On November 12, 2026, IC China was held in Beijing. Luo Xiaodong, Vice President of CXL Technology, pointed out in his speech that the AI computing power wave is reshaping the storage industry landscape, with the industry super cycle entering the "layered" phase; whether enterprises can cover layered demands with full-stack capabilities has become the key to competition.长鑫科技:存储迈入分层时代,全栈布局成胜负手

In his view, the first layer of stratification is reflected in the market structure: servers have become the main driver of storage demand, with their share exceeding 50%; automotive storage is expected to maintain a compound annual growth rate of about 30% by 2030. The second layer is the stratification of AI applications themselves: hot data remains in the cache, model parameters and KV caches reside in the core DRAM layer, while cold data sinks down to SSDs, forming a demand "pyramid." DRAM is precisely at the central hub of this pyramid and is also the fastest-growing segment. This "pyramid" structure implies that storage vendors can no longer rely on a single product to capture the entire market, but instead need to provide differentiated combinations of capacity, bandwidth, and power consumption tailored to the access characteristics of different data layers.

The stratification of demand also signifies a shift in the competition logic. Luo emphasized that the era of merely competing on parameters has passed, and system-level full-stack optimization has become the new battlefield. Different AI stages have varying demands for memory: the Decode stage outputs word by word, with the bottleneck in memory bandwidth; the Prefill stage processes long texts, with pressure concentrated on computing power; and large-parameter multi-card training is constrained by capacity and interconnect capabilities. In other words, storage solutions must deeply align with the specific load characteristics of AI inference and training to unleash performance in real systems.

Based on this assessment, CXL Technology expands its product boundaries through generational iteration and matrix-based layout. On the server side, DDR5 has advanced from 24Gb to 32Gb; RDIMM covers mainstream capacities from 32GB to 128GB, with speeds ranging from 5600 Mbps to 8000 Mbps, while MDRIMM offers 128GB and 192GB specifications, with speeds up to 12800 Mbps. As near-memory connected directly to the CPU, RDIMM's value is rapidly shifting toward the high-capacity segment to meet the growing demand for the core DRAM layer. In the mobile sector, the company globally first mass-produced LPDDR6, with a single chip of 16Gb and a maximum speed of 12800 Mbps, achieving a 60% increase in bandwidth and a 20% reduction in power consumption compared to the previous generation, providing higher energy efficiency support for edge AI applications.

Luo stated that CXL Technology will continue to uphold open cooperation, deeply understand application systems, and develop differentiated solutions to jointly strengthen the domestic storage foundation with industry partners.