AI算力需求：预期差在哪里？

Summary of Key Points from the Conference Call Industry Overview - The conference call discusses the AI computing power demand, highlighting a divergence in expectations between domestic (China) and overseas markets, with the latter showing a rebound in inference demand despite supply constraints [1][2]. Core Insights and Arguments - Divergence in AI Computing Demand: The domestic market remains conservative in its expectations for AI computing demand, while the overseas market, particularly in the U.S., is more optimistic, especially in the West [2]. - Anticipated Turning Point: A turning point in AI computing demand is expected in the second half of 2025, driven by increased dissatisfaction with current computing capabilities among large model developers and the arrival of new hardware [3][21]. - AI Agent Demand: The demand for AI agents is significantly higher than traditional chatbots, with increasing complexity and duration of tasks driving inference demand growth [3][40]. - Market Misconceptions: There is a misconception that training demand has peaked and does not require additional computing power, while actual demand is much higher than market expectations [5][18]. - Global Model Manufacturer Segmentation: Major players in the AI space are categorized into those focusing on model applications (Microsoft, Meta), those on inference computing (OpenAI, XAI), and those pushing next-generation model iterations (Anthropic, Google) [22][25]. - Investment in Computing Power: Significant investments in computing power are crucial for training large models, with OpenAI planning a 400,000-card cluster and Elon Musk aiming for a 1,000,000-card cluster, despite the high costs and uncertain outcomes [27][28]. - Stock Market Reflection: The stock market does not fully reflect the real state of AI development, with current stock prices potentially being high before actual market expectations shift [7][14]. Additional Important Insights - Token Volume as a Demand Indicator: The increase in token volume from major companies like Microsoft and Google serves as a key indicator of rising inference demand [15][17]. - Challenges for Chinese Companies: Chinese internet companies face challenges in large model training due to insufficient computing resources and funding, leading them to adopt a follow-the-leader strategy [30]. - Future Trends in AI Models: The second half of 2025 is expected to see a resurgence in pre-training activities, driven by new hardware and competitive pressures among major players [33]. - Market Dynamics: The dynamics of the stock market and AI development are closely linked, with the need for new narratives to support stock price increases [45][46]. Conclusion - The AI computing landscape is characterized by significant disparities in expectations between domestic and overseas markets, with a potential turning point on the horizon. The demand for AI agents is growing, and major investments in computing power are essential for future developments. The stock market's current performance may not accurately reflect the underlying trends in AI technology and demand. Investors should closely monitor these developments to identify potential opportunities and risks.