The Systemic Leap of Domestic Computing Power from “Building More” to “Using It Well”

The week of September 11–17, 2026, brought a rare “dense catalyst window” for the computing power field. On September 11, the State Council executive meeting specially deployed the construction of a computing power network; on the same day, Enflame Technology listed on the STAR Market, and the “four little dragons” of domestic GPUs all gathered in the capital market. On September 12, the 2026 China Computing Power Conference opened in Langfang, and the nationwide computing power “single network” was declared basically formed. On September 15, Huawei released the world’s first 3D data center. On September 16, MetaX and Moxing Intelligent announced that a 10,000-GPU cluster had officially entered the production environment of a “Token Factory.” Policy direction-setting, capital deployment, technological breakthroughs, and scenario implementation—four signals superposition—point to one core judgment: China’s computing power industry is accelerating from a stage of “scale expansion” to a stage of “systemic capability competition.”

The most direct driver of this shift comes from structural tension on both the supply and demand sides. Data from the China Academy of Information and Communications Technology (CAICT) show that in the first quarter of 2026, domestic AI computing power demand grew 417% year-on-year, while effective supply grew only 128%, and the gap continues to widen. At the same time, the focus of large-model competition is shifting from training to inference. The China Telecom Research Institute predicts that the share of inference computing power will rise to 80% by 2029, and annual token consumption will jump from the 10^17 level in 2026 to the 3.5×10^19 level in 2030, with a compound annual growth rate of nearly 12 times. This structural change in demand is forcing computing power infrastructure to move from “stacking cards” to “weaving a network.”

Capital Convergence: Domestic Chips Move from “Can They Be Built” to “Can They Deliver”

The landmark capital market event this week was that the “four little dragons” of domestic GPUs all listed on the capital market within less than a year. On September 11, Enflame Technology officially began trading on the STAR Market at an issue price of RMB 142.18 per share, corresponding to a market value of more than RMB 61.1 billion and raising about RMB 6.1 billion. At this point, Moore Threads, MetaX, Biren Technology, and Enflame Technology—four domestic computing power chip companies—completed initial technological verification and capitalization, and competition entered a new stage.

The revenue data of the four companies confirm that domestic AI chips are entering a performance delivery phase. In the first half of the year, Moore Threads achieved operating revenue of RMB 1.736 billion, up 147.42% year-on-year; MetaX achieved operating revenue of RMB 1.324 billion, up 44.67% year-on-year; Biren Technology achieved operating revenue of RMB 1.2355 billion, up 1,997.6% year-on-year; Enflame Technology achieved operating revenue of RMB 1.12 billion, up 279% year-on-year. Although none has yet crossed the profitability line as a whole, some companies have significantly narrowed losses. MetaX’s non-recurring-adjusted net profit attributable to the parent company in the second quarter was RMB 54 million, achieving non-recurring-adjusted profitability in a single quarter. MetaX Chairman Chen Weiliang previously predicted that the company could break even as early as 2026; Enflame Technology Chairman Zhao Lidong also said during the IPO roadshow that consolidated statement profitability is expected in 2026 or 2027.

However, the capital market’s attitude has gradually shifted from initial enthusiasm to scrutiny of performance delivery capability. All three already-listed companies experienced rapid valuation surges after listing, then generally retreated by about 40%–50% from their intra-year highs. Wang Peng, a researcher at the Beijing Academy of Social Sciences, believes that what lies ahead is systemic competition: large-scale delivery capability, full-stack hardware-software adaptation efficiency, rapid product iteration on an annual basis, and building dedicated advantages in specific industries. “Only companies that can first string together R&D investment, product performance, software ecosystem, and customer orders into a continuously profitable business chain can truly gain a firm foothold in the industry.”

The four companies also show a differentiated pattern in technology routes. Moore Threads, MetaX, and Biren Technology adopt the GPGPU route, compatible with Nvidia’s CUDA ecosystem, with relatively low developer migration costs; Enflame Technology chooses the DSA route and self-develops the TopsRider software platform. This differentiation itself is a sign that domestic computing power is maturing—no longer homogeneous “domestic substitution,” but a search for differentiated competitive advantages in different technological directions.

Computing Power “One Network”: From Isolated Data Centers to Full-Domain Scheduling

While capital focused on chip companies, the “networking” of computing power infrastructure also achieved a key breakthrough this week. At the opening ceremony of the 2026 China Computing Power Conference, the China Computing Power Platform announced the realization of nationwide integrated computing power coordination and monitoring. The platform is orderly promoting the connection of platforms in 31 provinces, autonomous regions, and municipalities to the network, and a unified, standardized, collaborative, and efficient national computing power situational awareness and monitoring system has basically taken shape.

At the conference, multiple academicians and business leaders unanimously proposed that the core of the computing power network is “connection,” not merely “construction.” Liu Yunjie, an academician of the Chinese Academy of Engineering, pointed out that the traditional point-to-point interconnection model should be abandoned, and a high-speed switched flattened backbone network should be built. Relying on optoelectronic convergence technology, a “computing power high-speed rail” should be built to achieve second-level computing power scheduling. Zhu Hanwu, vice president of China Unicom, introduced that China Unicom has fully deployed intelligent computing gateways with business-awareness capabilities and, relying on the “Unicom Xingluo” advanced computing power scheduling platform, completed cross-domain heterogeneous hybrid training verification over more than 2,000 kilometers, “truly moving ‘East Data, West Computing’ from blueprint to daily operation.”

The speed of local implementation of this “networking” movement is also noteworthy. Langfang has built 36 data centers, with total intelligent computing scale exceeding 200,000 PFLOPS, and its urban computing power sub-index has ranked first nationwide for two consecutive years. The country’s first provincial integrated computing power monitoring and scheduling platform to go online was put into use here; it took the lead in the province in launching computing power vouchers and is planning a “four-voucher linkage” of computing power, tokens, data, and models. Hebei innovatively introduced support measures such as computing power vouchers, token vouchers, model vouchers, and a “computing power+” industry fund, attracting nearly 200 outstanding upstream and downstream enterprises to land in Hebei, with the province’s computing power utilization rate exceeding 80%. Langfang also set up a RMB 2 billion “computing power+” industry fund to build a full-chain support system. Chengdu High-tech Zone offers subsidies of up to “RMB 200 million in computing power vouchers plus RMB 400 million in model vouchers” annually, targeting the core pain points of AI enterprises in computing power scheduling and model invocation in the era of large models.

The Ministry of Industry and Information Technology recently issued the 15th Five-Year Plan for the Development of the Information and Communications Industry, explicitly proposing to accelerate the construction of a nationwide integrated computing power network, build a multi-level computing power facility system of “hub-region-edge,” and deepen automated monitoring of computing power and supply-demand matching. The ministry also explicitly explores innovative businesses such as “computing power banks” and “computing power supermarkets.” The National Supercomputing Internet Platform has gathered more than 3.5 million CPU cores and 250,000 GPU cards, with more than 1.4 million registered users and over 7,300 application services connected. From “building centers” to “weaving a network,” and from “selling computing power” to “issuing computing power vouchers,” the operating model of computing power infrastructure is undergoing a profound paradigm shift.

Supernodes: The New Main Battlefield of Computing Power Competition

The most noteworthy industry signal this week was that “supernodes” moved from concept to large-scale deployment. Among the 15 “Annual Major Breakthrough Achievements in Computing Power China” announced at the 2026 China Computing Power Conference, domestic cluster supernodes and 10,000-GPU intelligent computing clusters were listed as core areas. A research report by Western Securities directly judged that “2026 is expected to officially become the first year of supernode deployment.” Traditional 8-GPU servers can no longer meet the communication needs of training and inference, and supernodes have become a “must-have” for computing power vendors.

The reason supernodes have become a focus is that they fundamentally change the physical form of computing power infrastructure. Gao Wen, director of Pengcheng Laboratory, gave a clear definition: a supernode is a computing system physically composed of multiple computing nodes tightly connected through efficient interconnect protocols, with unified memory addressing across physical nodes, and logically possessing the characteristics of “one computer.” This definition clarifies a key difference: a supernode is not simply “stacking cards,” but a system-level solution involving collaborative optimization of computing power, bandwidth, interconnection, management, and software stack.

The engineering complexity of implementation far exceeds imagination. Chen Zhenkuan, vice president of Lenovo Group, revealed in an industry exchange that in the past, a server production line needed a load-bearing capacity of about one ton per square meter, while supernode production “starts at three tons per square meter, and some even need to be designed for five tons.” Equipment must be placed on the first floor, and “there can be no underground garage below.” A supernode production base needs at least 30 MW or even higher power supply and requires local 110 kV transformer facilities. This means that the ticket of entry for computing power competition has upgraded from “whether there are GPUs” to “whether it is possible to solve the system engineering involving factory floor load-bearing, power supporting facilities, liquid cooling, and high-speed interconnection.”

Standards are another hurdle that must be crossed. In the Scale-up interconnect field alone, there is a situation of “12 vendors and 13 protocols,” including NVLink, C-Link, UB-Link, OISA, UA-Link, UEC, XPU-Link, and others. Hardware vendors need to be compatible with a multi-protocol ecosystem and match the complete-machine requirements of various supernodes. The OISA omnidirectional intelligent sensing interconnect architecture led by China Mobile is trying to provide a unified open protocol foundation for supernode Scale-up interconnection. At the computing power conference exhibition area, X-Epic and China Mobile jointly exhibited the “OISA Domestic Chips, Domestic Interconnect” system, building an end-to-end verification loop for computing chips, switching chips, and interconnect systems based on the OISA protocol, lowering the barriers for multi-type chip access, interconnect adaptation, and system integration in supernodes. Yang Gongyifan, founder of Zhonghao Xinying, summarized four core pain points for large-scale supernode deployment: cluster communication bottlenecks, multi-tenant business isolation and computing power balancing, pressure to upgrade cooling infrastructure, and insufficient hardware-software co-design maturity. Therefore, there is still a long way to go before supernodes reach standardized, stable commercial use. Whether standards can converge will directly determine the speed at which supernodes move from small-scale deployment by a few leading enterprises to large-scale commercial use across the entire industry.

Heterogeneous Architecture: A “Lane-Changing” Approach to Bypassing Process Constraints

While supernodes reconstruct physical infrastructure, another innovation with greater strategic significance is unfolding at the architecture level. At the computing power conference held on September 11–13, Mobile Cloud, together with the Nanhu Research Institute of China Electronics Technology Group Corporation, Lynxi Technologies, Iluvatar CoreX, Tsinghua University, and Peking University, jointly released China’s first domestically produced GPU + neuromorphic chip heterogeneous hybrid inference system for large models.

The core logic of this solution is to “use architectural innovation to compensate for process constraints.” Traditional pure-GPU inference clusters have long been constrained by the “memory wall” and “scheduling jitter,” while the supply of advanced processes in China is constrained, and relying solely on hardware process upgrades to improve computing power has limited room. The project team’s solution is to perform AF separation on the Transformer model: Attention computation is carried by domestic GPUs, while the FFN module is processed by neuromorphic chips, fully leveraging the architectural advantages of neuromorphic chips such as compute-in-memory and large on-chip SRAM, breaking through the inherent bottlenecks of traditional GPU inference.

Measured data verify the feasibility of this technology route. The project has completed full tuning and verification on the DeepSeek V4 Flash large model. Compared with similar domestic GPU computing clusters, inference output and energy efficiency both more than doubled, and business operating costs fell by more than 40%. The team self-developed a model compiler, high-speed interconnect protocol, and unified inference engine to achieve task decomposition, collaborative scheduling, and result aggregation between the two types of computing power. The project has accumulated 15 authorized invention patents, and the technology has reached the small-batch trial production stage. The entire solution “does not rely on overseas high-end chips or advanced processes.”

The significance of this technology route goes beyond a single product. The team will continue to iterate the heterogeneous inference system, carry out implementation verification in scenarios such as token factories, AI code development, text-to-video, multi-agent collaboration, intelligent manufacturing, and financial and communications security, and plans to gradually open the core heterogeneous inference framework to empower domestic GPU and neuromorphic chip vendors and computing power operators. It proves one thing: under constraints on advanced processes, through innovation at the system architecture level, domestic chips are fully capable of matching the business capabilities of the next generation of international heterogeneous inference solutions.

From “Building Well” to “Using Well”: The Last Mile of Computing Power Implementation

If chip listings, supernode deployment, and heterogeneous architecture innovation address the “supply side,” whether computing power can truly create value ultimately depends on the “demand side’s” ability to absorb it. Among the dense industrial information released this week, progress on the application side is equally noteworthy.

Lenovo’s “Computing Power in Action: A Practical Atlas,” released on September 16, includes 15 real-world computing power construction cases, covering manufacturing, healthcare, education, finance, the internet, and other industries. A set of data in the atlas reveals the core contradiction in current computing power applications: a VentureBeat Research survey of 573 enterprise technology leaders showed that more than 86% of enterprises’ self-built GPUs have utilization rates of no more than 50%; a KPMG survey showed that only 26% of organizations have complete, real-time visibility into AI system operating costs. Computing power planning, cost control, and governance capabilities are becoming key bottlenecks for enterprises to scale AI deployment.

In vertical industries, the real effects of computing power implementation have begun to emerge. In finance, the INESA Intelligent Computing Cloud Platform helped intercept more than RMB 200 million in suspicious transactions. In manufacturing, lithium battery industrial AI quality inspection technology compressed quality inspection model debugging and deployment time from the industry-typical 6 hours to 1 hour, and reduced the coating fluctuation indicator in key processes from 0.6 to 0.2. In healthcare, automatic electronic medical record generation shortened inpatient record writing time to one quarter of the original. Data from the National Data Administration show that in June 2026, China’s average daily token calls exceeded 500 trillion. IDC predicts that by 2031, the number of active intelligent agents in Chinese enterprises will reach 350 million, with a compound annual growth rate of more than 135%.

Meanwhile, the “financialization” trend of computing power supply became increasingly clear this week. CAICT data show that in the first quarter of 2026, the domestic computing power leasing market reached RMB 68 billion, and the full-year market is expected to exceed RMB 260 billion. In the first half of 2026, among China’s newly added intelligent computing center projects, the penetration rate of the financial leasing model reached 54.3%. With advantages in funding costs, asset ownership confirmation capabilities, and professional risk control systems, financial leasing companies have occupied an irreplaceable financial infrastructure position in the computing power industry chain.

The Next Stop for Computing Power: From Ground to Space, From Energy Consumption to Symbiosis

While supernodes and heterogeneous architecture accelerate on the ground, another notable direction this week is the outward extension of computing power boundaries. On September 13, Shanghai officially launched the “Pearl Constellation Plan” at the Pujiang Innovation Forum, with the long-term goal of building a “gigawatt-level space-based computing constellation.” It plans to build a globally covered, space-ground collaborative, independent and controllable space intelligent computing infrastructure and service network by 2030. The plan will explore directions such as on-orbit deployment of large models on satellites, intelligent meteorological warning, and ground-agent collaboration. After completing phased technical verification, it will gradually enter the stage of first-batch satellite deployment and constellation networking. Almost simultaneously, four localities in the Beijing-Tianjin-Hebei region jointly launched an initiative to build a space computing power industry corridor. Liu Jingjing, CEO of Dongfang Tiansuan, put it clearly: “Ground computing excels at low latency and high interaction, while space computing excels at wide-area coverage and ‘space-based data, space-based computing.’ Integrated scheduling allows ground and space computing to each play to their strengths.”

The value of space computing lies not in replacing ground computing, but in filling the blind spots of ground networks. Oceans, deserts, polar regions, and other areas that ground networks cannot reach thereby gain access to computing power services. At the same time, space computing can also serve as a strategic backup for ground computing. This direction is still in the technical verification stage and will not change the basic pattern of computing power supply in the short term. But it releases a signal worth watching: the competitive dimension of computing power infrastructure is expanding from “density” to “breadth,” extending from the physical limits of ground data centers to the space resources of low Earth orbit.

Equally strategically significant as space computing is the accelerated implementation of green computing power. Under the “dual carbon” goals, PUE has changed from a “bonus point” for data centers to an “entry threshold.” This week, Shijingshan-based company Chao Zhisuan globally launched the Net Zero Hub. It built an engineering system of “green power direct supply × energy storage regulation × full liquid cooling × 10,000-GPU clusters × waste heat recovery,” achieving a PUE as low as 1.05, a waste heat utilization rate of 100%, and carbon emissions approaching zero. At the industry level, cold plate liquid cooling currently has a market share of more than 80%; newly built intelligent computing centers basically configure it by default, and PUE can be stably kept between 1.15 and 1.25. It is currently the only liquid cooling technology path that can be “delivered ready to use.” Li Li, vice chairman of Runze Technology, offered a more intuitive comparison: after large-scale application of cold plate liquid cooling technology, PUE can reach 1.15; for every 1 kWh used for computing, only 0.15 kWh is consumed for cooling, and the cooling process can save 50% of electricity.

Cool-Headed Thinking Beneath the Boom

Among the dense catalysts this week, there were also calm voices. Wu Chaoze of China Securities Co., Ltd. provided a set of comparative data: in the past two decades of the mobile internet era, annual underlying capital expenditure in China and the United States was about USD 100 billion; this year, capital expenditure by North American CSP vendors is about USD 800 billion, while China’s is about RMB 1 trillion. “The investment scale has expanded 10 times, and including the supply chain, possibly 20 times.” The sharp expansion of capital expenditure has expanded funding sources from corporate free cash flow to bond issuance, financial leasing, long-term leases of third-party data centers, and even private credit. Project financing rates for projects backed by large cloud vendors are about 5.3% to 6.6%, while those without such backing are as high as 8.5% to 9%. A difference of a few percentage points in financing cost is enough to change the ROI of hundred-billion-yuan infrastructure projects. Wu Chaoze also admitted that currently “the computing power industry chain has obtained more than 90% of the revenue and profits of the AI industry chain, while models and applications are still in a stage of rapid competition, with commercialization relatively lagging.”

Discussion of “computing power oversupply” also continued to heat up this week. David Sacks, Trump’s technology adviser, previously used the “dark fiber” after the dot-com bubble burst as a metaphor to warn of the risk of “dark GPU” oversupply in AI data centers. But more industry figures tend to believe that the core contradiction is not total oversupply, but structural mismatch. Some analyses point out that “high-quality computing power remains scarce, while inefficient computing power is becoming oversupplied.” Moderately advanced construction should be more reflected in “slow-variable” infrastructure such as network interconnection, power supporting facilities, and land reservation. For “fast-variable” computing equipment such as GPUs, a strategy of “scaling on demand and building in rolling phases” is appropriate, to avoid resource idling caused by one-time large-scale investment. Tang Zhehui of EY judged that it is not that “computing power is no longer scarce,” but that valuation models are shifting from “capacity panic” to “input-output efficiency.”

Ultimately, the core narrative in the computing power field this week is not a single product launch or a single technological breakthrough, but a systemic signal: domestic computing power is bidding farewell to the “whether we have it” stage and entering a hard battle over “whether it is easy to use and cost-effective.” Supernodes test engineering delivery capabilities; heterogeneous architecture tests system-level innovation capabilities; space computing tests strategic patience and long-cycle investment capabilities; and green computing tests the sustainability of continued expansion under energy consumption constraints. Together, these four lines constitute the transformation trajectory of China’s computing power industry from “building more” to “using well and coordinating well.” On this trajectory, the enthusiasm and calmness of capital, the aggressiveness and prudence of technology, and the support and regulation of policy will continue to clash in the coming months, jointly shaping the next stage of China’s computing power industry.