By Peng Cheng
Sixty years ago, Moore's Law struck like the spring thunder of Jingzhe, triggering a global technological wave. Thirty years ago, computers began to enter personal lives, and the second-generation mobile communication system was widely adopted. As the times evolved, China's computing power industry grew rapidly. Over the past three decades, it has achieved a breakthrough from nothing to something in the field of high-performance computing, and gradually established a complete industrial chain from chips and complete machines to systems. In recent years, it has continued to break through, achieving a leap from "following" to "running alongside". In 2026, China's computing power industry has reached a critical historical node.
From its proposal at the Political Bureau meeting of the Central Committee in April, to its successive deployments at the executive meetings of the State Council in May and July, the computing power network has been listed alongside the water network, the new power grid, and the new-generation communication network, and included in the "Six Networks" major infrastructure system. The National Development and Reform Commission (NDRC) estimates that during the "15th Five-Year Plan" period, the construction of the computing power network will add 4 trillion yuan in direct investment. The scale of China's computing power construction has reached a level that draws global attention. Data from the Ministry of Industry and Information Technology (MIIT) shows that by the end of June 2026, the scale of China's intelligent computing power had reached 2,185 EFLOPS (FP16 precision), a year-on-year increase of 177%. Data released by the National Data Administration shows that China's daily average AI token invocation volume surged from about 100 billion in early 2024 to over 140 trillion in March 2026.
On the other hand, data from institutions such as the China Academy of Information and Communications Technology (CAICT) shows that the average GPU utilization rate in domestic intelligent computing centers is less than 30%. This is the current state of China's computing power industry: rapid growth in total volume, but long-term low effective utilization rate; globally leading construction capacity, but operational capacity still has shortcomings. If the past five years were the "construction period", with intelligent computing centers and supercomputing centers springing up across the country, then the next five years will be the "operation and application period". The industry's problem has become how to sell, utilize, and generate value from computing power.
This transition exposes the deep-seated soft underbelly of China's computing power industry: we have accumulated certain capabilities in hardware manufacturing, but there is still much room for improvement in ecosystem construction, service operation, and system collaboration. As AI brings explosive demand, how should China's computing industry proceed? To answer this question, we need to return to the source of the industry and find the rules in the thirty-year depth of the industry.
01. Three Major Rules of the Domestic Computing Power Industry
Looking back at the industrial history of China's computing power, three major evolutionary rules can be clearly sorted out: from complete machine integration to independent research and development of underlying core components, from single equipment to ultra-large-scale clusters, and from hardware delivery to service operation. These three rules represent the development trajectory of the entire Chinese computing industry.
Reverse growth of the industrial chain from complete machines to core components. The high-performance computing industry usually starts with underlying core components such as chips and gradually develops upward to complete machine systems. However, the growth path of China's computing industry is exactly the opposite of the global mainstream rule. It started with complete machine integration in the early stage, and was later squeezed by the external environment, pushing backward to underlying core components, taking a path of "reverse growth". Today, China's computing industry is undergoing a transition from "assembler" to "core definer".
From single nodes to clusters, the scale effect redefines computing power delivery. In the AI era, as the scale of model parameters evolves from tens of billions to trillions, the demand for computing power in the training process continues to grow, and the computing power supply of a single server or a single accelerator card is difficult to support independently. At the same time, the Mixture of Experts (MoE) architecture further amplifies the demand for large-scale parallel computing and high-speed interconnection communication, making it difficult for a single node to bear it independently. As a form of high-density interconnection of multiple cards, super nodes alleviate the single-point bottleneck to a certain extent, but when the model scale continues to expand, it is often necessary to expand through clustering to break through physical boundaries. Data shows that in 2025, China has built 42 intelligent computing clusters with 10,000 cards, and the delivery unit of computing power has changed from "unit" to "cluster". The performance competition of single equipment has given way to systematic competition in cluster scheduling, network interconnection, cooling systems, and overall energy efficiency.
From hardware delivery to service operation. Application generalization, as the AI industry enters the inference era, the demand for computing power shifts from centralized training to fragmented and continuous invocation. One-time hardware delivery is difficult to match dynamic demand, and computing power needs to be used on demand like water and electricity. The industry has begun to expand from one-time hardware delivery to computing power leasing and token billing. Taking the Sugon 8000 100,000-card AI super cluster deployed in Zhengzhou as an example, by accessing the National Supercomputing Internet, it has realized on-demand leasing and pay-as-you-go billing of computing power. Enterprises can obtain large-scale computing power support without building their own computer rooms. This shift in model marks that such computing power manufacturers are expanding from single equipment manufacturers and providers to computing power operators.
02. The "3+1" Flywheel: Breaking the Application Barriers of China's Computing Power
Today, China's computing industry is at the intersection of "chips, systems, and services". The construction capacity of computing power infrastructure has been basically formed, and the problem facing the next stage of the industry is how to break through application barriers. To solve this problem, it is necessary to clarify where the comparative advantages of China's computing power industry lie. The global AI computing power landscape has entered the era of "dual-track parallelism": the United States leads in high-end chip innovation and underlying framework definition, while China has the world's largest AI application market, the richest industry scenarios, the most complete electronic manufacturing industrial chain, and the fastest Token consumption growth rate. This means that China's computing power industry cannot copy NVIDIA's route, but needs to use comparative advantages to integrate huge domestic demand, diverse industry scenarios, and complete manufacturing capabilities into an efficient closed loop from chips to applications. The intuitive expression of this implementation, just as Li Jun, Chairman of Sugon, mentioned the "3+1" layout recently: focusing on computing chips, computing systems, and computing services, joining hands with ecological partners to grow bigger and stronger.
First is chips. Chips consolidate the foundation and determine the upper limit of China's computing power. The competition of computing power, in the final analysis, is the competition of underlying core components. Without independent underlying capabilities, supply chain security is out of the question, and the pricing power and iteration rhythm of computing power will be in the hands of others. At present, domestic AI chips are showing a trend of blooming everywhere. Domestic manufacturers are showing their strengths, forming differentiated layouts in different tracks such as training, inference, and graphics rendering. Hygon Information has shown unique advantages and provided a unique standardized interface. The Deep Computing DCU adopts the GPGPU architecture, and the software platform DTK is compatible with the ROCm ecosystem, naturally undertaking the migration path of CUDA developers. Developers do not need to learn a new set of toolchains specifically for Hygon, and existing models can be migrated with low code. At the same time, the Hygon C86 CPU, as a general computing base, forms a full-precision coverage of "general computing + AI acceleration" with the Deep Computing DCU. The whole machine collaboration of the two further reduces the scheduling complexity in heterogeneous computing. This "compatible and open" route makes it the smoothest interface between chips and systems in large-scale cluster deployment.
Second is systems. Systems build the skeleton. In the era of super nodes and clusters, it determines the capability of the computing power hub. During WAIC 2026, super nodes became the direction jointly bet on by chip manufacturers. Alibaba's Panjiu AL128 super node, ZTE's OEX super node, and H3C's UniPoD S80000 super node competed on the same stage. After the large-scale deployment of super nodes, how to efficiently provide external services has become a new industrial topic. Sugon 8000 (Dengfeng), which was awarded the "Treasure of the Museum" at the WAIC conference, is China's first fully domestic 100,000-card AI super cluster. It not only proves that we have the engineering capability to build large systems, but also proves the strength of this level of system to serve industrial needs by completing in-depth optimization of more than 300 key applications and submitting 15 large applications at the level of the Gordon Bell Prize.
Third is services. Services connect the veins, allowing China's computing power to flow into all walks of life. Domestic large-scale intelligent computing clusters have begun to provide external services through computing power leasing and Token-based billing. The "Token factory" model is gradually rising, and the three major operators have also successively launched Token-based billing services. Computing power is gradually becoming a standardized resource that can be obtained on demand. Under this trend, domestic computing power hardware manufacturers are breaking through application barriers by upgrading the computing service system. For example, relying on platforms such as the National Supercomputing Internet and Sugon Cloud, Sugon enables more users to use high-end computing power more efficiently and conveniently.
The establishment of the three major pillars of chips, systems, and services is only the starting point. As AI accelerates into the stage of large-scale application, the key to industrial competition lies not only in the strength of single-point capabilities, but also in whether the entire system can iterate quickly and strengthen itself in real scenarios, and more importantly, in the complementarity of the ecosystem. Chips need real scenarios for verification and optimization, systems need extensive adaptation requirements for polishing and evolution, and services need industry templates and developer communities for precipitation and expansion. To achieve this positive cycle, it is necessary to build an ecosystem: providing a testing ground and feedback loop for chips, so that hardware iteration can accurately respond to real needs; gathering diversified adaptation standards and optimization directions for systems to reduce collaboration costs in heterogeneous environments; precipitating industry know-how and application templates for services to shorten the transformation path from computing power to value.
The ecosystem is the hardest bone to chew in China's computing industry, but it is also the key to industrial collaboration. The Zhongzhi FlagOS community supports the adaptation of multiple domestic chips and has been verified in multiple large model training tasks, providing an important reference for the construction of the domestic computing power ecosystem. Sugon is one of the representatives that started early and has deep accumulation in the layout of the domestic computing power ecosystem. Relying on the Guanghe Ecosystem, Sugon joins hands with upstream and downstream enterprises in the industrial chain, universities, and research institutes to carry out collaborative technological research and build a full-chain innovation system covering from chips to applications. At present, the Guanghe Ecosystem has gathered more than 6,000 partners, completed more than 15,000 software and hardware adaptations, and precipitated more than 15,000 solutions. From the perspective of the ecological map, the Guanghe Ecosystem covers the full chain from BIOS, components, complete machines, operating systems, databases, middleware, cloud computing, big data, AI, network security, high-performance computing to industry applications, forming a complete industrial map from underlying hardware to upper-layer applications. This full-stack ecological layout with real landing capability is rare in the field of domestic computing power.
03. The Next Five Years: What Will China's Computing Power Look Like
The core of the "3+1" path of advancing "chips-systems-services" in parallel and integrating the ecosystem lies in: letting chips, systems, and services no longer fight alone, but forming a feedback closed loop through the ecosystem, enabling the entire industry to iterate quickly and strengthen itself in real scenarios. Following this path, the industrial picture of the next five years is gradually becoming clear.
First, ultra-large-scale clusters will become an important form of computing power infrastructure. The deployment of 100,000-card-level clusters may soon become popular, but the real technical barriers lie not only in the number of cards, but in the systematic efficiency of the cluster. The key lies in whether the network topology can support non-blocking communication, whether memory pooling can break the single-card video memory bottleneck, whether optical interconnection can break through the bandwidth ceiling, and whether overall energy efficiency can be continuously optimized. As the global AI industry enters the inference era, the competition indicators of computing power centers may shift from peak computing power to "Token output per watt". Domestic chips do not need to simply benchmark international similar products in single-point performance, but can find differentiated space in system-level optimization, using better interconnection, smarter scheduling, and lower energy consumption to make up for the single-card gap.
Second, the service-oriented computing power will accelerate and deepen, ultimately falling on the breakthrough of application capabilities. In the next five years, the mainstream form of computing power delivery will expand from one-time hardware sales to service-oriented operations such as on-demand leasing and pay-as-you-go billing. The popularization of the Token factory model means that computing power will be closer to the "on-demand" utility attribute. But the end point of service is not to deliver computing power itself, but to deliver computing power value. Application capability is the last mile in the development of China's high-performance computers. In this regard, we have seen this year that Sugon has basically successfully validated this business model.
Third, the ecosystem will shift from fragmentation to a unified base. The industry needs to promote the establishment of a more unified software stack and standardized interfaces, so that models can be deployed with one click and computing power can be used out of the box. This path is difficult to simply copy the 20 years of accumulation of the existing international ecosystem. A more realistic way is to force the convergence of standards through the form of ecological organizations, and have the platform side uniformly complete the underlying adaptation and middleware packaging to reduce the adaptation cost of the whole society. In this regard, we see that the Guanghe Ecosystem, which has a strong domestic ecosystem, is practicing this goal and benefiting greatly from it.
04. Conclusion
Everyone is talking about the "Golden Five Years" of China's computing power, but I think these five years are not waited for, but laid out. As the full industrial chain capabilities from core components to complete machines, and from system software to industry applications become increasingly mature, and with the continuous surge of the AI industry, on the one hand, the huge domestic market demand can allow localized deployment and scenario-based application of computing power; on the other hand, China's computing power will also usher in a window period to move towards the international market. From Southeast Asia to the Middle East, from Africa to Latin America, the digitalization process of emerging markets needs cost-effective and highly adaptable computing power infrastructure, which is exactly the unique competitiveness formed by China's computing industry after 30 years of local tempering.
Thirty years ago, Sugon-1 broke through the long night and illuminated the journey of China's high-performance computing; thirty years later, what the computing power industry needs is no longer another supercomputer, but an independent and efficient system. Only when the flywheel of the ecosystem truly turns and real applications truly run can China's computing power change from a participant to a contributor on the world stage. This is not an easy road, but it may be the only right one. Walking with the times, we will eventually achieve greatness!