In the future, as the computing power internet becomes fully integrated, upstream production capacity will be gradually released.

Author and source: Semiconductor Industry Verticals (ICViews)

In March 2026, the Outline of the 15th Five-Year Plan was officially released, incorporating the directive to "deeply advance the East Data West Computing project and build a multi-level computing power infrastructure system and a national integrated computing power network" into the nation’s highest-level planning document. In April 2026, the Central Political Bureau meeting proposed the construction direction of "six networks," including the computing power network among them (water network, new power grid, computing power network, next-generation communication network, urban underground pipeline network, and logistics network).

Around the same time, market rumors suggested that China plans to invest approximately 2 trillion yuan over five years to build a nationwide AI computing power network. In response, the National Development and Reform Commission stated at a press conference in June 2026 that it had not confirmed this specific figure, but clearly emphasized, "During the specific implementation process, market forces will play a decisive role; we will work with relevant departments to strengthen overall coordination and policy guidance." The strategic importance of computing infrastructure is now unquestionable.

From the 15th Five-Year Plan to the Computing Power Internet

The computing power network is a systematic project driven by national strategy.

The "15th Five-Year Plan" places it within the chapter on building a Digital China, calling for the promotion of cloud-edge-end collaboration and intelligent computing cloud services to reduce computing costs across society. Supporting policies continue to be implemented: The "Action Plan for High-Quality Development of Computing Infrastructure," issued by six ministries including the Ministry of Industry and Information Technology, proposes that by 2025, computing capacity exceed 300 EFLOPS, with intelligent computing accounting for 35%; the "Action Plan for Computing Interconnectivity," issued by the Ministry of Industry and Information Technology in May 2025, explicitly aims to achieve basic nationwide standardization and interconnection of public computing resources by 2028, forming a computing internet.

The rapid implementation of policies is translating into tangible infrastructure achievements. In May 2025, China Telecom, China Mobile, China Unicom, and the China Academy of Information and Communications Technology jointly launched the construction of a trial network for computing power internet. By the end of March 2026, China’s intelligent computing power capacity reached 1.882 million PFLOPS, 2.5 times that of the same period last year; by the end of 2025, 145 backbone optical fiber cables had been built connecting the eight major computing power hubs, forming a spatial layout of "8+10+3"—comprising eight national computing power hubs, ten national computing power clusters, and three coordinated computing-power-and-electricity development regions.

From a 100,000-GPU cluster to a world-class computing power center cluster

Under this context, the construction of intelligent computing centers is thriving.

On July 9, 2026, the National Supercomputing Internet core node in Zhengzhou officially commenced full-scale operations, launching China’s first ultra-large-scale AI and high-performance computing resource pool with over 100,000 accelerators, capable of providing more than 100,000 domestic AI computing units to external users. The node was deployed at an unprecedented pace: trial operations began in February with the deployment of three Sunway scaleX clusters, each with 10,000 accelerators; by April, a 60,000-accelerator AI for Science computing cluster went live; and by July, total computing capacity exceeded 100,000 accelerators. The node employs the domestic ParaStor F9000 all-flash storage system, which ranked first globally in both the Production Full Node and 10-Node categories on the IO500 list at the 2026 ISC High Performance Computing Conference in Germany.

On July 10, Sugon announced the completion and deployment of China’s fully domestic 100,000-GPU AI supercluster, the Sugon 8000 (Dengfeng), which has been connected to the National Supercomputing Internet. Over 300 optimizations for supercomputing and AI integration have been completed, covering more than twenty cutting-edge fields including large models and robotics. A second system of the same scale has already begun development.

The three major telecom operators are also accelerating their computing power deployment. China Mobile’s intelligent computing scale has reached 92.5 EFLOPS (FP16, by end of 2025), enabling full-range computing capabilities from hundreds to over ten thousand GPUs; China Unicom’s intelligent computing scale has reached 45 EFLOPS, with over 1.1 million standard racks and seven 100-megawatt AIDC campuses built; China Telecom’s total intelligent computing scale, including its own and connected resources, has reached 91 EFLOPS (as of end of 2025), aggregating over 10 trillion tokens of general-purpose large model training data and high-quality datasets spanning more than 14 industries, totaling over 500 TB.

Beyond telecom operators, regional intelligent computing centers are also expanding. Among all regional computing deployments, Inner Mongolia is emerging as a world-class computing center cluster, leveraging its cool climate and abundant wind and solar resources. Tech giants such as Huawei, Tencent, and Alibaba, along with the three major telecom operators and financial institutions like the Agricultural Bank of China, have all established operations there. Companies such as Moonshot AI and DeepSeek are training AI models on servers in Hohhot and other parts of Inner Mongolia; Sinopec is training its industrial large language model "Great Wall" there; and electric vehicle manufacturers like XPeng are also training autonomous driving algorithms in the region.

Significant orders have also been densely implemented in other regions. The Hubei Mobile Computing Network Jianghan Plain Big Data Center (Phase I) launched its EPC tender in June 2026, with a total investment of RMB 88.8231 million; the Qinghai Fangzhou Cloud Green Computing Center has a budget of RMB 1.004 billion and was won by China Construction Eighth Engineering Division, leveraging Qinghai’s clean energy to build efficient and low-carbon computing infrastructure. Demand from leading internet companies continues to surge: ByteDance projects domestic IDC demand of approximately 1.5 GW in 2026, while Alibaba plans for 2 GW, having already initiated its first tender for 500 MW in April. Inference computing has become a new frontier: China Mobile spent RMB 5.112 billion in the second half of 2025 to procure 7,058 inference-focused AI servers, setting the record for the largest single AI server purchase of the year.

Expansion of the entire chain for optical modules, PCBs, power chips, and liquid cooling

The rapid expansion of computing power infrastructure is quickly propagating upstream along the supply chain. From optical modules to PCBs, from power chips to liquid cooling systems, upstream segments are universally experiencing overwhelming order volumes and urgent capacity constraints.

Optical modules serve as the "highway" of computing power networks and are currently undergoing a generational transition from 400G to 800G and 1.6T. Tongling Xuchuang Optical Communication Technology Co., Ltd. plans to invest in the construction of the "Tongling Xuchuang High-Tech Zone Optoelectronic Industrial Park Phase I Project" and the "Tongling Xuchuang High-Tech Zone Optoelectronic Industrial Park Phase II Project." The total investment for both phases amounts to RMB 1.12 billion, with full-scale production scheduled for July 2026, achieving an annual output of 8 million high-speed optoelectronic devices supported by a workforce of 5,000. Dongshan Precision has acquired Solus Optoelectronics, which is implementing an expansion project for optical chips and optical modules in Changzhou and other locations with a total investment of $1.2 billion (approximately RMB 8.1–8.7 billion). XinYisheng is continuously expanding its production capacity throughout Q2–Q4 this year, with monthly capacity ramp-up, adopting a lightweight asset expansion model primarily based in Thailand, supplemented by facilities in Chengdu and Wuhan, focusing on high-end 1.6T LPO/silicon photonics computing modules. Huagong Technology is constructing facilities worth RMB 271 million, with total fixed assets reaching RMB 3.095 billion; fixed assets increased by RMB 291 million quarter-over-quarter in Q1, primarily due to the completion and transfer to fixed assets of the Wuhan campus, Phase I plant in Thailand, and automated optical module production lines. Hengtong Optoelectronics is building an RMB 8 billion super optical module base in Suzhou, with Phase I scheduled for completion and commissioning by the end of 2026, expected to deliver an annual capacity of 4 million 800G and 1 million 1.6T silicon photonics modules, enabling flexible co-production switching between 800G and 1.6T.

Demand for PCBs in AI servers is experiencing a surge in both volume and price. The current AI server architecture has transitioned from traditional CPU-based designs to GPU and ASIC computing clusters, imposing stringent requirements on PCB layer count, substrate performance, via structure, trace precision, and overall stability—directly driving market demand for high-layer-count and advanced HDI boards. Huaqin Technologies’ Thailand production base has completed its capacity ramp-up and has officially entered a phase of large-scale, efficient operations, enabling it to fully承接 high-value-added PCB orders for AI servers and high-speed communications. Pengding Holdings plans to raise no more than RMB 9.6 billion, all allocated to the Huai’an Qingding project for high-density interconnect (HDI) build-up boards targeting AI servers and high-speed optical modules, aggressively expanding high-end computing PCB capacity. Shenghong Technology is focused on high-value-added PCB segments including AI computing devices and data center switches; its four newly constructed factories in Huizhou have been rolled out in phases and are currently in the capacity ramp-up and mass production stage. Zhongjing Electronics plans to raise funds to advance two major projects: an intelligent PCB production base in Thailand and an expansion of high-density PCB capacity in Zhuhai Fushan, with the Thailand facility expected to commence initial production in the second half of this year. Shennan Circuits is raising capital through a private placement, primarily investing in the Wuxi AI computing electronic circuit products project, focusing on R&D and manufacturing of high-speed, high-density, high-layer-count PCBs tailored for AI hardware upgrades. Dongshan Precision’s legacy plant renovation project for its PCB business has been successfully commissioned, and new high-end capacity is being steadily implemented according to plan, with its production structure continuously shifting toward high-end sectors such as AI and high-speed communications.

AI server power chips are essential components for computing hardware. Data shows that global AI server shipments are expected to reach approximately 2.46 million units in 2025, a 24.3% year-over-year increase. Industry data indicates that the AI server power market will experience exponential growth from 2025 to 2027, expanding from $7.4 billion to $32.5 billion, with a compound annual growth rate of 110%. Previously, the high-end AI server power chip market was long dominated by overseas giants. Amid ongoing price hikes and extended delivery cycles for overseas chips, leading OEMs and cloud providers are increasingly prioritizing supply chain security and actively accelerating the adoption of domestic power chips. JHW's 90A high-current intelligent power stage is suited for high-performance computing power applications such as servers and GPUs. SG Micro has broken through key AI power chip technologies including 90A high-current DrMOS, 48V DC-DC converters, and high-precision monitoring AFEs, and has entered the supply chains of Ascend, leading domestic server ODMs, and high-speed optical modules. MaoRuiXin focuses on high-current intelligent power stages and has launched MK6841, MK6850, and other 90A high-power DrMOS products featuring advanced packaging and optimized topologies that significantly enhance power density and thermal performance while reducing PCB area—precisely meeting the dense power demands of AI servers. To meet industry capacity demands, specialty process semiconductor manufacturer UNICORN Integrated Circuit announced a new project to expand production, constructing a 12-inch mixed-signal chip production line with a monthly capacity of 50,000 wafers. UNICORN stated that the launch of Phase IV marks its systematic expansion into two high-growth sectors—AI server power and optical interconnect—building upon its established strengths in electric vehicles and industrial control, forming a dual-driver model of "core business stability and emerging sector expansion." UNICORN will further increase production capacity for high-end power management chips to address the surging global demand driven by AI computing growth.

As rack power consumption surges from 6kW to 18kW and beyond 100kW, and new intelligent computing clusters demand a PUE of ≤1.25, liquid cooling has become a necessity. China Mobile’s Hohhot Data Center has completed the world’s largest single-site liquid-cooled intelligent computing center operated by a telecom provider, with a total investment of RMB 4.66 billion. The facility deploys approximately 20,000 AI accelerator cards, achieving over 85% domestic chip adoption, including 19,800 training cards and nearly 10,000 inference cards from five domestic chip manufacturers: TianShu ZhiXin, Huawei Ascend, Birun, Kunlun Core, and others. The total intelligent computing capacity reaches 6.7 EFLOPS (67 quintillion floating-point operations per second). The project employs cold plate liquid cooling technology and intelligent busbar systems to achieve a PUE as low as 1.15 and utilizes renewable energy for 55% of its power consumption.

Hash power is national power; chain connectivity is the future.

From the inclusion of the computing power network in the "15th Five-Year Plan" as a national major project, to the official launch of the National Supercomputing Internet’s core node with a 100,000-GPU cluster; from the rise of Inner Mongolia’s world-class computing power center cluster, to the 88.82 million RMB EPC tender for Hubei Mobile’s Jianghan Plain Big Data Center and the successful implementation of Qinghai Fangzhou Cloud’s 1.004 billion RMB green computing power center; from Yunnan Germanium’s expansion of 450,000 indium phosphide substrates, to Sanan Optoelectronics’ mass production of 6-inch InP photonic chips—China’s AI industry chain is undergoing a profound transformation fueled by three key drivers: policy support, demand pull, and domestic substitution. In the future, as the computing power internet becomes fully interconnected and green computing hubs like Inner Mongolia continue to expand, upstream production capacity will gradually be unleashed.

This article is from the WeChat public account "Semiconductor Industry纵横" (ID: ICViews), author: Pengcheng, republished with permission from 36Kr.