Cisco exec sees China holding sway in industrial AI

China's vast manufacturing base could give the country an edge in industrial artificial intelligence by providing a rich source of operational data to train models to spot defects and predict equipment failures, a senior Cisco executive said.

"To develop physical AI, companies need manufacturing lines, sensors and the data generated by those production lines and sensors to train their models," said David Siroky, Cisco's AI architecture lead for Asia Pacific, Japan and Greater China. "China has spent more than two decades building that infrastructure."

Siroky spoke to China Daily on Wednesday on the sidelines of the 2026 World Internet Conference Wuzhen Summit in Wuzhen, Zhejiang province, which concluded on Thursday.

The economic logic is straightforward on the factory floor, he noted, pointing to automotive assembly. A paintwork blemish caught and buffed out on a domestic assembly line costs a fraction of what it would take to rectify once the vehicle has been shipped overseas to an export market such as Australia.

Major manufacturers are already executing this playbook. BMW Group said its production base in Shenyang, Liaoning province, deployed an AI-assisted computer vision system in its paint shop to scan photographs of every car for micro-scratches and surface imperfections, instantly marking affected zones with precision lasers for rapid repair.

Scaling such capabilities across industries is supported by China's adoption of automation. China accounted for 59 percent of the industrial robots installed worldwide in 2025, with 354,000 new installations, according to the International Federation of Robotics.

Turning the data generated on factory floors into material that can be used to train models takes further work. Guidelines published in January by the Ministry of Industry and Information Technology and seven other government bodies outline how manufacturers can collect, clean and label factory data. They set a target of building 100 high-quality industrial datasets by 2027 for uses including process control and fault diagnosis.

For Cisco, the rapid deployment of industrial models has opened up immediate commercial opportunities, particularly in high-speed networking.

As enterprise AI models expand, they quickly outgrow single-server limits. Training or fine-tuning them requires clustering dozens or hundreds of graphics processing units (GPUs) across multiple server racks, requiring massive data exchanges with near-zero latency.

"The fastest-growing demand we see in the China market is for high-speed networking infrastructure to connect all these GPUs," Siroky said.

Cisco's overall product revenue in China rose 23 percent year-on-year in the fiscal year ended July 25, according to its annual report.

Companies often discover how much network capacity they need only after they try to take a project beyond a pilot, he said. Many discover after initial proof-of-concept trials that their networking throughput is insufficient, or that their facilities cannot supply enough electricity to power dense GPU clusters.

Despite the dizzying pace of AI development over the past two years, Siroky believes the broader transformation has barely begun. "For all the speed we are seeing, we are still only scratching the surface," he said.

At the software layer, open-source models are playing an increasingly important role, driven primarily by enterprise cost discipline. Customers are scrutinizing recurring cloud expenses and turning to domestic open-source leaders such as DeepSeek and Alibaba's Qwen.

"Cost is driving the shift," Siroky explained. By deploying smaller, targeted open-source models directly on on-premise infrastructure, enterprises can cover their hardware investment upfront and process recurring workloads without incurring per-query cloud API fees. On-premise deployment also resolves strict data privacy demands for institutions like hospitals, which must analyze sensitive records without sending data beyond their internal firewalls.

lijing2009@chinadaily.com.cn