What's Behind China's Lead in Industrial AI

While everyone watched the models and marvelled at the Chinese AI miracle, China spent twelve years installing sensors, laying networks and writing policy documents. Now it has moved on to people.

The point I keep repeating in my books, in training rooms and in conversation sounds dull: China's success in AI is not a sprint. It is systematic work — long-term planning and governance that make it possible to invest steadily, year after year, in things that pay off a decade later. It now has something to train industrial AI on, and something to run it on. And it is now building the third pillar — people.

A serious advantage in industrial AI will emerge from this, I believe. Will emerge, not has emerged: below I show both what is already there and what is missing.

This started as an observation from a trip. Now I can back it with documents and dates — and show what the picture still lacks.

One framing note, to set expectations. I am not comparing countries and I am not predicting who wins the race. This is only about industrial AI — models that run on data from sensors, machine tools and rolling stock.

The three pillars everything else rests on

I have written many times that AI implementation failures are almost never technical. Boil the causes down and you are left with three.

Governance — strategy, portfolio, a review rhythm, the right to kill a project without being punished for it. Technology — data, its owners, its quality, the integration layer, the infrastructure it all sits on. People — from the top team down to the operators who understand what they are doing and can argue with a model. Failure in any one of the three wipes out the other two. Perfect data without trained people does not work. Trained people without data do not work either.

Below I go through the Chinese picture pillar by pillar, in the order China built them.

Pillar one: planning ten years out

Let me start with what we tend to dismiss as bureaucracy.

On 1 August 2013 the State Council issued the "Broadband China" strategy, document 国发〔2013〕31号. Its targets run seven years out, not one: by 2020, 400 million fixed broadband subscribers, coverage of over 98% of administrative villages, 50 Mbit/s in cities and 12 in the countryside. Guidance on the internet of things had already come out that February.

On 8 May 2015 — "Made in China 2025", document 国发〔2015〕28号. It contains a table I consider the most underrated item in the entire Chinese policy corpus. One line: the share of CNC machine tools in key production processes. 27% actual in 2013. 50% by 2020. 64% by 2025. The line below it: the adoption rate of digital design and development tools, from 52% to 84%.

Read those numbers again. This is not an AI plan. It is a plan to turn shop floors into a source of data, written seven years before the word "AI" became mandatory in every deck. A CNC machine is not just more precise than a manual one. It keeps a log.

On 1 July 2015 — "Internet+", 国发〔2015〕40号. Eleven action areas, one of them titled "Internet+ Artificial Intelligence". July 2015. Three and a half years before everyone started talking about AI.

On 19 November 2017 — the industrial internet, 国发〔2017〕50号, with 2020 targets: 300,000 enterprises on platforms, over 2 billion identifier registrations.

From there the documents get denser and more specific, and I have checked each of them against the primary source.

The national layer is the State Council's "AI+" plan to 2035: six application domains, eight enabling measures, milestones in 2027, 2030 and 2035. It answers "what" and "why". The technology layer beneath it is the "1397" strategy: one goal, three needs, nine directions, seven mechanisms; it answers "how" and "with what". The municipal layer is Beijing's ten measures on AI agents, where the city shifts from paying for compute consumed to paying for outcomes. The sectoral layer is the national rules on what an agent may decide on its own, defining an agent through five capabilities and sorting decisions into three risk categories. And a regulatory layer runs across all of it: 24 rules on AI have been in force since August 2023, and since September 2025 labelling of AI-generated content is mandatory under standard GB 45438-2025.

I will not retell these documents; each has its own article. What matters here is something else: thirteen years separate the first from the last, and they do not contradict each other. Each one stands on the previous rather than cancelling it. That is the systematic work I am talking about. Not brilliant foresight — writing a strategy document with a ten-year horizon and then walking it.

Timeline: what China laid down and when — technology and governance from 2013, the people loop from 2025

Pillar two: sensors and networks bought in advance

A plan can say anything. Let us look at what followed.

Networks. On 4 December 2013 the Ministry of Industry and Information Technology issued 4G licences to three operators; by the end of 2014 the country had roughly 100 million 4G users and 758,000 base stations. On 16 May 2015 the State Council's General Office set out the sums in document 国办发〔2015〕41号: over 430 billion yuan into networks in 2015 and no less than 700 billion across 2016–2017. 5G licences came on 6 June 2019, commercial tariffs on 31 October. By the end of 2025 the country had 4.838 million 5G base stations and 12.87 million base stations in total.

Robots. Here is the International Federation of Robotics series for annual installations in China: 36,560 in 2013, 87,000 in 2016, about 154,000 in 2018, 268,195 in 2021, and 295,000 in 2024 — 54% of global shipments. 2013 is the pivotal year: that is when China passed Japan and became the largest robot market in the world. The operational stock at the end of 2024 was 2,027,000 units, the largest anywhere.

One caveat, and I nearly got burned by it myself. The same federation reports robot density — robots per 10,000 employees — as 470 for 2023 and 166 for 2024, even though the stock grew over that period. The drop is an artefact: the denominator was revised against updated Chinese employment statistics, and the global average in the same report fell from 162 to 132. The series before and after 2024 are not comparable, and you cannot compare country densities across releases from different years. So from here on I talk about installations, not density.

Sensors. At the end of 2023 Chinese cellular networks carried 2.332 billion IoT subscriptions — 57.5% of all mobile network connections in the country. At the end of 2024, 2.656 billion. More than half of China's mobile connections serve devices, not people.

One more number. The 2015 plan required CNC on key operations to reach 64% by 2025. The ministry's report puts it at 69.5% as of the end of June 2026. Digital design tools stand at 86.3% against a planned 84%. A ten-year-old plan has been not just met but beaten on both lines.

That is what I mean by steady investment. Not "we allocated budget for AI", but: machines that keep logs; networks the logs travel on; sensors that outnumber people on those networks. Twelve years running.

What it produced: data nobody else has

Maintaining rolling stock and track is expensive and thankless. Everywhere in the world it is a cost centre, especially as equipment ages. The Chinese network meanwhile keeps growing and getting faster, and the higher the speed, the more the human becomes the weak link in the control loop: physically, a person cannot keep up.

China State Railway Group fitted sensors to infrastructure, track, wheelsets and carriages. The sensors measured vibration amplitude and frequency, accelerations and several dozen other parameters. That produced 195 terabytes — 57 data types across 23 categories, including dynamic waveform values from wheel-mounted sensors. Text and numbers, no images or video — the kind of volume no human can keep up with.

A model was trained on that data. The result was reported by the South China Morning Post, citing a peer-reviewed paper in the journal China Railway (《中国铁路》). The system warns maintenance crews of a developing fault 40 minutes ahead, with 95% accuracy. In the year before publication, not one operating high-speed line received a warning that required a speed restriction for track irregularities. Minor track faults fell by 80%. This covers the whole of the country's high-speed network — 45,000 km at the end of 2023.

Note the order. Sensors first. Then data. Then the model. Years separate "we installed sensors" from "the model warns forty minutes ahead", and the reverse order is physically impossible: you cannot train a model on data nobody collected.

This is where the line runs between industrial AI and everything else. Text for a language model can be taken off the internet — it has already been written. There is nowhere to get wheelset data from. It can only be collected: in advance, on your own hardware, at your own expense, without knowing exactly what you will do with it. Whoever started collecting ten years ago now holds something money cannot buy, because the real price was never money. It was time.

That is what "something to train on" means.

Pillar three: people — the one the money is going into now

Two pillars stand. China is building the third as we speak, and it is the newest part of the story.

Start with what already exists, because for some reason nobody mentions it. Chinese universities turn out more than 5 million STEM graduates a year — first in the world, according to Xinhua; the agency does not disclose its counting method, so I take the figure with that caveat. The Chinese Academy of Engineering estimates that China accounts for more than a third of the world's engineering graduates. In 2025 universities produced 5.23 million bachelor's graduates and 1.17 million master's and doctoral graduates.

Below the universities sits a layer that AI conversations ignore entirely. Around 29.4 million people are enrolled in secondary and higher vocational education: 6,741 secondary institutions, 1,554 colleges, 87 vocational universities. These are not the people who will write models. They are the people who will install and service the sensors — and without sensors there are no models.

The third layer is retraining the existing workforce. The "Skilled China" programme set a target of 40 million additional skilled workers over the current five-year plan. The result announced in September 2025: over 200 million skilled workers, of whom more than 60 million are highly skilled; over 92 million subsidised training placements; 72 new occupations added to the national occupational classifier.

And now the schools — the part every conversation about "the Chinese are teaching children AI" starts with.

On 12 May 2025 the Basic Education Teaching Guidance Committee under the Ministry of Education issued guidance on general AI education for primary and secondary schools. Here is the detail that gets dropped in the retelling: it is guidance, not an order, and it contains no hour requirements at all. The obligation appears one level down, in the regions:

Region	Date	Requirement
Shanghai	14 January 2025	"AI Fundamentals" in grades 4 and 7, one hour a week, at least 30 hours per grade
Beijing	6 March 2025	at least 8 hours per school year, all levels, from the autumn 2025 term
Guangdong	10 April 2025	grades 1–4: at least 6 hours a year; grades 5–6: at least 10; grade 7 and up: at least one hour every two weeks

The Shanghai requirement is nearly four times stricter than Beijing's, and it came first. In February 2026 a deputy education minister spoke about bringing AI into school curriculum standards, but the verb there is "promote", not "require".

What does this actually mean? Here I have to check myself. Eight or thirty hours a year is a signal of priority, not an effect. A child who sat down at a desk in 2025 reaches the labour market in the middle of the next decade. A school programme is an investment that pays back later. It says nothing about the present.

The present state is measured by a different number, and it is an unpleasant one. China's shortage of AI specialists is estimated at more than five million people, with demand exceeding supply ten to one. The figure comes from People's Daily citing the Ministry of Human Resources; I could not find the ministry's own document behind it, so I trust it less than the education statistics. But the direction is clear: the people pillar is the least finished of the three.

Hence a simple conclusion, and the numbers below back it up: the advantage has not been realised yet. Two pillars are ready, the third is in progress. Which is why I say "will emerge" rather than "has emerged".

What is already visible, and what is not

Now let us test all this against what can actually be measured — in both directions.

What is visible. The World Economic Forum keeps a list — its Global Lighthouse Network — of production sites that an independent expert panel has picked as leaders in digital transformation. As of June 2026 it lists 238 sites across more than 30 countries. Chinese sources, including state media and the Party journal Qiushi, count 109 of them in China — 45.8% of the list and first place in the world. The Forum itself does not publish a country breakdown; the tally is the Chinese outlets' own, and I could not verify it against a primary source. According to the Ministry of Industry, the country has built around 43,700 "smart factories" of various tiers, and the smart manufacturing sector together with industrial software has passed 4.5 trillion yuan in scale. Add 54% of global robot shipments and a CNC target beaten ahead of schedule.

What is not. The share of large manufacturing enterprises using AI is over 30%. Which means 70% do not. Robot density, mentioned above, is 166 robots per 10,000 employees — 22nd in the world. Small and medium-sized business is barely touched: the 2025–2027 plan covers the digitalisation of 40,000 small enterprises, and the target for cloud adoption among SMEs is "over 40% by 2027". The base layer belongs to someone else: the CAD and CAE markets are held by foreign vendors, with the largest domestic CAD vendor at around 9.6%.

One backdrop applies to everyone, not only China. On the same Forum's estimate, more than 70% of companies investing in advanced analytics and AI never get past the pilot stage.

Put the two columns together and you get precisely what I am describing. The scale is there because the first two pillars were laid over twelve years. The depth is not there yet because the third pillar is being built right now. This is not a contradiction in the picture. It is what the middle of the journey looks like.

What is already visible and what is not: 109 of 238 lighthouse factories and 54% of robot shipments against 22nd place on density and 70% of enterprises without AI

And separately — why a plan on its own guarantees nothing. In November 2017 a company called HSMC was registered in Wuhan with a declared investment of 128 billion yuan, about 18.5 billion dollars, and the goal of reaching 14-nanometre and sub-7-nanometre processes. There was local government backing, there was construction, there was a single ASML lithography machine — pledged to a bank in January 2020 against a 580 million yuan loan. In February 2021 the entire staff were told to hand in their resignations. Of the 128 billion yuan, 112.3 billion never arrived. The district's state platform had paid in 200 million yuan for a 10% stake, while the private shareholder holding 90% paid in nothing at all. The warning was written not by the press but by the district administration itself: a report dated 30 July 2020 spoke of a large funding gap and the risk of the financing chain breaking. The report was later taken down from its website.

The money, the plan, the construction — all of it was there. What was missing were people who could tell chip fabrication from a building site. Failure in one pillar wiped out the other two — exactly as it does in a 500-person company, only with six more zeros. The link to the pillars is mine, not the sources': they speak of founders lacking relevant experience and of signs of fraud.

And the boundary none of this crosses. With text data, China's position is the opposite. On estimates published in China itself, Chinese-language content makes up 1.2–1.3% of the global internet against 49.9–59.8% for English. Roughly one to forty. Add China's own statistics: of the 52.26 zettabytes of data produced in the country in 2025, 2.53 were retained — about 5%. Everything I have said about advantage is about data from sensors and machine tools. It does not carry over to generative models.

What this means for your company

Chinese scale does not transfer to you. The order does.

First. Data cannot be bought after the fact. It is the only one of the three pillars paid for in time alone. A model can be bought off the shelf, a platform purchased, an integrator hired — but last year's telemetry cannot be bought, and it cannot be hurried. Hence a rule I take from the Chinese story and see confirmed in my own practice: instrumentation and data collection start before everything else, even when it is unclear what the data will be used for. China installed sensors without knowing about large language models. It worked precisely because the decision had nothing to do with AI.

Second. The people track does not come after data. It is just as long. You do not produce someone who argues with a model and turns out to be right in a single quarter — not with courses and not by hiring. The timescales here and below are my own estimate from industrial projects, not a measurement. If data will be ready in two years and training takes two to three, you start on people today, not afterwards. China, incidentally, started this pillar last — which is exactly why it now has a shortage of five million people and 70% of enterprises without AI. Its story works here as a warning, not a model.

Third. The platform is bought last, and cheaply. It ages fastest, and it is the only one of the three you can replace in months. The temptation runs the other way: a platform looks like a way to close all three pillars with one purchase. It does not.

Data and people are paid for in time; governance and the platform are bought: the order of investment runs top down

A table to take with you

Take your portfolio and fill in five columns for each initiative. Nothing more is needed — this is enough to make the mismatches visible on one screen.

Initiative	Which pillar it stands on	When that pillar will be ready	Confidence level	Waiting on another pillar?
Predictive maintenance, pump station	technology: telemetry	Q2 2028	grounds to believe	yes — operators not trained
Dispatcher training	people	Q4 2027	assumption	no
Portfolio review procedure	governance	Q1 2027	confirmed	no

The confidence column is not decoration. Three words: assumption, grounds to believe, confirmed. Percentages are banned: "60% ready" sounds solid and means nothing, while "assumption" is honest: there is no date, only hope.

The "waiting on another pillar" column is the important one. The moment a single "yes" appears in it, you have a frozen budget: the money is spent and the result is on hold.

How to read the filled-in table:

1. Not a single readiness date — you have no schedule, only a list of initiatives. That is the first task, not the fifth.

2. Every date at "assumption" — the table was filled in for show; go back to the pillar owners.

3. A "yes" in the last column — a decision on each: accelerate the pillar or freeze the initiative. There is no third option; "let's wait" is the frozen budget by another name.

4. Is the earliest start date on the longest pillar? If not, the mismatch is already built in and will surface in a year or two.

Why this does not happen on its own

At your next meeting with the leadership team, ask for the readiness date for data. You will most likely hear a status rather than a date: "in progress", "being scoped", "depends on the integration work". Not because people are dodging. Because the link between an initiative and its parent reason — a goal, a commitment, a constraint, a risk — is recorded nowhere and lives in the head of whoever created the initiative. While it lives there, the question "what is waiting on whose pillar" has no answer, and nowhere to look for one.

The remedy does not depend on software. Every initiative must have a derivable parent reason and a stated readiness date. "Derivable" is the key word: not a mandatory field on a form. The moment the link becomes a mandatory field, you are demanding a goal before the first useful action — and people either stop creating initiatives or type whatever fits. The right way round is this. An object may exist without a link, but the system must be able to show its parent reason — and to offer the link once the object has produced something useful.

In BAEOS, the executive system we are building, both rules are invariants: they cannot be switched off in the settings. Every inference carries its grounds and one of the three verbal confidence levels, and an object's link to its parent reason is not a mandatory field but a derivable one.

What it does not do, and this is worth saying plainly: it does not work out how long your pillars will take to mature, and it does not assemble your portfolio. Only the person accountable for data can name the readiness date for data, and prioritisation is yours. A tool that can show grounds is useful exactly to the extent that a real person has named those grounds.

What I am not claiming

Three caveats, so nothing gets read into this that I did not say.

I am not claiming that China has already won or is bound to win. The advantage I describe is still forming, and it may never form at all — there is a talent shortage, the base software layer is foreign, and 70% of enterprises have no AI. I am not claiming that state planning beats markets: the state set the direction and created the demand; the execution came from a mass export market; and I cannot separate their contributions from open data. And I do not carry any of this over to generative models — there the picture is reversed, as the numbers above show.

Here is what I am claiming. Installing sensors and laying networks for twelve years running is dull work. It produces no quick headlines and it looks bad in a report. Which is precisely why almost nobody does it. China did — and now holds something that cannot be bought later.

Your horizon is shorter than China's and your budget is smaller, but the rule is the same. Start collecting data before you understand why you need it. In three years you will understand, and by then there will be no way to collect it after the fact.

If you want to go deeper. The three-pillar frame and the eight barriers to implementation are covered in the book Artificial intelligence. Freefall, third edition. You can download it free, no registration required. The practical half — selecting use cases, data readiness, pilot thresholds — is in the second book, Artificial intelligence. Practical guide for implementation.

Observations from the trip this all started with are in an earlier piece, What We Brought Back from Our Trip to China. The market in numbers is in my review of the artificial intelligence market in China.