未知机构：AI储存调研-20260211

Summary of Conference Call Notes Industry or Company Involved - The discussion revolves around the AI and storage industry, particularly focusing on the optimization of agent execution processes and the implications of storage costs on AI performance. Core Points and Arguments 1. Optimization in Agent Execution - During the understanding phase, the most powerful models are used for planning agents, while various models are utilized in the execution phase to support operations. A tool matrix is employed to optimize resource allocation, saving computational power during inference by caching results. The cache hit rate can reach up to 67%, allowing for significant efficiency gains in processing similar queries [2][4]. 2. Storage Architecture - The storage system is structured in layers, including HBM, DRAM, and SSD, to manage hot, warm, and cold data effectively. This architecture is widely adopted in large companies [3]. 3. Cache Hit Rate Dynamics - The cache hit rate tends to improve with increased daily active users (DAU) and user engagement, but it approaches a ceiling of 60% to 70% due to the need for diverse responses in AI services [4]. 4. Data Storage Practices - For consumer users, data is modeled on an individual basis, but common queries can be identified. The system stores both questions and their corresponding key-value (KV) pairs to reduce computational load during the prefill phase [5][6]. 5. Cost Efficiency in Computation - The computational load is generally reduced from a potential 1:5 ratio to about 2-3 times due to storage optimizations, avoiding a simple linear relationship in resource consumption [7]. 6. Rising Storage Costs - The increase in storage prices, particularly for SSDs, is attributed to the demand for long-chain caching solutions. SSD prices are rising faster than DRAM, which serves as a bridge for data [8]. 7. Log Storage and Data Lifecycle - Logs are stored on HDDs, while KV pairs from inference processes are stored on SSDs. The lifecycle of KV data typically requires retention for at least 90 days for long-chain applications [11]. 8. Impact of SSD Read/Write Frequency - High read/write frequencies on SSDs do not significantly affect their lifespan, which is designed to handle several GBs of throughput per second [12]. 9. Government Policies on Chip Imports - Current policies regulate the import of advanced chips like H200, allowing only top enterprises engaged in large model training to apply for procurement, with a focus on narrowing the gap between domestic and foreign capabilities [15][16]. 10. Economic Viability of Storage Solutions - If SSD prices increase to 2-2.5 times their current levels, the cost-effectiveness of storage-based computation will be challenged. New technologies may emerge to mitigate these costs, but significant price hikes could necessitate a reevaluation of pricing strategies [17][21]. Other Important but Possibly Overlooked Content 1. Diverse Supply Chain Strategies - Companies are encouraged to diversify their supply chains to avoid reliance on overseas markets, especially in light of rising prices [20]. 2. Technological Advancements in Cost Reduction - Continuous advancements in AI infrastructure are crucial for reducing inference costs, which is a key focus for cloud service providers [14]. 3. Market Dynamics and Future Predictions - The market is expected to see a shift in the balance of GPU and storage costs, which will influence the overall cost structure of AI applications [21].