CITIC SEC: ASIC chip industry chain is about to enter a period of performance realization, top manufacturers are expected to continue to lead.

CITIC SEC released a research report stating that companies in the ASIC chip industry chain are expected to enter a period of performance realization by 2025, with the potential for continued high-speed growth in the future. The proportion of shipments for the two sub-products in the future is expected to remain at around 8:2 in the medium to long term, with ASIC having a relatively limited impact on GPUs. From the perspective of product supply, the ASIC industry is expected to maintain an oligopoly competitive landscape in the future, with U.S. suppliers leading the way. It is recommended to pay attention to investment opportunities in ASIC and leading companies in the industry chain. The main points of CITIC SEC are as follows: ASIC is expected to be an important direction for the AI industry chain by 2025. 1) In 2023-2024, as the generative AI industry develops, various technology companies have increased their layout for large model training, leading GPU vendors such as NVIDIA to enter a phase of rapid growth in performance. However, with the development of AI Agent technology, the market's demand for inference is also increasing. The demand for AI ASICs in the inference stage of generative AI is expected to significantly increase. 2) In December 2024, Broadcom's CEO raised the AI revenue expectation for XPU + network business by 2027 at the FY24Q4 earnings conference, further increasing market attention to the ASIC field. 3) Compared to commercial GPUs, ASIC chips have significant advantages in cost, power consumption, etc., and cloud providers have clear demand and motivation. The report primarily analyzes the industrial development trend of the ASIC industry, industry competitive landscape, and the progress of various companies. CSP companies' self-developed AI accelerator cards are expected to drive long-term growth in the ASIC industry market size. 1) Currently, high-end AI ASIC products worldwide have iterated to 3-5nm, with supply mainly from Broadcom and Mellanox. Currently, ASIC product suppliers have actively cooperated with overseas tech giants (Amazon, Google, Microsoft, Meta, Tesla), and related companies are expected to enter a period of performance realization by 2025. 2) The precondition for the mass production of the ASIC industry lies in CSP companies' self-development. The main reason is that CSP companies hope to continue reducing costs and break free from dependence on a single supplier through self-developed ASIC products. 3) Market size: Mellanox estimated at the AI DAY investor conference in April 2024 that the global market size for acceleration customers in data centers is expected to grow from $6.6 billion in 2023 to $27.5-42.9 billion in 2028, with a corresponding CAGR of 63%-110%. 4) The expected impact on GPUs is relatively limited. The market is currently concerned about the competition between GPUs and ASICs. By splitting NVIDIA's customer structure and predicting future development trends, the proportion of shipments for the two sub-products is expected to remain at around 8:2 in the medium to long term. The expected impact of ASICs on GPUs is relatively limited. The entry barrier is relatively high, and future industry product supply will be dominated by a few manufacturers. 1) The entry barrier for ASIC products is relatively high. (a) From the algorithmic perspective, although the core architecture of AI algorithms is stable (transformer), the optimization methods for engineering are different, and on the post-training side, the model (parameter weights, etc.) needs to be continuously updated based on new data, and the application scenarios on the inference side are also different (and the computing power consumption is much higher than traditional small models). Therefore, the report judges that the threshold for flexibility and computing power of AI chips on the inference side is much higher than people's imagination. (b) From the perspective of product iteration speed, cloud providers hope that ASICs can be used not only for inference but also for training (Amazon has merged chip models), so chip design pursues training & inference integration. Also, from the iteration speed of the competitor NVIDIA, with a product pace of one to two generations every two years, it means that everyone needs to follow a similar pace to avoid losing business value. 2) Referring to NVIDIA's development history, future ASIC products will not simply compete in computing power or cost but will need to focus more on system advantages. Chip manufacturers need to accumulate expertise in chip design, networks & interfaces, etc., and have long-term cooperation with cloud providers at the system level. Based on this, the report predicts that future ASIC industry product supply will be mainly provided by U.S. manufacturers, with Taiwanese manufacturers supplementing. Investment strategy: Looking forward to 2025, the AI industry chain is expected to continue to maintain a high level of prosperity, and industry opportunities will further expand around NVIDIA. With North American CSP companies developing ASIC chips on their own, companies related to the industry chain are expected to enter a period of performance realization by 2025 and are likely to maintain high-speed growth in the future. From the perspective of product iteration speed and product performance, the report judges that the future ASIC industry will be mainly dominated by a few manufacturers, with U.S. suppliers leading the way. It is recommended to pay attention to companies with a significant presence in the ASIC field. Additionally, as ASIC products are mass-produced, there is strong demand for advanced processes. It is advised to pay attention to: 1) leading companies in advanced process fields; 2) leading companies in the PCIe Retimer field; 3) companies in the network switch field. Risk factors: The risk of a significant decrease in downstream IT spending; the risk of slower-than-expected development in artificial intelligence; the risks of potential ethical, moral, and user privacy issues in AI; the risk of intensifying industry competition; the risk of product R&D falling short of expectations, etc.