Overview of Domestic AI Computing Power Ecosystem

Core Viewpoints

ChatGPT’s large model computational demand is rapidly expanding.1) The large model represented by ChatGPT has seen a significant increase in parameter and data volume, with the GPT-3 model reaching 175 billion parameters, requiring strong computational support for training.2) Currently, Google has a daily search volume of 3.5 billion, and we believe there is significant room for growth in daily active users for ChatGPT, with computational demands likely to continue to be released.3) In the future, under the trend of multimodal applications, a broader range of data forms, more application scenarios, and deeper user experiences will greatly increase the computational demands supporting artificial intelligence, ushering in an era of rapid expansion in computational power.

AI chips create the foundation for computational power, leading companies accelerate their pursuit.Leading companies such as Huawei, Haiguang Information, Cambricon, and Baidu are accelerating their layout in the AI chip market, continuously improving product performance, with some nearing international advanced levels:

Huawei’s Ascend 910 integer precision performance reaches 640 TOPS, and half precision reaches 320 TFLOPS, comparable to leading international products; its Atlas 300T training card is mainly used in scenarios requiring AI training and high-performance computing, such as telecommunications, internet, and finance;
Haiguang’s Deep Computing No. 1 DCU has 60-64 computing units, with up to 4096 cores, featuring strong parallel computing capabilities and high energy efficiency, and has achieved large-scale sales;
Cambricon’s SiYuan 370 chip uses 7nm process technology and chiplet technology, integrating 39 billion transistors, with a maximum computational power of 256TOPS(INT8);
Baidu’s Kunlun 2nd generation AI chip has a general computing core performance improvement of 2-3 times, with half precision reaching 128 TFLOPS, supporting training and inference;
Jingjiawei’s GPU can be widely used in PCs, servers, graphic workstations, etc., meeting the display computing needs of geographic information systems, image matching, signal processing, and airborne, vehicle-mounted, and shipborne display control.

Soft and hard ecosystems build barriers, focus on the leading ecosystems of Haiguang and Ascend.Considering that besides the technical barriers of the GPU itself, the software ecosystem also becomes an important barrier to enhance the competitiveness of GPU manufacturers, we recommend focusing on Haiguang and Kunpeng and their partners, which have outstanding product performance, complete ecosystems, and rich downstream applications:

Haiguang ecosystem: The Haiguang DCU co-processor can adapt well to NVIDIA‘s CUDA ecosystem, reducing development and migration difficulties, and also alleviating promotion pressure; it builds a relatively complete AI toolchain ecosystem, maximizing the use of existing mature AI algorithms and frameworks; CPUs and GPGPUs also receive support from mainstream manufacturers across the industry chain, and we recommend paying attention to Haiguang Information, Zhongke Shuguang, etc.
Ascend ecosystem: The Ascend computing industry ecosystem is built on the Ascend series processors and basic software, constructing a full-stack AI computing infrastructure, industry applications, and services. In terms of the software and hardware system, the Atlas hardware, MindSpore framework, and AI development platform form a complete cooperation system; in terms of complete machines, Digital China and Tuowei Information are among Huawei’s first partners in the Ascend computing field; in terms of industry applications, Beiming Software joined the Ascend Wanli Partner Program in 2022, clearly indicating comprehensive cooperation intentions in finance, internet, electricity, etc., and the Ascend computing industry ecosystem is gradually improving. We recommend focusing on Digital China, Tuowei Information, Changshan Beiming, etc.

Risk Warning:Risks of AI technology iteration not meeting expectations, risks of economic downturn exceeding expectations, risks of intensified industry competition.

Report Body

ChatGPT‘s large model computational demand is rapidly expanding.

ChatGPT, based on a large model, performs well in translation, Q&A, content generation, and other fields.1) ChatGPT is a form of generative AI, and Gartner ranks it as the top strategic technology trend for 2022. 2) According to research by Tencent Research Institute, current AI is mostly trained for specific scenarios, making it difficult for generated models to transfer to other applications, falling into the category of ‘small models’. The entire process requires a lot of manual tuning and feeding the machine with massive labeled data, which lowers the research and development efficiency of AI and incurs high costs. The support behind ChatGPT is the AI large model. Large models are usually trained on unlabeled large datasets using self-supervised learning methods. Later, in applications in other scenarios, developers only need to fine-tune the model or use a small amount of data for secondary training to meet the needs of new application scenarios. This means that improvements to large models can benefit all downstream small models, significantly enhancing the applicability and R&D efficiency of AI. 3) Therefore, large models have become a key investment direction in the industry, with OpenAI, Google, Facebook, Microsoft, and domestic companies like Baidu, Alibaba, Tencent, Huawei, and Zhiyuan Research Institute all launching super-large models. In particular, the impressive performance of OpenAI’s GPT-3 model in translation, Q&A, content generation, and other fields has given the industry hope for achieving general artificial intelligence. The current version of ChatGPT is GPT-3.5, which is an optimization on top of GPT-3, further enhancing its capabilities.

Overview of Domestic AI Computing Power Ecosystem

Based on large models, the parameter and data volumes are rapidly expanding, leading to a sharp increase in computational demand.Under the framework of large models, the parameter volume of each generation of GPT models is rapidly expanding; at the same time, the demand for pre-training data is also increasing rapidly. We believe that the rapid penetration and application of ChatGPT will also significantly boost computational demand.

Overview of Domestic AI Computing Power Ecosystem

ChatGPT’s monthly active users exceed one billion, making computational power a key metric for investment..According to data from Similarweb, in January 2023, ChatGPT had over 100 million cumulative users, setting a record for the fastest app to reach 100 million users on the internet, surpassing TikTok’s previous record of 9 months.

1) Access Phase: Initial investment of nearly one billion dollars, with daily electricity costs of tens of thousands of dollars.

NVIDIA A100: According to OneFlow, currently, NVIDIA A100 is the most cost-effective GPU choice on AWS.
NVIDIA DGX A100 server: Each machine is equipped with 8 A100 GPUs, with AI computational performance of approximately 5 PetaFLOP/s, and a maximum power of about 6.5kw, priced at about 199,000 USD/machine.
Daily Consultation Volume: According to data from Similarweb, as of the end of January 2023, the chat.openai.com website (i.e., the official ChatGPT website) attracted a daily visitor count of up to 25 million during the week of 2023/1/27-2023/2/3. Assuming a stable state where each user asks about 10 questions daily, the daily consultation volume would be approximately 250 million.
A100 Running Hours: Assuming each question averages 30 words, each word on the A100 GPU consumes about 350ms, thus a total of 729,167 A100 GPU running hours are needed per day.
A100 Demand: Corresponding to the need for 30,382 NVIDIA A100 GPUs to be computed simultaneously each day to meet the current access volume of ChatGPT.
Initial Computational Investment: Based on the aforementioned NVIDIA DGX A100, it requires 30,382/8=3,798 servers, corresponding to 3,798/7=542 cabinets. To meet the current consultation volume of millions of users for ChatGPT, the initial computational investment cost is approximately 542*140=759 million USD.
Monthly Electricity Costs: In terms of electricity consumption, 542*45.5kw*24h=591,864kwh/day. Referring to Hashrate Index statistics, we assume that the average industrial electricity price in the US is about 0.08 USD/kwh. Thus, the daily electricity cost is approximately 591,864*0.08=47,349 USD/day.

Additionally, considering that Google‘s daily search volume has reached 3.5 billion, we believe there is significant room for growth in daily active users for ChatGPT, with computational demands likely to continue to be released.

2) Training Phase: Under public cloud, a single training costs between one million to ten million USD.

Each token training cost is usually about 6N (while inference cost is about 2N), where N is the number of parameters of the LLM;
Assuming that during the training process, the model’s FLOPS utilization is 46.2%, consistent with the training of the PaLM model (which has 540 billion parameters) on the TPU v4 chip.
According to OneFlow, the cost of training GPT-3 once is approximately 1.398 million USD; for some larger LLM models (such as Gopher with 280 billion parameters and PaLM with 540 billion parameters), using the same calculation formula, the training cost is between 200 thousand to 1.2 million USD.

Overview of Domestic AI Computing Power Ecosystem

The current text interaction is just the tip of the iceberg for ChatGPT and AIGC application scenarios, with voice, image, video, and other forms of input and output potentially bringing revolutionary changes to content creation. Moreover, a broader range of data forms, more application scenarios, and deeper user experiences will significantly increase the computational demands supporting artificial intelligence, ushering in an era of rapid expansion in computational power, with manufacturers of servers, chips, IDC, and optical communication likely to benefit significantly.

Overview of Domestic AI Computing Power Ecosystem

AI chips create the foundation for computational power, leading companies accelerate their pursuit.

AI chips optimize machine learning and deep learning computations, representing a technological shift compared to traditional CPUs. Based on CPUs, AI chips optimize computations commonly used in machine learning and deep learning, exhibiting technological changes in parallel computing, low-precision computing, and memory optimization, performing different functions from CPUs to meet the computing needs of the new era.

Overview of Domestic AI Computing Power Ecosystem

According to technical architecture classification, AI chips include Graphics Processing Units (GPU), Field Programmable Gate Arrays (FPGA), and Application-Specific Integrated Circuits (ASICs).1) GPUs were originally designed for processing parallel computations of images. Since 2012, GPUs have increasingly been used for training AI systems; by 2017, GPUs had become the dominant AI chip. According to the Haiguang Information prospectus, GPGPU still remains the mainstream architectural choice, accounting for 90% of the market.2) However, GPUs still adopt a general computing design, while FPGAs and ASICs have become more prominent in training and inference.ASICs include hard-etched circuits customized for specific algorithms, as ASICs are optimized for specific algorithms, they usually have higher performance and speed than FPGAs; FPGAs can be reconfigured by programmers after manufacturing to adapt to specific algorithms, offering higher versatility than ASICs for secondary programming and application modification.

Overview of Domestic AI Computing Power Ecosystem

According to the tasks they undertake, AI chips include training chips and inference chips..Training chips learn through large amounts of labeled or unlabeled big data to build neural network models, requiring stronger computational power and often leading to higher power consumption; the latter infers conclusions based on the trained model. According to SCET calculations, training chips and inference chips have efficiency and speed improvements of 10~1000 times compared to equivalent power-consuming CPUs.

Overview of Domestic AI Computing Power Ecosystem

Huawei Ascend, Haiguang Information, Cambricon, Baidu, and other leading companies are accelerating their layout in the AI chip market, continuously improving product performance, with some nearing international advanced levels:

Huawei Ascend (Training+Inference):1) Inference Card: The Ascend 310 chip is Huawei’s first full-stack, all-scenario AI chip, with a power consumption of only 8W, capable of outputting integer precision (INT8) performance of 16 TOPS and half precision (FP16) performance of 8 TOPS; its Atlas 300 inference card is widely used in scenarios such as smart cities, smart transportation, and smart finance.2) Training Card: The Ascend 910 has a power consumption of 310W, with integer precision (INT8) performance reaching 640 TOPS, and half precision (FP16) performance reaching 320 TFLOPS, comparable to leading international products; its Atlas 300T training card is mainly used in fields requiring AI training and high-performance computing, such as telecommunications, internet, and finance.

Overview of Domestic AI Computing Power Ecosystem

Haiguang Information (Training): The company’s main products include general processors (CPU) and Haiguang co-processors (DCU). The Haiguang DCU corresponds to the Haiguang 8000 series, which is an AI training chip designed and developed by Haiguang. The company started the design of the “Deep Computing No. 1” product in October 2018 and has achieved large-scale sales. This chip is equipped with 60-64 computing units, with a maximum of 4096 cores, featuring strong parallel computing capabilities and high energy efficiency, suitable for compute-intensive applications such as vector and matrix calculations. The Haiguang DCU is compatible with “CUDA-like” (ROCm) environments, with a rich software and hardware ecosystem, and can be widely used in big data processing, artificial intelligence, commercial computing, and other compute-intensive application fields. In January 2020, the company initiated the research and development of the second-generation DCU “Deep Computing No. 2”.

Overview of Domestic AI Computing Power Ecosystem

Cambricon (Training+Inference): 1) Integrated Training and Inference: The SiYuan 370 chip is a training and inference AI chip launched by Cambricon, using 7nm process technology and chiplet technology, integrating 39 billion transistors, with a maximum computational power of 256TOPS(INT8), which is 2 times that of the previous generation SiYuan 270 in terms of computational power and 3 times in terms of memory bandwidth.2) Inference Card: Cambricon’s SiYuan 270 is an inference chip capable of processing non-sparse AI models, with peak performance reaching 128TOPS(INT8). SiYuan 270 also supports various precision operations including INT4 and INT16, as well as floating-point and mixed-precision operations. It is suitable for various AI applications, including vision, speech, natural language processing, and machine learning. Additionally, the SiYuan 290 chip is Cambricon’s first AI training chip, integrating 460 billion transistors, with HBM2 memory providing the high memory bandwidth required for AI training, and vMLU technology helping customers achieve cloud virtualization and resource isolation.

Overview of Domestic AI Computing Power Ecosystem

Baidu Kunlun Chip (Training+ Inference):1) Inference Card: The first and second generation Kunlun AI chips are named K series and R series respectively. Among them, the first generation Kunlun AI chip is a cloud inference chip supporting general AI algorithms. The chip has powerful performance, with integer precision (INT8) reaching 256 TOPS and half precision (FP16) reaching 64 TFLOPS, deployed in thousands of units across Baidu’s search engine, Xiaodu, and other businesses, empowering industries such as internet, industrial manufacturing, smart finance, and smart transportation.2) Integrated Training and Inference: Compared to the first generation product, the Kunlun 2nd generation AI chip has improved general computing core performance by 2-3 times, with half precision (FP16) reaching 128 TFLOPS, supporting both training and inference, providing strong AI computational power for high-performance computing in data centers, supporting virtualization, inter-chip connectivity, and video encoding and decoding.

Overview of Domestic AI Computing Power Ecosystem

Jingjiawei (Inference): Jingjiawei is a leading enterprise in the domestic high-performance GPU field. The company started developing the first domestically reliable, low-power GPU chip JM5400 in 2014, succeeded in developing the second generation high-reliability, high-performance GPU JM7200 in 2018, which has been widely applied in the market, and completed the third generation product JH920 upgrade by the end of 2021. JH920 is Jingjiawei’s third generation high-performance GPU, with significantly improved performance compared to the previous two generations, mainly used in mid-to-high-end graphic display, general computing, and embedded fields. JH920 fully supports domestic CPUs, domestic operating systems, and domestic firmware, and can be widely applied in PCs, servers, graphic workstations, etc., meeting the display computing needs of geographic information systems, image matching, signal processing, and airborne, vehicle-mounted, and shipborne display control..

AI chips create the foundation for computational power, leading companies accelerate their pursuit.

3.1Software strengthens GPU competition barriers, improving the ecosystem becomes key to development.

NVIDIA’s CUDA ecosystem strengthens chip high barriers. CUDA is a general parallel computing architecture launched by NVIDIA in 2006, which includes the instruction set (ISA) used for NVIDIA GPUs and the internal parallel computing engine of the GPU. CUDA provides an easy interface for GPU programming, allowing programmers to compile applications based on GPUs and utilize the parallel computing capabilities of GPUs to solve complex computational problems more efficiently. According to data from Jon Peddie Research, as of the fourth quarter of 2022, NVIDIA maintains a leading position as the world’s independent GPU supplier with a market share of 82%, while Intel and AMD each account for about 9%.

Overview of Domestic AI Computing Power Ecosystem

Software ecosystem becomes an important barrier for GPU manufacturers.NVIDIA’s GPU is still the mainstream solution for cloud-based artificial intelligence acceleration globally; fundamentally, other AI chip companies find it difficult to compete with NVIDIA‘s CUDA ecosystem: On the one hand, it depends on the complete programming and AI toolchain, which requires long-term accumulation; on the other hand, it relies on its wide range of applications and partners.

1) A complete AI toolchain ecosystem facilitates chip promotion. In the early stages of promoting any new computing platform, developers need to port existing applications to the new platform, thus requiring advanced toolchains and development environments; as applications expand, data center-level support will also require more tools. Taking CUDA as an example, NVIDIA relies on the CUDA platform, having formed a complete toolchain through long-term accumulation while collaborating with third parties to provide developers with a comprehensive ecosystem component, which is deeply tied to hardware and more conducive to the expansion of NVIDIA chips.

2) Applications and partners influence chip deployment.The software ecosystem built on chips profoundly affects chip usability, and the upper-layer application programs and partners directly determine whether chips can genuinely land and are worth investing in.NVIDIA continuously launches initiatives into new computing fields, expanding from cloud computing and healthcare to autonomous driving, robotics, language models, and even NASA’s Mars landing program.

ROCm benchmarks CUDA, providing source-level support for CUDA programs. In 2015, AMD developed ROCm, an open-source software development platform for HPC and large-scale GPU computing, aiming to establish an alternative to the CUDA ecosystem, providing source-level support for CUDA programs. Although AMD’s ecosystem is based on the open-source ecosystem OpenCL, AMD has also created a programming model called HIP, which almost completely replicates the CUDA API; in 2016, AMD demonstrated the automatic migration of the deep learning framework CAFFE from CUDA to HIP, achieving an automatic migration rate of 99.6%.

Overview of Domestic AI Computing Power Ecosystem

Learning from AMD‘s development strategy, in the short term, domestic GPUs that are compatible with CUDA can facilitate promotion, while developing proprietary core technology should be the long-term strategy.1) In the short term, domestic GPUs that are compatible with CUDA and other international ecosystems can leverage NVIDIA’s well-established software ecosystem, reducing development and migration difficulties, thereby lowering promotion pressure.2) In the long term, as the CUDA architecture undergoes minor adjustments, if domestic GPUs are developed entirely based on the CUDA ecosystem, hardware updates will be tied to NVIDIA’s development process. Therefore, improving their own toolchain and downstream applications, building company ecosystem barriers, and developing proprietary core technology should be the long-term strategy.

3.2Haiguang Ecosystem: Compatible with international mainstream computing ecosystems, with rich downstream applications.

Haiguang DCU products are compatible with international mainstream ecosystems, facilitating rapid migration. During the cross-platform migration process, operator loss and precision degradation can lead to low migration success rates. The Haiguang DCU co-processor is fully compatible with ROCm GPU computing ecology; since ROCm and CUDA are highly similar in terms of ecosystem and programming environment, CUDA users can quickly migrate to the ROCm platform at a low cost, and ROCm is also known as “CUDA-like”. Therefore, the Haiguang DCU co-processor can adapt well to NVIDIA’s commercial computing software and AI software, with a rich software and hardware ecosystem, and can be widely used in big data processing, artificial intelligence, commercial computing, and other compute-intensive application fields, mainly deployed in server clusters or data centers, providing high-performance, high energy efficiency computational power for high-complexity and high-throughput data processing tasks.

Overview of Domestic AI Computing Power Ecosystem

Improving the AI toolchain ecosystem to maximize the use of existing mature AI algorithms and frameworks.1) Provide a unified underlying hardware driver platform that supports common computing frameworks, libraries, and programming models;2) Provide a hierarchical software stack that adapts to different API interfaces and compilers, maximizing the use of existing mature AI algorithms and frameworks.

Overview of Domestic AI Computing Power Ecosystem

Forming a synergistic effect with CPUs, the CPU+GPGPU heterogeneous computing architecture increases flexibility.1) There are multiple technical routes for computational co-processors, including GPGPU, ASIC, and FPGA. Among them, representative enterprises of GPGPU include NVIDIA and AMD; utilizing ASIC technology, many large companies have developed co-processor products, including Intel’s Phi and NNP, Google’s TPU, Huawei Ascend, Cambricon, etc.; many dedicated co-processor products have emerged based on Intel and Xilinx’s FPGAs.2) Considering performance, energy efficiency, and programming flexibility, GPGPU has a significant advantage in co-processor applications, occupying over 90% of the AI market share, with extensive market space in smart factories, autonomous driving, and smart cities.3) Haiguang adopts the GPGPU route, and the CPU+GPGPU heterogeneous computing architecture allows the system to have greater flexibility, meeting the different needs of complex scenarios, significantly improving the task execution efficiency of using either CPU or GPGPU alone; the CPU and GPGPU can also interact through shared memory, leveraging the advantages of heterogeneous computing.

CPUs and GPGPUs receive support from mainstream manufacturers across the industry chain, with an increasing number of partners. Currently, the company has a complete industry chain ecosystem, supporting mainstream manufacturers’ products and services in operating systems, cloud computing, databases, big data, artificial intelligence, and commercial computing software.

Zhongke Shuguang: As of the third quarter of 2022, Zhongke Shuguang holds 27.96% of Haiguang Information’s shares. Zhongke Shuguang is a leading enterprise in domestic server solutions, with mature server solutions that help Haiguang expand its industry market.

Other OEM customer support: Haiguang’s products have received support from many OEM customers such as New H3C, Lenovo, etc., forming a comprehensive and complete machine instance, promoting subsequent customer purchases of the company’s products.
Support for mainstream BIOS: Currently, the company’s products support mainstream BIOS manufacturers such as BaiAo, Kunlun, Insyde, etc.

Overview of Domestic AI Computing Power Ecosystem

Accelerating the autonomous ecosystem centered around Haiguang, establishing the “Light and Harmony Organization” for industry chain ecosystem construction:

In April 2020, the company established the “Haiguang Industry Ecosystem Cooperation Organization”, abbreviated as “Light and Harmony Organization”, aiming to unite upstream and downstream enterprises, universities, research institutions, and industry enterprises around the domestic autonomous general computing platform to achieve collaborative technological breakthroughs, jointly create secure, user-friendly, and open products and solutions, and carry out a series of activities such as testing and certification, technical training, program incubation, application demonstration, and promotion and communication to promote the common development of cooperative organization members and build an inclusive and prosperous information technology ecosystem.
The achievements of the Light and Harmony Organization have been remarkable. Currently, the organization has over 1,000 members, over 500 certified manufacturers, over 1,000 product certifications, and has established 10 regional branches and 15 adaptation centers.

Overview of Domestic AI Computing Power Ecosystem

Rooted in the domestic market, massive demand will continue to accumulateKnow-how, with plans to expand into more downstream fields in the future. Currently, the Haiguang DCU mainly targets compute-intensive application fields such as big data processing, commercial computing, and artificial intelligence, as well as general artificial intelligence applications. Compared to leading international chip companies, the company is rooted in the Chinese domestic market and understands Chinese customer needs better, enabling it to provide safer and more controllable products and more comprehensive and detailed solutions and after-sales services, possessing localized competitive advantages. As the demonstration effect of the company’s products in the aforementioned fields gradually becomes apparent and the company’s marketing efforts continue to strengthen, its products will expand into more fields, capturing a larger market share.

3.2Ascend Ecosystem: Building full-stack AI computing, with in-depth ecosystem partners.

The Ascend computing industry ecosystem is built on the Ascend series processors and basic software, constructing a full-stack AI computing infrastructure, industry applications, and services, which can be divided into three layers: Ascend computing software and hardware system, partners, and industry applications.

1) Hardware System: Atlas series hardware products, such as embedded modules, boards, small stations, servers, clusters, etc. Atlas partners include Digital China, Xiangjiang Kunpeng (Tuowei Information), Anqing, Baode, Huakun Zhenyu (Changhong), Yangtze Computing, Yellow River Technology, New H3C, Baixin, Tsinghua Tongfang, and Guangdian Wuzhou, etc.

Digital China: In 2021, Digital China became one of Huawei’s first partners in Ascend computing, and according to the company’s official WeChat account, the KunTai A722 inference server based on “Kunpeng + Ascend” at its core can provide computing power for 128 processing cores within a compact 2U space, while supporting up to 8 Huawei Atlas 300 inference cards, providing 256GB inference cache and a maximum of 704 TOPS INT8 AI computational power.
Tuowei Information: In 2021, the company became one of the first partners in Ascend, and in April 2022, the Zhaohan inference server RA2300-A series was developed based on Ascend processors, completing compatibility tests with Huawei’s Atlas 300I Pro inference card and Atlas 300V Pro video parsing card, capable of carrying up to 8 Atlas 300V Pro video parsing cards or Atlas 300I Pro inference cards.

2) Basic Software:

The heterogeneous computing architecture CANN and corresponding drivers, runtime, acceleration libraries, compilers, debugging and tuning tools, development toolchains such as MindStudio, and various operation and maintenance management tools are open to a wide range of developers and customers;
AI computing frameworks, including the open-source MindSpore, as well as various popular frameworks in the industry, as an organic part of the ecosystem: MindSpore’s partners include Pengcheng Laboratory, Shenzhen Bay Laboratory, Peking University, Tsinghua University, Harbin Institute of Technology, Douyu, etc.
AI development platforms such as ModelArts, HiAI Service, etc., with partners including Fourth Paradigm, Yitong Technology, Zhongke Hongyun, etc.

3) Industry Application Partners: Numerous partners, together with Huawei, have launched various AI solutions, widely applied in telecommunications, finance, internet, energy, transportation, education, healthcare, and other industries, creating significant industry value in practice.

Changshan Beiming: According to the official account of Beiming Software, a wholly-owned subsidiary, in 2021, Beiming Software officially signed a contract with Nanjing Jiangbei New District to assist Huawei and Jiangbei New District in building the Nanjing Ascend AI Computing Center; in 2022, Beiming Software officially joined the Ascend Wanli Partner Program, becoming an application software partner of Ascend, clearly indicating comprehensive cooperation intentions in finance, internet, electricity, etc. With Huawei’s leadership and the collaboration of Huawei’s ecosystem partners, the Ascend industry ecosystem is gradually improving.

Investment Targets

With the arrival of the large model era represented by ChatGPT, multi-modal AI technologies such as voice, image, and video are rapidly rising, leading to a sharp increase in computational demand due to broader data forms, more application scenarios, and deeper user experiences. AI chips, as the core of computational power, are currently dominated by overseas manufacturers, while domestic leaders such as Huawei Ascend, Haiguang Information, Cambricon, and Baidu are accelerating their layout. Considering that, in addition to the technical barriers of GPUs themselves, the software ecosystem has also become an important barrier to enhance the competitiveness of GPU manufacturers, we recommend focusing on Haiguang and Kunpeng and their partners, which have outstanding product performance, complete ecosystems, and rich downstream applications: 1) In the Haiguang ecosystem, focus on Haiguang Information, Zhongke Shuguang, etc.; 2) In the Ascend ecosystem, focus on Digital China, Tuowei Information, Changshan Beiming, etc.

Risk Warning

Risks of AI technology iteration not meeting expectations: If AI technology iteration does not meet expectations, and the NLP technology’s understanding of human intent does not achieve breakthroughs, it will have some adverse effects on companies related to the industry chain.

Risks of economic downturn exceeding expectations: If the macroeconomic situation declines, with a slowdown in fixed asset investment, it will affect companies’ willingness to reinvest, thereby impacting consumers’ willingness to spend and the production willingness of the industry chain, which will adversely affect the entire industry, and the application of NLP technology will be limited.

Risks of intensified industry competition: If related companies accelerate their technology iterations and application layouts, the overall industry competition will intensify, posing threats to the growth of companies within the industry.

For detailed analysis, please refer to the report published on March 5, 2023, titled “Overview of Domestic AI Computing Power Ecosystem”.

Analyst Liu Gaochang Analyst ID S0680518090001

Research Assistant Sun Xingzhen Analyst ID S0680122020018

Special Statement: The “Measures for the Management of Appropriateness of Securities and Futures Investors” officially came into effect on July 1, 2017. This material produced in WeChat form is only aimed at professional investors among clients of Guosheng Securities. Please do not forward this material in any form. If you are not a professional investor among Guosheng Securities clients, to ensure service quality and control investment risks, please cancel your follow, and do not subscribe to, accept, or use any information contained in this material. As it is difficult to set access permissions for this subscription number, we apologize for any inconvenience caused! Thank you for your understanding and cooperation.

Important Statement: This subscription number is established by the computer team of Guosheng Securities. This subscription number is not a platform for publishing research reports from the Guosheng computer team. The information contained in this subscription number is only aimed at professional investment institutions and is only for timely communication of research views in the context of new media. The information contained in this subscription number is all excerpted from research reports that have been published by the Guosheng Securities Research Institute or are subsequent interpretations of published reports. In case of ambiguity arising from the excerpt from the report, the complete content of the report on the day of publication shall prevail. This material only represents the judgment on the day of the report’s release, and relevant analytical opinions and speculations may change without notice. Readers should also keep track of the latest research progress in a timely manner.

This material does not constitute a judgment or investment advice regarding specific securities at specific price points, specific times, or specific market performances. It cannot be equated with operational opinions guiding specific investments. Ordinary individual investors using this material may misunderstand key assumptions, ratings, target prices, etc., in the report due to a lack of interpretation services, thereby causing investment losses. Therefore, individual investors should seek guidance from professional investment advisors. This material is for reference only, and recipients should not rely solely on the information in this material to replace their independent judgment, and should make their own investment decisions and bear investment risks.

Related posts

Leave a Comment Cancel reply