Staff inspected the computer room of China Mobile (Qingyang, Gansu) data center. Photo by Xinhua News Agency reporter Lang Bingbing

  China Mobile recently announced the construction of the first computing power network in the country that covers the integration of four computing power (communication, intelligence, super, and quantum computing). The scale of computing power servers exceeds 1 million, and the scale of intelligent computing exceeds 43EFLOPS (10 billion floating-point operations per second). It has 21 intelligent computing centers, 3 national supercomputing centers, and 3 quantitative computing centers connected to the network. The dispatchable computing power resources account for one-sixth of the country.

  Experts believe that a new round of scientific and technological revolution and industrial transformation with digital intelligence as the main characteristics have in-depth development, data has become a new production factor, computing power has become a new infrastructure and basic energy, and artificial intelligence has become a new production tool, promoting the accelerated transformation of the economy and society from "Internet +" and "5G+" to "artificial intelligence +", putting forward new requirements for the evolution of computing network infrastructure, and also bringing new opportunities.

  The scale of computing power continues to expand

  As the cornerstone of the development of artificial intelligence, computing power has become one of the important infrastructures, and the demand for computing power has shown explosive growth. Data shows that as of the end of 2024, the total computing power scale in my country reached 280EFLOPS, of which the intelligent computing power scale reached 90EFLOPS, accounting for more than 30%. At the same time, the development of computing power infrastructure has achieved remarkable results, the computing power supply system with a stage optimized computing power supply has been initially built, the comprehensive capabilities of computing power infrastructure have been significantly improved, and the computing power industry has been continuously innovated and developed, and it has continued to empower thousands of industries.

  The "2025 China Artificial Intelligence Computing Power Development Evaluation Report" jointly released by International Data Corporation (IDC) and Inspur Information shows that the improvement in algorithm efficiency brought by DeepSeek has not suppressed the demand for computing power. Instead, due to the addition of more users and scenarios, it has promoted the popularization and application of large models, driven the construction of computing power in data centers, edges and ends, and driven the growth of computing power demand. China's smart computing power growth rate is higher than expected. It is expected that China's smart computing power will continue to maintain rapid growth in the next two years. In 2025, the scale of China's intelligent computing power will reach 1037.3EFLOPS, an increase of 43% from 2024; in 2026, the scale of China's intelligent computing power will reach 1460.3EFLOPS, twice the scale of 2024.

  At the 2025 China Mobile Cloud Intelligent Computing Conference held recently, Yang Jie, chairman of China Mobile, said that the innovative breakthroughs and wide application of general artificial intelligence technology will greatly improve the efficiency of computing power use and release new demand for computing power services. The total computing power will increase exponentially. In the next three years, my country's intelligent computing power scale will grow by more than 2.5 times, with an average annual compound growth rate of nearly 40%. The computing power structure will also change significantly, and the inference computing power demand will exceed the training computing power demand. In the next three years, the annual compound growth rate of inference computing power will reach nearly 4 times that of training computing power; by 2028, the scale of inference computing power will exceed the scale of training computing power. This process will inject new impetus into the development of computing networks and promote the industry to enter a new round of growth cycle.

  According to reports, China Mobile has built 13 national and regional smart computing center nodes and created multiple super-large smart computing centers. The computing network brain has been commercialized on a large scale and has been implemented in four hub-level and regional computing power nodes including Beijing-Tianjin-Hebei and the Yangtze River Delta. Especially in the Wuhu cluster computing power public service platform of the Yangtze River Delta hub, it has created the first national hub computing power scheduling platform in the country that "four calculations are integrated" is being built.

  The first "four calculations in one" computing power scheduling platform recently built in the country can support up to hundreds of millions of computing power calls per day. Sun Shaoling, deputy general manager of China Mobile Cloud Capacity Center, told reporters that the core of the computing power network is to use a powerful network to gather and connect the scattered computing power resources, and then dynamic global intelligent scheduling and task distribution of resources through the computing network brain. The computing network brain is the intelligent central system of the computing power network. Through global computing network scheduling, it can provide lower-cost computing power services, create a new computing power supply ecosystem that accurately matches supply and demand, support the construction of a national integrated computing power system, and empower the digital upgrade of thousands of industries.

  Optimization and reconstruction of computing power structure

  "The rapid development of artificial intelligence drives the optimization and reconstruction of computing power layout." Xia Bing, deputy director of the National Data Administration, said that since this year, big models such as DeepSeek have been released, triggering a new round of thinking on the development and changes in the global artificial intelligence field's demand for computing power. From the perspective of layout, large-scale clusters adapted to demand are still the strategic focus of the development of artificial intelligence, and the scale rule of artificial intelligence is still effective. At the same time, the computing power layout of "scale node deployment + hot spot area optimization + edge and end adaptation" has become a clear trend. Deeply gaining insight into the development of artificial intelligence technology, deeply subdividing application scenario types, and deeply promoting the development of chips of independent computing power is becoming more important for the current layout of computing power and optimizing computing power structure.

  The trend of green and intelligent drives the reconstruction of computing power supply methods. It has become an industry consensus to develop green, energy-saving, efficient and intelligent computing power. To maximize the intensive application and universal supply of social computing power, we need to focus on solving the current problems such as the coexistence of surplus and shortage of computing power supply side and the high threshold for application on the small and medium-sized demand side. By exploring the integrated intelligent scheduling of computing and networks, we can provide key groups, especially small and medium-sized enterprises with the best adaptation solutions for computing power, transportation capacity and storage capacity, and realize the ubiquitous and universal application of computing power.

  Scenario-based applications drive the reconstruction of computing power service models. Driven by the application of artificial intelligence, the leasing logic of computing power and the business model of Internet data centers will usher in reconstruction. Computing network services are expanding and upgrading from traditional resource-based services to task-based services and MaSS (model-as-a-service) services, and deeply penetrate from Internet finance to industries such as industry, transportation, and natural resources. The application needs of computing power to enter households, enterprises, parks and schools will promote the ubiquitous and universal development of computing power services, and form a new ecosystem of computing power development that is accessible to everyone, available everywhere, and on-demand services.

  The "2025 China Artificial Intelligence Computing Power Development Evaluation Report" believes that DeepSeek is based on great innovations at the algorithm level, bringing profound changes to the artificial intelligence industry in China and even the world. Algorithms have become the core engine driving the development of artificial intelligence, driving the development of computing power, and driving the transformation of computing architecture and data centers.

  Among them, in terms of computing architecture, in order to meet the high demand for computing resources of large models, it is crucial to improve the computing performance of a single node. In terms of data centers, the node failure rate increases with the growth of cluster size, and data centers need a more efficient monitoring system and advanced failure recovery mechanism. In addition, some intelligent computing centers have experienced the use of computing power in actual operations that have not met expectations. The development of intelligent computing centers needs to solve the problem of low utilization of computing power resources.

  Reduce the cost of computing power utilization

  After enterprises apply the artificial intelligence model, the demand for computing power has grown rapidly and systematic planning is required to balance performance and cost. Zhou Hongyi, founder of 360 Group, believes that when enterprises apply artificial intelligence, the choice of base models is crucial. A large model that can be deployed exclusively and privatically should be chosen, which can not only prevent data leakage and respond quickly, but also customize the enterprise knowledge base. Since the model is open source and free, the cost is almost zero. After selecting the base model, you must plan the computing power. Since multiple models usually work together within the enterprise, distributed computing power networks are generally used. Large enterprises can build their own computing power centers or purchase all-in-one machines, while small and medium-sized enterprises can deploy distillation models on their computers, that is, reduce the computing volume and storage needs of the models while maintaining high performance.

  Fang Xinping, deputy director of the Information Development Bureau of the Central Cyberspace Affairs Office, said that the integrated development of multi-computing power of integrated computing, intelligent computing and supercomputing should be strengthened, the computing power resource structure should be further optimized, the computing power utilization level should be improved, and the computing power utilization cost should be reduced.

  "The National Data Administration will continue to deepen the reform of market-oriented allocation of data elements. Continue to carry out pioneering trials, give full play to the decisive role of the market in resource allocation, expand the scope of pilot application of public transmission channels, reduce the cost of computing power used by small and medium-sized enterprises through multiple paths, and form a joint force for the government and the market to efficiently coordinate the allocation of resources." Xia Bing said.

  Yang Jie introduced that China Mobile will build a super-large "computing power factory". Promote the multiplication and expansion of 100,000-caliber intelligent computing centers, actively carry out forward-looking research on 100,000-caliber intelligent computing centers, build "centralized + distributed" inference computing resources, and form an intelligent computing system with "centralized, edge-based ubiquitous, middle-level training and side-promotion, and training and promotion".

  Fang Xinping suggested that high-quality computing power supply should be increased and the foundation for computing power should be consolidated. Accelerate the research and development of core software and hardware of computing power, promote breakthroughs in key technologies such as integrated circuits, parallel architectures, and high-speed storage, accelerate the development of computing power computing, computing power scheduling, and intelligent operations, promote systematic innovation of computing power technology, continuously optimize the layout of computing power infrastructure, and promote the linkage and complementarity between the east, central and western regions.

  "We will adhere to five coordinations including the integration of national diversified computing power and the integration of computing power in the east, central and western regions, optimize the layout of national computing power resources, and promote the coordinated development of computing power layout with regional economic and social development, and the construction of computing power infrastructure and energy." Xia Bing said. (Reporter Huang Xin)

[Editor in charge: Ran Xiaoning]