According to a research report from CITIC Securities, artificial intelligence (AI) is evolving from "basic chat interactions" to "agent-based functionalities," prompting a substantial surge in context length requirements. Data sourced from Epoch AI reveals that the maximum context window has been expanding at an approximate rate of 30 times per year. Concurrently, the KV Cache memory capacity escalates linearly with the increase in context length, outstripping the growth rate of hardware configurations by a significant margin. At present, leading large-scale model developers and hardware providers are primarily tackling storage bottlenecks through techniques such as quantization, hierarchical storage solutions, and architectural optimizations of models. Nevertheless, the demand for memory continues to skyrocket.

CITIC Securities posits that memory optimization can effectively diminish the cost associated with generating each Token, thereby motivating users to enhance concurrency levels and extend context lengths. This, in turn, will further fuel the overall demand for storage solutions. The enhancement of storage capacity has thus become a fundamental prerequisite for agent-based reasoning processes. Against this backdrop, CITIC Securities maintains a bullish outlook on the growth prospects of the storage industry.