At the beginning of this year, the news that only 2,048 H800GPU were used in DeepSeek-V3 training came out, like a nuclear bomb, bringing unprecedented shocks to the industry. The amazing performance of DeepSeek has undoubtedly injected a "shot in the arm" for the inference chip and inference computing power market, and the resource utilization rate of computing power centers in various regions has shown a good trend of significant improvement.

Recently, the country's first "four-in-one" computing network scheduling platform undertaken by China Mobile was officially put into operation. "Four-in-one" refers to a computing power system that integrates the four computing capabilities of general computing power, intelligent computing power, super computing power and quantum computing power. It is understood that the computing power scheduling platform can support hundreds of millions of computing power calls per day, can schedule 1/6 of the country's computing power scale, and the efficiency of computing network integration is increased by 20%. The localization rate of chips in the platform's own intelligent computing center exceeds 90%, and it is compatible with 8 kinds of domestic AI chips.

At present, AI computing chips, as the core infrastructure supporting various AI applications, are playing an irreplaceable key role. ACCORDING TO IDC'S FORECAST, THE GLOBAL COMPUTING POWER SCALE IS EXPECTED TO GROW FROM 1397 EFLOPS IN 2023 TO 16ZFLOPS IN 2030, WITH A COMPOUND GROWTH RATE OF 50% FROM 2023 TO 2030.

1. Overview of AI computing power chips

With the rapid progress of artificial intelligence technology, especially the rise of A-large model technology, the demand for computing power has shown explosive growth. As the core hardware platform supporting generative AI applications, the computing power behind AI servers mainly comes from AI computing chips.

From the perspective of chip design concept and application field, AI computing power chips can be subdivided into two categories: general-purpose and special-purpose. General-purpose AI chips, such as CPUs (central processing units), GPUs (graphics processing units), and FPGAs (field-programmable gate arrays), are designed to meet a wide range of computing needs. Specialized AI chips, such as TPU (Tensor Processing Unit), NPU (Neural Network Processing Unit), ASIC (Application Specific Integrated Circuit), etc., are tailored for the field of artificial intelligence.

AI computing chips are the cornerstone of computing power. CPU+GPU is the mainstream heterogeneous computing system solution for AI servers, and according to the data of server cost composition in IDC2018, the cost of CPU+GPU in inference and machine learning servers accounts for 50-82.6%, of which the GPU cost of machine learning servers accounts for 72.8%. AI computing chips have powerful parallel computing capabilities, which can quickly process large-scale data and complex neural network models, and realize AI training and inference tasks.

According to different application scenarios, AI computing chips can be further divided into cloud AI computing chips, edge AI computing chips, and terminal AI computing chips. Whether it is the extension from the cloud to the edge or the penetration into the terminal, various application scenarios of artificial intelligence are inseparable from the powerful computing power provided by AI computing chips. It is worth noting that the three scenarios of cloud, edge and terminal put forward differentiated requirements for AI computing chips in terms of computing power and power consumption.

2. Analysis of the industrial chain

The AI computing chip industry chain covers multiple links such as artificial intelligence algorithms, chip design, chip manufacturing, and downstream applications. In the upstream of the industrial chain, it mainly involves artificial intelligence algorithms and chip design tools. Among them, AI algorithms have a wide range of applications, including vision algorithms, speech processing algorithms, natural language processing algorithms, and various machine learning methods (such as deep learning). Chip design and chip manufacturing are the core of the AI computing chip industry. Chip design tool manufacturers, wafer foundries, and packaging and testing manufacturers have provided necessary tool support for the research and development of AI computing chips, and laid a solid foundation for the development of the entire industry.

From the perspective of downstream application scenarios, AI computing chips are widely used in cloud computing and data centers, edge computing, consumer electronics, intelligent manufacturing, intelligent driving, smart finance, intelligent education and other fields.

Third, the development trend

In today's era, computing power has become the core new engine driving the rapid development of the digital economy, and the field of artificial intelligence has also entered a new era driven by computing power, and the scale of global computing power is continuing to expand at an unprecedented rate.

According to the forecast data of IDC, Gartner, TOP500 and the China Academy of Information and Communications Technology, the global computing power scale is expected to grow significantly from 1397EFLOPS in 2023 to 16ZFLOPS in 2030, and the compound growth rate of global computing power is expected to be as high as 50% during this period (2023-2030). From the specific situation of the Chinese market, according to IDC statistics, the scale of China's intelligent computing power has reached 725.3EFLOPS in 2024, and is expected to further climb to 2781.9EFLOPS by 2028, which means that during the period from 2023 to 2028, the compound growth rate of China's intelligent computing power scale is expected to be 46.2%.

Fourth, some listed companies that have deployed AI computing chips

With the United States increasing export controls on high-end GPUs in recent years, domestic AI computing chip manufacturers have ushered in an extremely rare golden period of development opportunities.

At the same time, DeepSeek has achieved extremely cost-effective large-scale model training and inference through technological innovation, and the domestic computing power ecological chain has achieved all-round adaptation with DeepSeek. Through continuous technological innovation, DeepSeek has effectively improved the operation efficiency of AI computing chips, which has greatly accelerated the process of achieving independent and controllable AI computing chips. Domestic AI computing chip manufacturers are expected to continue to increase their market share in the market competition and gradually occupy a more important position in the global AI computing chip market.

After combing, many outstanding companies have emerged in the field of cloud AI computing chips among A-share listed companies.

Cambrian is a leading enterprise of artificial intelligence chips in China, and the cloud-edge-end integration is developing in a coordinated manner. Founded in 2016, the company focuses on the R&D and technological innovation of artificial intelligence chip products, and is committed to building core processor chips in the field of artificial intelligence. Cambrian can provide a series of intelligent chip products and platform-based basic system software with cloud-edge-end integration, software and hardware collaboration, training and inference integration, and a unified ecology.

According to the financial report, according to the company's 2024 performance report, the company will achieve operating income of 1.174 billion yuan in 2024, a year-on-year increase of 65.56%, and the net profit attributable to the parent company will be -443 million yuan, a year-on-year loss of 47.76%.

(Cambrian's operating income from 2020 to 2024 Source: Flush Finance).

Haiguang Information(301325. Founded in 2014, Haiguang Information Co., Ltd.'s main business is the R&D, design and sales of high-end processors, which are mainly used in computing and storage devices such as servers and workstations.

According to the financial report, Haiguang Information will achieve revenue of 9.162 billion yuan in 2024, a year-on-year increase of 52.40%; The net profit attributable to the parent company was 1.931 billion yuan, a year-on-year increase of 52.87%. Due to the continuous iterative upgrading of the company's products, the increase in the gross profit margin of new products, and the continuous improvement of the company's profitability, the company will achieve a gross profit margin of 63.72% in 2024, an increase of 4.05% year-on-year, and the company will continue to increase R&D investment, with R&D investment of 3.446 billion yuan in 2024, a year-on-year increase of 22.63%, and R&D investment accounting for 37.61% of operating income.

(Haiguang Information's net profit from 2020 to 2024 Source: Flush Finance).

VeriSilicon (688521. SZ) is an enterprise that relies on independent semiconductor IP to provide customers with platform-based, all-round, one-stop chip customization services and semiconductor IP licensing services. According to IPnest data, VeriSilicon's IP licensing market share in 2023 ranked first in China and eighth in the world. In 2023, VeriSilicon's IP licensing fee revenue ranked sixth in the world.

According to the 2024 performance report, the company will achieve operating income of 2.323 billion yuan in 2024, a year-on-year decrease of 0.66%; The net profit attributable to the parent company was -605 million yuan, an increase year-on-year, and the company increased the proportion of R&D investment, and the R&D expenses in 2024 increased by about 32% year-on-year.

(VeriSilicon's operating income from 2020 to 2024 Source: Flush Finance).

SMIC (688691. SZ) was established in 2000, mainly to provide customers with integrated circuit wafer foundry and supporting services based on a variety of technology nodes and different process platforms. The company is one of the world's leading integrated circuit wafer foundry enterprises, and also a leader in the integrated circuit manufacturing industry in Chinese mainland, with leading process manufacturing capabilities, production capacity advantages and service support; According to the latest 2024 sales announced by the world's pure-play foundry companies, SMIC ranks second in the world and first among Chinese mainland companies.

According to the financial report, in 2024, the company will achieve operating income of 57.796 billion yuan, a year-on-year increase of 27.7%; The net profit attributable to the parent company was 3.699 billion yuan, a year-on-year decrease of 23.31%. The company continued to increase R&D investment, and in 2024, the company's R&D investment reached 5.45 billion yuan, a year-on-year increase of 9.1%, accounting for 9.4% of operating income.

(SMIC's operating revenue from 2020 to 2024 Source: Flush Finance).

(The content of this article is from public information, for reference only, does not constitute investment advice, and you operate at your own risk).