AI Empowers Qingdao Metro to Build the Industry's Largest Intelligent Computing Center

Planning for Qingdao Metro began in 1987, and construction of its first line broke ground in 2010. By the end of 2012, the Qingdao Metro Group was established to undertake the city's rail transit construction, operations, investment, financing, and business development. Guided by a concept of smart, green, integrated, and innovative urban rail, the group focuses on high-quality and sustainable development and is committed to the coordinated development of its two core businesses: rail transit and low-altitude economy. In the rail transit sector, its long-term plan envisions a network of 21 lines totaling 900 kilometers. To date, a total of 13 lines spanning 506 kilometers have been approved, with 8 lines covering 352 kilometers currently in operation.

In November 2023, Qingdao Metro formulated a next-generation smart urban rail plan based on AI, and partnered with 19 organizations, including Huawei, Shandong University, and Beijing Urban Construction Design & Development Group, to integrate large model technology into the technical system of smart urban rail. By building an intelligent computing center, an AI platform, and a model evaluation platform, Qingdao Metro has advanced data governance and developed large models and agents specifically for the urban rail industry, aiming to enhance operational quality and efficiency. The plan came to fruition in April 2025, when, under the guidance of the China Association of Metros (CAMET), Qingdao Metro released China's first AI model for the urban rail sector. This model serves as the technical foundation for a 26-agent cluster spanning four core business domains: construction, operation, maintenance, and enterprise management.

To address the complex and diverse business scenarios in urban rail transit, Qingdao Metro and Huawei have deepened their collaboration to drive the rapid deployment and continuous optimization of the AI model and multi-scenario agents. Together, they have built a fully autonomous intelligent computing center for all business scenarios using Huawei's Atlas SuperPoD solution.

Low-threshold, agile deployment for rapid business ROI validation: Eight small supernodes (64 NPUs) were built in the first phase, significantly lowering the initial investment and the threshold for equipment room renovation. This enables fast iteration and on-demand capacity expansion, drastically shortening the cycle from project initiation to business value realization.

Shared pool for training and inference resources, increasing resource utilization by 30%: Training and inference tasks can share the same compute pool, with dynamic and elastic resource allocation. This approach is particularly well-suited to urban rail transit businesses, which feature multi-scenario concurrency and pronounced tidal traffic patterns.

Full-stack high reliability, meeting the strict 24/7 online requirements of rail transit: The Atlas AI cluster can locate faults within minutes and automatically complete resumable training within 10 minutes, ensuring core businesses are uninterrupted.

Open and compatible architecture for smooth evolution to large-scale supernodes: Relying on a standardized interconnection architecture, the eight-unit cluster can be seamlessly expanded to create supernodes with 384 or more NPUs through future node additions.

Acting as the "computing engine" of Qingdao Metro's digital transformation, the intelligent computing center has reliably supported the training of the domain-specific AI model, as well as the development and rapid deployment of agent applications. Piloted on Qingdao Metro Line 6, various agents have cut manual operations by approximately 70%, boosted overall workforce efficiency by over 30%, and accelerated fault handling by more than 70%. Such improvements help to drive business restructuring, process reengineering, and organizational transformation.

Accelerated troubleshooting: Leveraging the huge computing power of Atlas supernodes in the intelligent computing center, the troubleshooting process for complex faults such as device trips has been streamlined from 26 steps with an average duration of 12.67 hours down to 8 steps, shortening the average fault handling time to 2.67 hours and boosting overall efficiency by over 70%.

Closed-loop management of the entire business chain: The intelligent computing center provides the computing backbone for core systems such as line network–level smart operation, maintenance, and energy management, enabling closed-loop management of the entire business chain spanning dispatching, O&M, service, safety, and decision-making.

As China's first computing platform to position large models as the core infrastructure of smart urban rail, Qingdao Metro's intelligent computing center provides a replicable model for the digital transformation of the nation's urban rail industry. Serving as a testing ground for advanced computing technologies and a platform for collaboration among government, industry, academia, research, and application, it builds a bridge for communication between the rail transit industry and the advanced computing sector, accelerating the cultivation of new forms of productivity for urban rail transit. Furthermore, the adoption of a fully China-developed computing platform comprehensively supports the implementation of Qingdao Metro's digital and intelligent transformation system, ensuring independent intellectual property rights and security of the entire urban rail digital system.