东海证券晨会纪要

Group 1 - The report highlights the rapid development of AI large models, which is driving a significant increase in demand for AI servers, projected to grow at a CAGR of approximately 29% over the next three years [20][10] - The global AI server market is expected to expand from approximately 855,000 units in 2023 to around 2,369,000 units by 2026, indicating a strong growth trajectory for the AI computing chip and optical module industries [20][10] - The report identifies that GPUs are the most common type of AI chips, accounting for about 89% of the AI computing market in China, and they represent approximately 70-75% of the value in AI servers [8][20] Group 2 - The optical module market is anticipated to benefit from the high growth in computing demand, with a significant increase in demand for high-speed optical modules expected as AI servers proliferate [9][10] - The report notes that China has a leading position in optical module packaging capabilities, with five of the top ten global optical module companies based in mainland China [9] - The domestic market has achieved a 90% and 60% localization rate for 2.5G and 10G optical chips, respectively, but only a 20% localization rate for 25G optical chips, indicating substantial room for domestic substitution [9] Group 3 - The report emphasizes the importance of High Bandwidth Memory (HBM) in addressing the memory bottleneck in AI computing, with HBM offering significant advantages such as high bandwidth, low power consumption, and reduced size [21] - HBM is becoming the preferred solution for GPU memory, with a bandwidth of up to 1TB/s, which is crucial for supporting the rapid growth of AI computing chips [21] - The report suggests that domestic companies are beginning to develop HBM products, which could enhance the local supply chain and reduce reliance on foreign technology [21][24]