CITIC Securities: Storage capacity upgrade is the core requirement for current intelligent reasoning; firmly optimistic about the growth trend of storage

People’s Finance Network, March 31—CITIC Securities’ research report says that AI is evolving from “simple conversation” to “agents (Agent),” driving a surge in context length. According to Epoch AI data, the longest context window grows by about 30x every year on average; the relationship between KV Cache GPU memory capacity and context length is linear, far outpacing the growth rate of hardware configurations. At present, major large-model vendors and hardware vendors mainly address the compute/memory capacity bottleneck through quantization, hierarchical storage, and model-architecture optimization, but this still does not change the explosive trend in GPU memory demand. CITIC Securities believes that GPU memory optimization could reduce the cost of generating each token, thereby encouraging users to run higher concurrency and longer contexts; total memory capacity demand will not decrease but instead increase. Memory capacity upgrades are set to become the core need for current Agent inference, and it remains firmly optimistic about the growth trend in storage.