算力市场专题解读

Summary of Conference Call Notes Industry Overview - The current domestic computing power market is experiencing uneven demand, particularly due to limited resources of NVIDIA's GPU chips influenced by policies and government guidance, affecting companies like Baidu [1][3] - Government policies and strategic projects by state-owned enterprises are driving significant demand for computing power, with project budgets ranging from hundreds of thousands to tens of millions, indicating a hot market [1] - Overall, the demand for computing power remains strong, especially benefiting domestic chip companies [1] Structural Changes in Computing Power Market - The slowdown in innovative teams like OpenAI has led to a decrease in demand for computing power required for model training, while demand for inference, especially in visual and video models, remains robust [1][2] - 2024 and 2025 are expected to be significant years for application deployment, with increased availability and cost optimization driving demand across various applications [1][4] Capital Expenditure Expectations of Major Internet Companies - ByteDance's capital expenditure in 2024 is expected to be no less than in 2023, focusing on model functionality and domestic/international procurement [1][4] - Baidu may rely more on its self-developed Kunlun chips, reducing dependence on foreign chips like NVIDIA [1][4] - Overall, capital expenditures for Tencent, Alibaba, Baidu, ByteDance, and Meituan may decrease but remain at a high level, with a potential reduction of 20%-30% [1][4] Development and Adaptation of Domestic Chips - Domestic chips are widely used in inference but still rely on foreign chips for training [1][4] - The future adaptation of domestic chips will involve collaboration between major companies and chip manufacturers to meet demand and conduct testing [1][7] - The outlook for domestic chips is optimistic due to strong government support and improvements in chip design capabilities, despite existing supply chain and technical challenges [1][4] Overseas Market and Application Prospects - The overseas demand for computing power is substantial, with many companies purchasing GPUs to enhance business performance, although investment may not be as aggressive as expected due to the slowdown of companies like OpenAI [2][8] - Key overseas application areas include AI search, intelligent companionship, education, and smart hardware, showing rapid growth and high revenue potential [2][8] Key Considerations in Chip Procurement - Companies focus on several parameters when purchasing chips, including process technology, FP16 precision, storage capacity, communication performance, and adaptability of operators [5][6] - Domestic chips face challenges in process technology and supply chain, with a need for software optimization to maximize performance [5][6] Future Trends in Inference and Training Demand - Demand for inference is expected to grow significantly, potentially surpassing training demand by late 2024 or early 2025 [6][8] - The focus for training will be on cluster networking, communication stability, and pre-training parameters, while inference will prioritize model distillation, precision adjustments, and memory requirements [6][8] Conclusion - The domestic computing power market is poised for growth, driven by structural changes and strong government support for domestic chip development [1][4][8] - Major internet companies are adjusting their capital expenditures and procurement strategies in response to market dynamics and policy changes [1][4][5]