AI storage flash memory has a 332 layer leader at Kioxia while Samsung, SK Hynix, and Micron spent two years investing in HBM, the stacked DRAM next to AI training GPUs.
Kioxia began shipping engineering samples of its 10th-generation BiCS FLASH on July 3, a 1-terabit triple-level-cell 3D NAND chip with 332 active layers and a 4.8 gigabit-per-second interface. The chips are built for the enterprise and data-center SSDs that hold training checkpoints between GPU runs, model weights in production, and the inference logs that flow through the same flash during serving (Kioxia news release).
That category, the flash memory layer under the GPUs and high-bandwidth memory (HBM) that dominate AI coverage, is what EE Times Asia calls the next battleground for AI infrastructure spend. Kioxia's news release leads with the 332-layer count and the 4.8 Gb/s interface. The more useful read is who chose to invest in scaling that count while the rest of the memory industry looked the other way.
For the past two years, the three memory companies treated as inevitable AI winners, Micron, Samsung, and SK Hynix, redirected engineering and capital toward HBM, the stacked DRAM chips that feed GPUs. NAND flash was the lower-margin cousin. The capital math made the redirection rational: HBM ships in higher-margin, multi-quarter contracts than commodity flash, and the 2024-25 NAND oversupply punished anyone running at the back of the cost curve. Kioxia kept pushing its NAND roadmap through the downturn instead. IwaiCosmo analyst Kazuyoshi Saito frames the consequence as Kioxia being roughly two years ahead of peers in stacking layers, a layering lead presented as analyst opinion rather than benchmark-verified fact (EE Times Asia).
EE Times Asia, citing a Reuters interview with consultant Satoru Oyama, a former Tokyo Electron executive, frames the gap as the residue of a multi-year NAND under-investment cycle that concentrated AI-era NAND demand on the supplier that did not pull back. That is the mechanism. The lead is the product of choices inside competitor balance sheets, not a permanent technical moat.
The structural effect is concentration. The AI storage lane of the memory market now has a clear leader, and it is a single Japanese supplier whose IPO has been delayed multiple times during the cycle that produced its lead.
The technology is part of the story, with caveats. The 10th-generation BiCS uses CMOS Bonded Array (CBA) wafer bonding, a process that manufactures the control logic and the memory cells on separate wafers and then bonds them together, lifting both speed and density versus the older approach of stacking both on a single wafer. Kioxia's reported 4.8 Gb/s interface speed is what comes out of that architectural split. A Kioxia-patented OPS architecture splits NAND planes into independent sub-operations so reads and writes can overlap (Kioxia news release). The 332-layer count and the 4.8 Gb/s interface are Kioxia's reported specifications. No independent benchmark or teardown appears in the public source basis as of Friday.
Volume production is gated by customer qualification, and Kioxia has not named SSD partners. The corporate story behind the roadmap is not a straight climb. Kioxia is the 2018 Bain-led carve-out of Toshiba's memory business, with an IPO delayed multiple times as the NAND market moved from shortage to glut. The current cycle is the first time in years the company has held both technological and pricing leverage at once.
NAND pricing tracks inventory cycles measured in quarters, and the same under-investment that redirected AI-era demand onto Kioxia now shows up in sell-side notes as the setup for the next capacity wave. Samsung and SK Hynix retain sizable NAND franchises, and the relevant comparison is capital priority across product lines. Omdia analyst Akira Minamikawa, cited in the same coverage, expects hyperscaler sourcing decisions in 2026 to be shaped by Kioxia's lead, though those decisions can shift on a single quarter of supply normalization.
Customer qualification is the next marker. Kioxia said samples shipped Friday. The company has not named SSD partners, and public SSD platforms built on 10th-generation BiCS are not expected before partners are named or volume production is committed. If the layering lead holds, AI-storage NAND pricing has a single dominant supplier through the second half of next year. If the lead compresses, it compresses against the same HBM-priority decisions that created it.