计算机计算机行业跟踪周报：从字节大会看算力产业的三个趋势

Industry Investment Rating - Overweight (Maintained) [11] Core Views - The edge side is expected to embrace "+AI" as a key landing scenario and traffic entry point for AI ByteDance has already launched AI earphones and is expected to expand its presence in more AI hardware terminals [1][25] - Domestic computing power capital expenditure is expected to accelerate ByteDance's Doubao model has seen a significant reduction in computing power costs, with the visual understanding model priced at 0.03 yuan per thousand tokens, 85% lower than the industry average [11][22] - The structure of computing power cards is changing, with ASIC and inference scenarios increasing It is expected that Doubao's inference scenarios will rapidly materialize in 2025, leading to a surge in demand for inference computing power [12][28] Key Points by Section Domestic Computing Power Capital Expenditure - ByteDance's Doubao model has achieved a daily token usage exceeding 4 trillion, a 33-fold increase from its initial release seven months ago [45] - ByteDance's capital expenditure for 2024 is projected to reach 80 billion yuan, significantly higher than the combined capital expenditure of BAT (Baidu, Alibaba, Tencent) at approximately 100 billion yuan [23] - The rapid development of Doubao is expected to drive a new wave of capital expenditure among major internet companies, with growth in computing power cards, data centers, and related electrical infrastructure [11][46] Edge Side "+AI" - ByteDance has released AI earphones and is expected to expand into more AI hardware terminals, with potential opportunities for ecosystem partners [1][25] - Flashcore Technology launched the first AI glasses, "Flashcore AI A1," priced at 999 yuan on December 19, 2024 [1] - FoloToy, in collaboration with ByteDance's Volcano Engine, Espressif, ToyCity, and Soulmate Technology, announced the AI+Hardware Leap Plan at the 2024 Winter FORCE Original Power Conference [1][25] ASIC and Inference Computing Power - NVIDIA reported that over 40% of its revenue comes from inference tasks, with significant growth potential in the inference chain [30] - Broadcom's CEO Hock Tan predicts that the market demand for custom AI chips will reach 600 to 900 billion dollars by 2027 [50] - Domestic AI chips are continuously iterating, with new products entering the testing phase, and the gap between domestic and overseas computing power is narrowing [31] Investment Recommendations - Domestic computing power capital expenditure increase: Runze Technology, Oulutong, Yakang, Accelink, GDS, and Cloudwise [33] - Edge side "+AI": Intellifusion, Espressif, BES, and DAR Electronics [51] - ASIC and inference computing power: Cambricon, Hygon, Jingjia Micro, Anbotech, Aibru, Digital China, Sugon, and Inspur [34]