. A description of where things stand tells you which tier you are on. It does not tell you when the paper is due.

What actually made me sit down and write this was not the policy. It was another set of numbers.

Wanting it and paying for it are seventy points apart

CAICT, the China Academy of Information and Communications Technology, surveyed more than 6,000 small and medium-sized enterprises: 50.2 per cent already use AI somewhere in their operations; 26.7 per cent plan to bring it in within two years; 83 per cent say they are willing to pay for it. But in the first quarter of 2026, only about 10 per cent of industrial enterprises above the designated size actually booked any AI-related spending.

Between 83 per cent and 10 per cent there are seventy-odd points. That gap is the real cross-section of SME transformation.

One thing I have learned in business: "are you willing" and "will you pay" are two completely different questions. The first asks about attitude. The second asks about cash flow. What sits between them is not a lack of conviction. It is three thresholds.

Three thresholds: can you count it, can you afford it, does it come back fast

The first: can you count it. The owner wants a concrete input and output — how much money goes in, how many months until it comes back. Not a vague line about yield rising by some number of points. The vaguer the description, the more certain I am that whoever wrote it has never stood on the shop floor.

The second: can you afford it. Customised retooling runs to a million yuan and more. For a factory whose margin is scraping along in single digits, that number means pledging a year of profit. Pledging a year of profit for an unproven return is not ambition. It is a bet.

The third: does it come back fast. If the payback period runs past twelve months, most owners choose to wait a little longer. Waiting is not unwillingness. The arithmetic does not clear.

Of the three, the third is the hardest. The first two are sums: work them out and you have an answer. The third is confidence: you have to believe the number is real before you have seen the result.

Two samples: fifty thousand yuan, and just under a million

But people have already walked this road. Two samples: ten times apart in size, twenty times apart in spend, and both routes work.

A thirty-person hardware factory in Dongguan, on a 50,000-yuan plan. An industrial camera plus a free algorithm platform for visual inspection; a laser displacement sensor for critical dimensions; a large-model tool for production data analysis. The result: within two months, first-pass yield rose from 92 per cent to 98 per cent, customer complaints fell 70 per cent, and the spend paid for itself in two months.

A five-hundred-person machining plant in eastern China, on just under a million yuan. Three scenarios: AI visual inspection, intelligent scheduling, predictive maintenance. The result: the defect rate fell from 8 per cent to 2.3 per cent, scheduling efficiency rose 35 per cent, unplanned downtime fell 40 per cent, and the spend paid for itself in a little over six months.

What the two have in common is not money. It is a small incision. Neither began with "smart factory, whole plant". Both began with one process, one workstation, until the arithmetic cleared — and only then added the second.

The threshold is falling

Black Lake Technology, the basic edition of its digital work-order system: 10,800 yuan a year.

China Mobile Hebei's "train as you roll out" model: a single deployment costs around 50,000 yuan.

Taizhou's Luqiao district, where the local government built a shared AI foundation: compute costs 95 per cent lower than a company building its own, and agent applications at 20,000 to 30,000 yuan.

For a factory turning over tens of millions a year, starting money at that scale has dropped into the range where the owner will sign it himself. No board meeting, no waiting for the next funding round — money he can commit on his own signature.

Where the decision starts: three specific wounds

That machining plant in eastern China did not start from "we must go digital". It started from three specific wounds: a blended defect rate of 8 per cent, an on-time delivery rate of 82 per cent, and about 600 hours of unplanned downtime a year.

All three map directly onto cash leaving the business. That is why the payback on every purchase could be worked out in advance, month by month. This is what "can you count it" looks like in practice: not an indicator, a cash-flow table by month.

Put the incision on the visible pain. Keep the spend inside the batch size the company's cash flow can carry. Compress the payback period to within twelve months. Those three lines are worth more than any transformation roadmap I have seen.

Zheng Ge's Take

"We don't know how to do it" is still the hardest of the three thresholds to get around.

74 per cent of enterprises say they are short of digital talent, and the overall shortage of people who can span both sides is put at 25 to 30 million. Someone who understands the process and understands AI is extremely rare; the data sits scattered across paper ledgers and incompatible systems, and a model that arrives has nowhere to start.

So I turn all of this into three lines that are useful to a small or medium manufacturer.

One: do not try to rebuild the whole plant with AI on day one. Get the arithmetic to clear at one workstation, on one process first. Win a template you can copy.

Two: the equipment end is where to cut in first. It is the part that directly decides yield and delivery reliability, and it costs far less than rebuilding every system in the plant.

Three: government platforms are lowering the threshold. Shared AI foundations, subsidy schemes and looser credit are making the "can you afford it" question easier.

The ladder of tiered cultivation will not ask every company to climb to the top, but the bottom rungs are turning into the ticket in. For a small manufacturer wedged inside somebody else's supply chain, being needed comes before making money.

Source: 今日頭條 (Toutiao), by 樊鴻傑, a doctoral candidate in statistics at Beijing Technology and Business University, 20 October 2026. The policy material quoted here is what the Ministry of Industry and Information Technology said at its press conference on 20 October; the measures on a new generation of intelligent manufacturing are still being studied and drafted and have not been issued. The CAICT survey figures, the two factory cases and all cost figures are as cited in that article. They are restated and commented on here, not independently verified for this piece.