Economic Observer reporter Zheng Chenye

Entering the third quarter report release period, domestic computing chip companies have successively submitted their performance reports, with generally impressive growth.

For example, on the evening of October 15th, Muxi Corporation (688802.SH) released a performance forecast, expecting a year-on-year increase of 105.60% to 151.29% in revenue for the first three quarters, and a turnaround in net profit attributable to the parent company; On the evening of October 19th, Haiguang Information (688041. SH) released its third quarter report, showing a year-on-year increase of 57.58% in revenue for the first three quarters and a new high of 5.856 billion yuan in revenue for the third quarter. Cambrian (688256. SH) and Moore's Thread (688795. SH) have not yet released their third quarter reports, but their revenue for the first half of 2026 also increased by 108.13% and 147.42% year-on-year, respectively.

Behind the impressive performance, there is not only a rapid growth in demand for AI computing power and manufacturers' own technological accumulation, but also external environmental support - in the past few years, domestic substitution has continued to advance, and US export controls have restricted the sales of Nvidia's high-end chips to China. The overseas competition for domestic chips is relatively limited, and orders are also relatively guaranteed.

At the same time, the coordination between domestic large models and domestic computing chips is becoming increasingly close, and the gap in model adaptation among chip manufacturers is also narrowing. After DeepSeeker R1 is open sourced in January 2025, most domestic chips will be gradually adapted after more than ten days; In September of the same year, DeepSeeker V3.2-Exp was open sourced, and Huawei Ascend, Cambrian, Haiguang Information, and others were able to complete the adaptation on the same day.

On October 15, 2026, the big model company Step 5 Star opened up the new generation model Step 5 Preview, and domestic chips such as Huawei Ascend, Moore Thread, Muxi Shares, Boren Technology (06082. HK), and Haiguang Information's DCU (Deep Computing Processor) were all adapted on the same day. The industry refers to this type of adaptation that is completed on the day the model is open sourced as "Day 0 adaptation", and this situation has gradually become popular among top domestic chip manufacturers. This also means that the differentiation advantage brought by model adaptation is gradually narrowing, and the competition among domestic manufacturers is becoming increasingly fierce.

From the interviews with reporters, it can be seen that the competition for domestically produced computing chips is extending from the chips themselves to various aspects of computing power implementation, such as software migration, system efficiency, and supply guarantee. Or in other words, the stage where domestic computing chips relied on external environments to "win by lying down" is coming to an end.

Software Migration

The 'Day 0 adaptation' is mainly applicable to newly released models. The competition at the software level has not ended here, and a large amount of manpower is still needed to migrate existing applications to domestic computing platforms.

On the open source day of Step 5 Preview, Huawei Ascend and Boren Technology were both adapted based on open source inference frameworks. After the basic software of chip manufacturers is integrated into such frameworks, the main task when the new model is released is to supplement the operation modules (known as operators in the industry) used in the new model, and then optimize the accuracy and performance. But the situation of existing applications is much more complex, as many industry users' software was initially developed on foreign hardware and software environments.

Bu Jingde, Chief Engineer of High end Computing at Zhongke Shuguang (603019. SH), told Economic Observer reporters that in recent years, with the improvement of the company's basic software, mainstream software environments for scientific computing have been basically supported, and the workload of adaptation has been greatly reduced. However, traditional CPU based applications still face certain challenges and workloads in migration and adaptation.

According to Bu Jingde's introduction, commercial software does not have source code, making adaptation difficult. The source code and development team are in China, making cooperation relatively simple; The algorithm written by the research team itself may take some time to get started when transplanted, and deep adaptation and optimization often require a team of dozens of people, some close to one or two hundred people.

The user experience is more direct. Zhang Meng, Chief Expert of Sinopec Shengli Oilfield Geophysical Research Institute, stated in an interview with reporters that the institute's original self-developed exploration software was all based on foreign commercial hardware, and moving to a domestic platform "is not simply taking things from the left hand to the right hand". The program architecture, language, and interface all need to be modified.

Lowering the threshold for migration cannot be achieved without the continuous investment of vendors in basic software. For example, in August 2025, Huawei's rotating chairman Xu Zhijun announced that Ascend's basic software CANN (software that connects Ascend chips with upper level AI frameworks) would be fully open source and open to the public.

A person in charge of a leading domestic chip manufacturer told reporters that the increasing variety of computing hardware and software models will give rise to specialized middleware service providers that can adapt between different hardware and software. In his view, new forms such as AI Infra (AI infrastructure software) have emerged in this way. What chip manufacturers can do is to ensure the most basic part, and the software ecosystem still needs to rely on more partners to improve together. According to its introduction, the proportion of domestically produced software and hardware is still relatively low in industrial manufacturing and other fields.

Cao Zhennan, Vice President of Zhongke Shuguang, also told reporters that in order to make hardware more user-friendly, Zhongke Shuguang has invested heavily in software personnel, mainly focusing on the system software level. He believes that building a computing power platform based on domestic chips requires a sufficiently thick software stack, including various libraries, development environments, and even compilation tools. In addition, computing power systems require continuous iteration, and hardware is often slightly faster than software and applications. With a hardware platform, application software can develop.

In Cao Zhennan's view, in addition to performance, scale, and cost-effectiveness, whether it is easy to use and whether the application threshold is high may become a key indicator for measuring domestic computing power in the future.

System Efficiency

In addition to software, the efficiency of computing power systems themselves is also diverging. The efficiency of a system composed of the same number of chips may vary greatly, and this difference may further widen as large-scale clusters gradually become a common form of deployment for large models.

Wan Wei, Chief Engineer of Zhongke Shuguang High speed Network Interconnection Product Department, told reporters that as the model becomes larger, eight computing cards (a common configuration of a single AI server) are no longer able to accommodate a complete model, and a distributed architecture is needed, which is jointly carried by multiple nodes. Relevant data needs to be frequently exchanged between nodes.

Whether it is training or inference, multi card parallel computing relies on communication synchronization. Once some computing cards experience abnormal delays, it may cause other computing cards to wait, thereby dragging down the computing efficiency of the entire cluster. In Wan Wei's view, the development of large-scale AI computing towards clustering is an inevitable trend, and the collaborative efficiency of computing, networking, and storage is a key factor affecting the performance of the entire cluster.

A semiconductor analyst from a large securities firm in southern China also told reporters that supernodes (systems that connect a large number of AI chips through high-speed interconnection for collaborative work) have gradually become the threshold for deploying large models, and the configuration required for inference has been upgraded from last year's eight card all-in-one machine to supernodes.

For example, in July, the parameter scale of the open-source Kimi K3 reached 2.8 trillion. The official recommendation of the Dark Side of the Moon is to use a supernode composed of 64 or more accelerators for local deployment. The analyst believes that supernodes have reached a turning point of mass production in the third quarter of 2026, and the penetration rate is expected to accelerate in 2027.

He also stated that the value of supernodes lies in improving computational efficiency through innovative system architecture. Due to the denser integration of computing by supernodes, the memory of the chip can also be addressed uniformly and accessed directly. Compared with traditional horizontal expansion methods (expanding the scale by increasing the number of servers), communication loss can be reduced from 30% -40% to within 10%, computing power utilization can be increased from 60% -70% to over 90%, and the training cycle of trillion parameter level models can be shortened from months to weeks.

In its view, as there is still a gap in single card performance between domestic and overseas chips, this is a path for domestic computing power to catch up. He also stated that the capital expenditure of cloud vendors is expected to gradually shift from simply purchasing chips to purchasing complete systems, and the delivery of supernodes is also upgrading from single servers to complete cabinets or even multi cabinet levels, involving computing, networking, and cooling power supply, resulting in a significant increase in the complexity of the solution. In this case, market share will be more concentrated towards the top server manufacturers.

In fact, this change has already been reflected in the procurement of cloud vendors. For example, on September 9th, Cao Peng, Chairman of the Technical Committee of JD Group (09618. HK) and President of JD Cloud, announced that JD Cloud plans to build a 100000 card intelligent computing cluster, using Moore's Thread GPU and opening up computing power to the entire industry. Cao Peng stated that JD Cloud is interested in the mature delivery experience of Moore's Thread on large-scale clusters.

In addition, data from market research firm IDC shows that in the first half of 2026, Nvidia's market share in the Chinese data center AI chip market has dropped to 34%, while Huawei's market share has reached 28%. IDC also believes in the report that competition in the Chinese AI chip market is shifting from the performance of individual chips to the efficiency of large-scale computing systems.

Of course, the expansion of cluster size also puts higher demands on the stable supply of chips.

Supply Capacity

AI chips cannot do without high-performance storage, and storage chips continue to be in short supply. As a result, the stability and continuity of supply have gradually become a competitive advantage for domestic computing chip manufacturers.

Renowned market research firm Jibang Consulting recently predicted that the contract price of general DRAM (i.e. conventional memory) will increase by 10% -15% month on month in the fourth quarter of 2026. DRAM manufacturers continue to invest in advanced processes for high-performance products used in servers, and the overall market remains in short supply, although the upward trend has slowed down; In addition, the overall average selling price of HBM (high bandwidth memory paired with AI chips) in 2027 can increase by 121% year-on-year. The organization believes that the demand for AI servers continues to expand, and HBM and general DRAM are still competing for limited advanced process and wafer production capacity.

The relevant person in charge of the aforementioned chip manufacturer also told reporters that from wafers to substrates, the upstream of chips is facing pressure from rising prices. Currently, chip production capacity is affected by external supply chain fluctuations, with memory being the most affected. According to its introduction, chip products have already taken this factor into consideration during the design phase, and future products will try to be compatible with more types of memory without increasing or increasing costs, in order to cope with the supply and price fluctuations of memory.

The aforementioned semiconductor analysts also believe that storage supply is still one of the main bottlenecks currently restricting domestic AI chips. With the technological progress and production increase of domestic storage manufacturers, this bottleneck is expected to gradually ease, thereby driving the growth of domestic computing chip shipments.

Domestic DRAM manufacturer Changxin Technology (688825. SH) recently announced that its new DRAM architecture has made progress, and the new generation of server memory is expected to be launched by the end of the year, with storage density expected to increase. For now, the company stated in its performance forecast for the first three quarters of 2026 that the demand for AI computing power is rapidly increasing, DRAM products are in short supply, prices have risen significantly, and the company's capacity utilization rate remains high.

Increasing inventory is a direct response from many chip manufacturers.

The relevant person in charge of the aforementioned chip manufacturer explained that their company's stocking plan often needs to be formulated one year in advance, and the next year's shipment volume is predicted based on market research. Stocking cannot be blind, but if there is insufficient stocking and production capacity cannot keep up, market supply will be tight.

According to financial report data, multiple domestic computing chip manufacturers are increasing their inventory. As of the end of June 2026, the inventory of Cambrian increased by 66.83% compared to the end of the previous year, mainly due to raw materials and commissioned processing materials; During the same period, the inventory of Moore Thread increased from 1.332 billion yuan at the end of 2025 to 3.55 billion yuan.

As of the end of September 2026, Haiguang Information's prepayments (payments made in advance to suppliers) increased from 2.885 billion yuan at the end of 2025 to 5.885 billion yuan, and inventory increased from 6.406 billion yuan to 8.921 billion yuan. The company stated in its third quarter report that in order to match market sales planning, ensure supply capacity, and seize market opportunities, the company has increased its stocking scale according to its business strategy, and the expenditure on raw material procurement has significantly increased year-on-year.

The aforementioned semiconductor analyst also believes that there will still be strong demand support for domestic computing chips in the next two years, but the recent supply problem of domestic computing cards has suppressed some of the demand for computing infrastructure.

(Author Zheng Chenye)