Foreword:

In October 2026, the 36Kr Research Institute released the "2026 China Industrial AI Industry Development Research Report," noting that China's industrial AI has transitioned from the proof-of-concept stage to a phase of large-scale implementation. Driven by the dual core logic of "mechanism + data," the report reveals that industrial AI is accelerating its penetration across the entire manufacturing sector.

Author | Fang Wensan

Image source | Internet

Market Explosion: Penetration Rate Quintuples in Five Years

Data shows that in 2025, the adoption rate of large models and intelligent agents among Chinese industrial enterprises reached 47.5%, nearly five times the 9.6% recorded in 2024; the share of enterprises deploying AI simultaneously across multiple links jumped from 1.7% to 35%. In 2026, the addressable market size of China's industrial AI is expected to exceed 116.3 billion yuan, and is projected to reach 303.4 billion yuan by 2030, with a compound annual growth rate of approximately 25.4%. Comprehensive investment per project ranges from 2.5 million to 6.2 million yuan, and the return on investment cycle is generally 6 to 18 months.

Process Industry Emerges as a High-Value Arena

Process industries such as petrochemicals, chemicals, power, and steel have become the "touchstone" for testing the deep value of industrial AI, given their continuous production, strong multivariable coupling, and strict safety and low-carbon constraints. The industry as a whole is currently in the transition stage from L2 to L3—L2 represents single-point predictive detection, while L3 marks the watershed where AI intervenes in decision-making but humans provide the fallback. Leading enterprises have already validated the "autonomous operation factory" model in areas such as chlor-alkali and petrochemicals, moving toward L4 and L5-level unattended operations.

Implementation Bottlenecks and Breakthrough Pathways

The report acknowledges that the large-scale implementation of industrial AI still faces four mutually reinforcing challenges: extremely scarce anomaly samples (fault data at one factory accounts for only 0.03%), the conflict between black-box models and industrial traceability requirements, high upfront investment that leads small and medium-sized enterprises to wait and see, and a shortage of compound talent who understand both processes and AI.

The key to breaking through lies in five major synergies: building a high-quality industrial data system, promoting the deep integration of mechanism knowledge and data models to suppress AI hallucinations, reducing delivery costs through pretraining plus fine-tuning, cultivating compound talent and unifying the understanding of management and frontline staff, and creating a closed-loop validation mechanism that is "executable and accountable."

Future Trends: Model-Data Resonance and Physical AI

Looking ahead, the report proposes three major directions: first, "model-data resonance," using model training to force data governance and using on-site data to feed back into model iteration, forming a two-way closed loop; second, the rise of physical AI, with intelligent capabilities embedded in the equipment itself, enabling devices to integrate perception, analysis, decision-making, and execution; third, upgrading from single-point intelligence to agent factories, where agents for production scheduling, logistics, operations and maintenance, and other functions negotiate autonomously, promoting full-chain digital and intelligent autonomy. As national standards gradually take shape, industry competition is shifting from "demo shows" to competition in engineering delivery capabilities.

The following is part of the report: