NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China Named 34 "Lighthouse" Factories. 70% of Their AI Is Already on the Floor.

China named 34 more lighthouse factories. Together they report 1,000-plus industrial agents and more than 300 million token calls a day, and almost no return data.

2026-11-19 · 4088 words · NeuroAI
China Named 34 "Lighthouse" Factories. 70% of Their AI Is Already on the Floor.

The number to hold onto is not 34. It is 300 million.

That is the daily volume of token calls reported across the second batch of 34 lighthouse factories, named on 19 November by the Ministry of Industry and Information Technology and five other departments at the World Intelligent Manufacturing Conference in Nanjing. The same group reported more than 1,000 industrial agents built and AI applied in more than 70% of its production scenarios.

A four-rung ladder

China grades smart factories in four tiers: basic, advanced, excellent and lighthouse. Lighthouse is the top, and it is small on purpose. The second batch adds 34 plants, bringing the cumulative total to 49.

That ratio is the first thing to understand about the programme. It is not a claim that the manufacturing base has been upgraded. It is a reference fleet, selected to be copied.

The 34 companies jointly issued a "Lighthouse Factory Cultivation Action" in Nanjing, which reads as an intent to diffuse rather than to decorate. A ceremony would not need an action plan attached to it.

From assistance to the floor

The wording of the release is more interesting than the totals. AI, it says, has moved from operational support into research and development design, process optimization, production operations and production management.

That is a claim about where the model sits in the loop. A system that schedules maintenance is an adviser. A system that sets process parameters is inside the product.

Three figures carry that claim, and they are measured in three different units. Seventy percent is breadth, the share of scenarios with an AI component. One thousand agents is structure, the number of separate tasks given their own software. Three hundred million tokens a day is intensity, the volume of inference actually being run.

Breadth and intensity are not the same claim, and only intensity costs money every day. Every number released describes adoption or activity; none describes the return. A factory can cover 70% of its scenarios with tools that nobody uses twice, and the token count is the closest thing here to evidence that they do.

The apparel case

Heilan Home had its demand-insight flexible fast-response factory selected, the first in China's garment industry, according to Xinhua. Chairman Zhou Lichen attended the Nanjing event.

The disclosed base is large. More than 7,500 stores in China and abroad, over 200 million online followers, and more than 1 billion records of RFID and 5G-A passive IoT data. On top of that sit an apparel-specific large model, generative design tools and a process knowledge graph, with a cross-modal system the company says turns a design brief into an image in seconds.

The store count and the follower count are usually read as scale. Here they are better read as the input layer: a demand signal detailed enough to brief a model before a pattern is cut.

The number that travels furthest is 90%. Heilan Home says sampling fabric loss fell by 90%.

That is a material cost, measured at the point of cutting. It is far harder to dispute than a percentage gain in efficiency, and it is the kind of figure that persuades another manufacturer in the same trade.

Why an apparel factory matters

The list could have been all semiconductors and electric vehicles. Xiaomi Auto is on it, with a plant described as an AI-driven agile manufacturing site for high-end new energy vehicles.

Instead, the most detailed disclosure came from a clothing retailer with thin margins and about 300,000 people across headquarters, stores and the supply chain. If the large model and the knowledge graph hold up there, the argument that these tools only pay off in high-value electronics weakens.

Apparel is also the harder test. A car plant runs one product architecture for years; a garment factory renegotiates its product every season, which is precisely where a generative design loop earns its keep.

What 49 buys

My judgment is that the programme's value is diagnostic rather than statistical. Forty-nine plants cannot move a national industrial aggregate, and nothing in the release claims they do.

What they can supply is a tested configuration: which scenarios to automate, which data to instrument, what a working knowledge graph looks like once it is loaded with a decade of store-level records. That is portable knowledge, and it is the scarce input for everyone below the top tier.

The figure to watch is the 300 million. If daily token calls climb while the factory count stays near 49, the models are going deeper into the same plants. If the count climbs quickly and token volume per plant falls, the label is spreading faster than the practice.

One of those two outcomes is a manufacturing upgrade. The other is a naming ceremony, repeated.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "China Named 34 "Lighthouse" Factories. 70% of Their AI Is Already on the Floor," 19 November 2026, https://neuroai.site/a/na-policy-china-lighthouse-factories-1119b

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories