In recent years, artificial intelligence technology has shown a trend of evolving from “cloud centralization” to “device-edge collaboration”.
In 2026, the Chinese government work report continues to emphasize the “artificial intelligence +” action and clearly promotes the spread of large models to terminal devices, covering smartphones, intelligent connected cars, and Industrial Internet of Things (IIoT) scenarios.
This policy orientation and technological evolution have jointly given rise to the large-scale implementation of “On-Device AI”, especially in the industrial field, triggering structural changes.
Traditional industrial automation relies on programmable logic controllers (PLCs), preset rules and centralized monitoring systems, and lacks environmental adaptation and real-time decision-making capabilities.
On-device AI deploys lightweight neural network models on edge devices to achieve low-latency, high privacy, and robust local intelligence, providing a new paradigm for the evolution from Industry 4.0 to 5.0.
Device-side AI has a multifaceted core impact on industrial automation.
In terms of production line control, traditional production lines operate with fixed logic and are difficult to deal with dynamic disturbances.
Device-side AI enables equipment such as industrial robots and AGVs to have local reasoning capabilities.
For example, a robotic arm equipped with a Distilled Vision-Language model can understand natural language instructions (such as “grab scratched parts”) and combine them with real-time images to complete the task.
Multi-agent systems realize task redistribution and path optimization through edge collaboration, significantly improving system resilience.
In terms of edge intelligent detection, end-side AI models (such as MobileViT, YOLOv9 – Tiny) can complete defect identification within <50ms, meeting the tempo requirements of high-speed production lines.
More importantly, the data does not need to be uploaded to the cloud, and it complies with the compliance requirements of the "Data Security Law" and the "Industrial Data Classification and Grading Guidelines" for sensitive manufacturing data to "not leave the factory area".
The case of a photovoltaic company shows that using domestic edge AI chips to deploy cell crack detection can save more than 3 million yuan in annual labor costs for a single line.
In terms of flexible manufacturing capabilities, in the face of small batch and multi-variety production needs (such as customization of new energy vehicles), end-side AI supports rapid model change.
Through prompt engineering (Prompt Tuning) or local fine-tuning, the same device can switch detection logic or assembly strategy in minutes.
Gartner predicts that by 2026, more than 70% of the world’s discrete manufacturing companies will deploy flexible production line modules with end-side AI capabilities.
Based on the above influences, five types of core business opportunities can be identified.
However, despite its promising prospects, on-device AI still faces multiple challenges in its industrial implementation.
There is a conflict between computing power and model complexity. Large models with tens of billions of parameters are difficult to deploy directly on edge devices and need to rely on efficient compression technology. The ecosystem is fragmented, including chip architecture (ARM/RISC-V), operating system (Linux/RTOS), and communication protocols (OPC).
UA/Modbus) are not unified, increasing development costs; there is a large gap in comprehensive talent, and there is a serious shortage of engineers who master industrial control, AI algorithms, and edge deployment. According to the Ministry of Industry and Information Technology, the gap is estimated to be more than 500,000; the business model needs to be verified, and small and medium-sized enterprises have a weak willingness to pay, so sustainable models such as SaaS subscription and performance sharing need to be explored.
It is recommended to systematically break the situation by establishing industry standard interfaces (such as AI-PLC communication protocol), promoting the interdisciplinary construction of “AI + manufacturing” in colleges and universities, and increasing subsidies for equipment updates.
On-device AI is reconstructing the technical base of industrial automation.
It not only improves detection accuracy and response speed, but also gives the manufacturing system autonomous perception, reasoning and decision-making capabilities. It is a key carrier to realize “artificial intelligence + manufacturing”.
As the “first year of device-side AI”, 2026 marks that my country’s industrial intelligence has entered the deep water zone.
In the future, as the performance of domestic chips improves, model compression technology matures, and safety compliance systems are improved, on-device AI is expected to become a standard configuration of the new generation of smart factories, helping China’s manufacturing industry move to the high end of the global value chain.
