Memory-hungry AI servers are the target of Kioxia’s newest pitch, expansion cards that pool fast NAND flash into the same address space as DRAM and cost far less per gigabyte.
The Japanese company presented the approach at a briefing in Tokyo on September 3, with Nikkei and other local media covering the event. Its XL1 module is built around XL-Flash, a NAND variant tuned for read speeds far above ordinary flash.
Compute Express Link is the glue. The standard lets a server treat different memory types as one shared pool that software carves up on demand, so capacity can be added without populating every slot with expensive DRAM. Kioxia says its module reads data in under a tenth of the time of earlier products and, versus an all-DRAM setup, can deliver twice the memory for the same outlay while lifting performance by 30 percent.
The pitch lands as AI buyers cope with surging DRAM and HBM prices driven by a supply crunch. Cheaper NAND-based expansion is one of the few ways operators can grow memory pools without paying premium rates, and Kioxia sees the category as an escape from the boom-and-bust cycle of its core storage business.
New AI-focused solid-state drives are also planned, and Kioxia says they could sit alongside Nvidia GPUs in server memory hierarchies, working with HBM rather than against it. Executives at the Tokyo briefing argued that generative AI workloads will need memory capacity that DRAM alone cannot supply at reasonable cost.
