GF SEC: eSSD storage space is vast for AI & storage servers. It is recommended to pay attention to the core beneficiaries of the industry chain.

GF SEC issued a research report, stating that with the rapid growth of long context reasoning, RAG database, and token scale, the demand for high-bandwidth, large-capacity eSSD in AI workloads will continue to strengthen, and the market space for eSSD used in AI servers and storage servers will further expand. According to the above assumptions, the market space for eSSD used in AI servers in 2024, 2025, and 2026 will be 59EB, 89EB, and 120EB respectively. AI servers and storage type eSSD applications will contribute mainly to the incremental space, driving the industry focus towards AI servers and data-driven storage scenarios. It is recommended to focus on core beneficiaries of the industry chain. The main points of view of GF SEC are as follows: AI inference and data storage constitute the core incremental demand scenarios for eSSD. In the AI application system, eSSD is mainly used for training, inference, and data storage. During the training phase, eSSD is used for Checkpoint saving, model loading, and system boot. During the inference phase, the KVCache generated by long context is stored in HBM/DRAM as the preferred hot storage, and hot data exceeding the memory capacity will overflow to high-performance eSSD; in addition, the RAG database is used for large-scale knowledge retrieval and context expansion, also relying on eSSD to provide large capacity and high throughput support. During the data storage phase, eSSD is used for token storage, data cleaning, and ETL processes. With the rapid growth of long context reasoning, RAG database, and token scale, the demand for high-bandwidth, large-capacity eSSD in AI workloads will continue to strengthen, and the market space for eSSD in AI servers and storage servers will further expand. The shipment volume of eSSD used in AI servers continues to grow. The bank has calculated the market space for eSSD used in AI servers based on the following key assumptions: according to the Nvidia NVL72 by HPE white paper, the total capacity of the cache disk tray for each computing tray is 30TB (composed of 8 3.84TB NVMe SSDs), and the operating system disk is one 1.92TB NVMe SSD, that is, one GPU corresponds to two local SSD slots. The bank assumes that one GPU corresponds to two local SSD slots; the capacity of eSSD corresponding to one slot is 4TB. According to the above assumptions, the market space for eSSD used in AI servers in 2024, 2025, and 2026 will be 59EB, 89EB, and 120EB respectively. This calculation reflects the theoretical maximum market space for eSSD used in AI servers, and in actual deployments, some servers may not be fully equipped, so the actual shipment volume may be lower than this number. The shipment volume of eSSD used in storage servers is growing more significantly. The bank has split the eSSD market structure based on the following key assumptions: (1) According to TrendForce data, by 2024, the SSD procurement capacity related to AI will exceed 45EB; according to yole data, the shipment volume of eSSD used in AI servers accounts for 1/3 of the total server eSSD shipment volume; and according to McKinsey data, the market size of eSSD in 2024 is 181EB, from which it is calculated that the market size of eSSD used in general servers, AI servers, and storage servers in 2024 is 90EB, 45EB, and 46EB respectively; (2) According to yole data, the global eSSD shipment volume will have a CAGR of 26% from 2024 to 2030. Structurally, eSSD is mainly used in general servers, AI servers, and storage servers, assuming a CAGR of 12%, 20%, and 54% for the shipment volume from 2024 to 2030, the market size of eSSD used in general servers, AI servers, and storage servers in 2030 is calculated to be 178EB, 134EB, and 614EB respectively. AI servers and storage type eSSD applications will contribute mainly to the incremental space, driving the industry focus towards AI servers and data-driven storage scenarios. Risk warning: The development of the AI industry and demand may not meet expectations; the shipment volume of AI servers may not meet expectations, and the technological progress and product development of domestic manufacturers may not meet expectations.