Published by ChinaMNC · Ningbo

Ningbo–Hangzhou AI Pilot Bases Show Foreign Founders How Industrial AI Works Inside China's Petrochemical Plants

A practical look at how Ningbo and Hangzhou are pairing petrochemical expertise with large-model and embodied-AI capacity, and what the two-city model signals for foreign founders building or operating companies in China's industrial AI market.

Two Cities, One Industrial AI Test Bed

China has been setting up a national network of pilot and demonstration bases to scale up artificial intelligence applications. Zhejiang province has so far received approval for six such bases. One of them, the only national AI application pilot base focused on the petrochemical and chemical sector, is located in Ningbo. Another, based in the Hangzhou high-tech zone, focuses on embodied intelligence.

For foreign founders, the interesting part is not the approval count. It is the operating model that has emerged between the two cities, and how it addresses two problems that any industrial AI company in China eventually faces: data cannot leave the factory, and laboratory demos rarely survive contact with a real plant.

Why Ningbo Is the Test Case

Ningbo is one of China's seven major petrochemical bases. It hosts large enterprises including Zhenhai Refining & Chemical, Daxie Petrochemical and Wanhua Chemical. The city's green petrochemical industry recorded above-scale output value of 546.8 billion yuan last year.

That concentration of plants creates a dense set of real production scenarios. It also creates a set of constraints. At a petrochemical enterprise in Daxie, a hydrocracking unit with annual capacity of 1.5 million tons runs at full load, with an automatic control rate above 95%. For a long time, however, operators still relied on experience: when feedstock quality shifted, temperature and pressure had to be adjusted manually, and unstable process parameters could send off-spec semi-finished product back for reprocessing.

The obstacle to using autonomous decision agents was not willingness. It was data privacy. Production networks are physically isolated, process formulas are core corporate secrets, and raw production data cannot leave the plant. Under a conventional public cloud model, sending production data to a Hangzhou cloud for model training was unacceptable to most enterprises.

The 'Data Stays in the Factory' Solution

The Ningbo pilot base team worked with Hangzhou partners including Alibaba Cloud and Zhijiang Laboratory on an alternative. According to Li Hongye, a relevant Alibaba Cloud executive, the Tongyi Qianwen open-source large model can be deployed privately on-site at the Daxie petrochemical campus. All raw production data remains on the enterprise intranet, and secondary model training and iteration happen only inside the factory.

On top of that private deployment, Hangzhou-headquartered Supcon Group connected its time-series model to the pilot base's multi-model collaboration service platform. Supcon specializes in industrial time-series forecasting, predicting how a unit will behave in the next hour or the next day. Engineers from the Ningbo base and Supcon traveled repeatedly between the two cities, reading historical operating records at the plant and importing a petrochemical general-knowledge corpus, blending chemical process mechanism knowledge with the time-series model.

After several months of on-site tuning, the results on the hydrocracking unit were reported as follows:

• Process parameter fluctuation range fell by 48.89%.
• About 7,500 tons of reprocessed product were avoided each year, generating roughly 1.2 million yuan in efficiency gains.
• Annual energy cost savings reached 450,000 yuan.
• Total added benefit for the unit is estimated at more than 2.7 million yuan per year.

Li Hongye's summary of the division of labor is direct: Ningbo has no shortage of petrochemical experts, while Hangzhou's strength is in large models and computing power. The cooperation is described as two-way empowerment.

Embodied AI Meets the Chemical Plant

The second track is embodied intelligence. Chen Jianxiong, a marketing engineer at Hangzhou-based Bawitong, a partner unit of the national embodied intelligence pilot base, found after visiting Ningbo that wheeled robots are often more practical than humanoid robots in chemical plant scenarios. Existing wheeled robots that meet flameproof and explosion-proof requirements tend to be bulky and struggle in narrow spaces inside chemical workshops.

The division of labor is clear. Hangzhou handles the robot body, motion control hardware and the on-board brain. Ningbo supplies real scenario requirements, chemical mechanism knowledge and time-series knowledge models. Both sides concluded that robot inspection involves more than spotting a visible leak. Pipeline leaks can be gaseous or liquid, and some damage is hidden inside pipes, requiring acoustic vibration detection or optical image recognition. After a signal is obtained, the chemical process knowledge base is needed to locate the source and produce emergency response recommendations.

A Ningbo pilot base official said that, nationally, embodied intelligence has not yet been systematically simulated in real chemical scenarios. In August this year the two sides formally launched joint research. The Ningbo base completed the country's first round of scenario recruitment, collecting more than 30 industrial scenario requirements and shortlisting more than 10 quality scenarios for implementation. Next, the Hangzhou base will iterate robot appearance, sensors and explosion-proof configuration based on Ningbo's requirements, while Ningbo will package chemical industry knowledge models and connect them to the robot brain.

A Commercial Intelligence Center for the Industry

A third workstream involves Zhejiang University's Pan Yunhe academician team. Ningbo is building a petrochemical industry commercial intelligence center. The center has integrated more than 20 domestic and international information sources and aggregated about 2.6 billion pieces of data across four dimensions: innovation chains, industrial chains, talent chains and capital chains. It conducts automated monitoring, analysis and early warning around the clock.

The raw material includes 210 million global patent records, information on 130 million business entities, and data on imports and exports, industrial policy and investment and financing. The Zhejiang University team handles external source connection, data screening, cleaning and refinement. The Ningbo team translates that into platform services that enterprises can call up. A manager at a small and medium-sized chemical enterprise in Zhenhai said companies no longer need to search for information themselves; they can track industry product quotations, port logistics and upstream and downstream capacity changes in real time, then combine that with their own plant capabilities to decide which product to make.

What This Means for Foreign Founders

For foreign founders, the Ningbo–Hangzhou case offers several practical reference points when registering or operating a company in China's industrial AI space.

• Design for private deployment from day one. If your product touches production data, expect customers to require on-premise or intranet-only operation. A cloud-first architecture may disqualify you in petrochemical and other sensitive sectors.
• Partner for domain knowledge. The model alone was not enough; chemical process mechanism knowledge and time-series forecasting were essential. Foreign teams should plan for local domain partners rather than assuming software alone wins the account.
• Use the pilot base system as an entry point. National AI application pilot bases function as matchmaking and testing platforms. Engaging with them can shorten the path from demo to real scenario.
• Think in two-city terms. Hangzhou contributes models, computing and talent; Ningbo contributes industrial scenarios and application knowledge. A China entity structure that can operate across both may access capabilities that a single-city setup cannot.
• Treat data compliance as a product feature. The entire architecture here was shaped by the requirement that raw data never leaves the factory. Founders who build that constraint into their pitch will find it easier to talk to Chinese industrial customers.

This summary is based on information published by the relevant Chinese government website and is general information, not legal or tax advice.