Editor|信创Zongheng public account (ID: xinchuangzh)
Produced by | Anton Studio
Source: Guolian Securities
1. The demand for intelligent computing power is growing strongly
1.1 AIGC applications greatly increase the demand for computing power
In November 2022, OpenAI released ChatGPT, which triggered many companies to deploy large model training. Subsequently, domestic manufacturers such as Baidu, Huawei, Alibaba, SenseTime, and iFLYTEK actively participated in the large-scale model training and successively released large-scale model products. According to the 2024Q3 financial report released by NVIDIA on November 22, the data center business achieved revenue of $14.514 billion in the third quarter, a year-on-year increase of 278.66%.
1.2 The demand for intelligent computing power extends to multiple fields
According to the "China Computing Power Development Index White Paper" jointly released by the Academy of Information and Communications Technology and IDC in September 2023, the scale of computing power in mainland China continues to grow, and the demand for intelligent computing power is growing strongly. The Internet is still the industry with the largest demand for computing power, accounting for 39% and 53% of general computing power and intelligent computing power respectively. Government, telecommunications, finance, education, and manufacturing account for 8.67%/4.16%/3.67%/4.00%/4.16% of the intelligent computing power, respectively, and the application scenarios of intelligent computing power extend to industrial manufacturing, urban governance, intelligent retail, intelligent scheduling and other fields.
The improvement of intelligent terminal computing power has become a new growth demand
Privatized AI models that can be directly run on smart terminals such as mobile phones and computers have become a new demand in the "large model era", which puts forward higher requirements for the intelligent computing power level of terminals and promotes the iterative upgrading of terminal product computing methods. The penetration rate of smart computing power in mobile phone terminals continues to grow rapidly, and mobile phone manufacturers such as Huawei and Xiaomi have entered the large model one after another, Huawei has directly embedded the large model capabilities at the mobile phone system level, and the HarmonyOS 4 system will be supported by the Pangu large model, and Xiaomi has successfully run the 1.3 billion parameter AI large model locally on mobile phones. On November 13th, vivo officially released the first AI large model mobile phone vivo X100 series, which is the industry's first batch of mobile phones with tens of billions of large models in the terminal, which is equipped with a blue heart model, which is a general AI large model matrix developed by vivo, and users can use the "blue heart small V" intelligent assistant in the new machine.
1.3 The concentration of domestic AI servers is high
According to IDC data, Inspur, H3C, Super Fusion, Ningchang and ZTE ranked among the top 5 in China's server market in 2022, with a total market share of nearly 66.9%. According to the statistics and collation of Guanzhi Domestic Information Network, in 2022, Inspur will occupy the first place in the mainland AI server market (by sales) with a share of 47%, Xinhua III and Ning Chang will occupy the second and third places with 11% and 9% respectively, and the top 3 will account for about 67%.
2. The domestic computing infrastructure industry has ushered in multiple catalysis
2.1 The new U.S. export control regulations will accelerate the localization of computing power
On October 17, 2023, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) issued new export control rules for chips, imposing new export controls on semiconductor products, including NVIDIA's high-performance AI chips, which have come into effect on October 23, 2023. Nvidia's filing with the U.S. SEC shows that the banned products that take immediate effect include A800, H800 and L40S, the most powerful AI chips. The early implementation of the export control will make it impossible for existing AI server manufacturers to deliver due to NVIDIA GPU supply problems, or they will seek domestic GPU alternatives. We believe that this move will accelerate the localization of AI computing power and further improve the construction of the domestic GPU ecosystem.
Policies promote the high-quality development of computing infrastructure
On October 8, 2023, six departments, including the Ministry of Industry and Information Technology, the Cyberspace Administration of China, the Ministry of Education, the National Health Commission, the People's Bank of China, and the State-owned Assets Supervision and Administration Commission of the State Council, jointly issued the "", proposing that by 2025, the scale of computing power in mainland China will exceed 300EFLOPS, and the proportion of intelligent computing power will reach 35%. The Action Plan proposes six key tasks, including improving the comprehensive supply system of computing power, improving the efficient carrying capacity of computing power, strengthening the efficient and flexible guarantee of storage capacity, deepening the application of computing power to empower industries, promoting the development of green and low-carbon computing power, and strengthening the construction of security capacity, and strives to promote the high-quality development of computing infrastructure.
2.2 Nvidia GPU domestic substitution space calculation
NVIDIA's business in China accounts for more than 20%
According to the financial data released by NVIDIA, from NVIDIA's 2019 fiscal year to the present, its revenue in China has remained above 20%, and the data center business in China accounts for about 20%-25% of NVIDIA's related business revenue.
The space for domestic substitution has been significantly improved
Prior to the implementation of the U.S. Department of Commerce's latest export control measures, Nvidia's data center business is expected to have total revenue of about $45 billion in 2023. According to the proportion of data center business revenue in China is about 20%-25%, the revenue of Nvidia's data center business in China in 2023 will be about 90-11.25 billion US dollars, and the neutral estimate is that the demand for data centers in China will increase by 15% year-on-year in 2024, so the replacement space of NVIDIA GPU in China in 2024 will be about 74.2-92.7 billion yuan.
2.3 国产算力谁将受益？
China's AI computing power ecosystem may undergo great changes
Taking into account factors such as the U.S. ban, the current product line layout, R&D strength, and product progress of various domestic companies, we believe that China's AI computing ecosystem may undergo great changes in the future, from the current Nvidia monopoly to the international ecology (NVIDIA, AMD, Intel, etc.) and the domestic ecology competing against each other. NVIDIA is still expected to occupy a certain market share with its good CUDA ecosystem and good cluster performance. However, due to the US ban, the strict restrictions on the performance of a single SIM card may seriously affect its market share in China.
Relying on the strong backing of Sugon and the Chinese Academy of Sciences, Haiguang Information has a complete closed loop of "core-device-cloud", which corresponds to Huawei's "Kunpeng + Ascend" industry chain, and is expected to become an important domestic computing power ecosystem. Other independent third parties such as Cambrian, Bichen Technology, Moore Threads, Jingjiawei, etc. will also fully benefit from the opportunity of computing power localization and are expected to gain a certain market share by virtue of their cost-effective advantages. On the whole, with the continuous iteration of domestic computing power, it is expected to achieve corner overtaking in the future.
The GPGPU technology route may gain a larger market share
Originally, the GPU was designed to improve the performance of the computer in processing data such as images and videos, and solve the performance bottleneck of the CPU. However, with the gradual emergence of GPU's performance advantages in parallel computing and the expansion of the application range of parallel computing, GPU has evolved into two branches: one is the traditional GPU, which continues to be used exclusively for graphics and image processing, and has built-in video codec acceleration engine, 2D acceleration engine, 3D acceleration engine, image rendering and other special computing modules; The matrix operation instruction improves the accuracy and performance of floating-point arithmetic and can meet the needs of more computing scenarios.
With the technological progress and ecological improvement of GPGPU, its application fields continue to expand, in addition to traditional graphics and image processing, it also involves many fields such as business computing, big data processing, and artificial intelligence. In the field of artificial intelligence, its advantages are mainly reflected in:
(1) High training efficiency, accelerating model iteration and optimization;
(2) low energy consumption and operating costs of artificial intelligence;
(3) The software ecosystem is rich and supports mainstream artificial intelligence frameworks, which is convenient for the transplantation of existing applications and the development of new algorithms. At present, GPGPU has become the mainstream choice for AI-related solutions.
Summary:
How much room is there for domestic GPU replacement?
According to the 2024Q3 financial report released by NVIDIA on November 22, the data center business achieved revenue of $14.514 billion in the third quarter, a year-on-year increase of 278.66%. Prior to the implementation of the U.S. Department of Commerce's latest export control measures, Nvidia's data center business is expected to have total revenue of about $45 billion in 2023. According to the proportion of data center business revenue in China is about 20%-25%, the revenue of Nvidia's data center business in China in 2023 will be about 90-11.25 billion US dollars, and the neutral estimate is that the demand for data centers in China will increase by 15% year-on-year in 2024, so the replacement space of NVIDIA GPU in China in 2024 will be about 74.2-92.7 billion yuan.
Who will benefit from domestic computing power?
Relying on the strong backing of Sugon and the Chinese Academy of Sciences, Haiguang Information has a complete closed loop of "core-device-cloud", which corresponds to Huawei's "Kunpeng + Ascend" industry chain, and is expected to become an important domestic computing power ecosystem. Other independent third parties such as Cambrian, Bichen Technology, Moore Threads, Jingjiawei, etc. will also fully benefit from the opportunity of computing power localization and are expected to gain a certain market share by virtue of their cost-effective advantages. On the whole, with the continuous iteration of domestic computing power, it is expected to achieve corner overtaking in the future.