IN Brief:
- Kioxia has started sampling 1Tb TLC devices based on 10th-generation BiCS FLASH.
- The 332-layer 3D NAND technology reaches a 4.8Gb/s interface speed and higher bit density than the previous generation.
- The devices are intended primarily for enterprise and data-centre SSDs supporting AI storage demand.
Kioxia has begun sample shipments of 1Tb triple-level-cell memory devices using its 10th-generation BiCS FLASH 3D NAND technology.
The devices are intended primarily for enterprise and data-centre SSDs, strengthening Kioxia’s storage portfolio for AI infrastructure. Production is planned at Fab2 of the company’s Kitakami plant in Iwate Prefecture, Japan.
The new generation uses 332 layers and improves lateral density to increase bit density compared with eighth-generation BiCS FLASH. The technology also reaches a 4.8Gb/s NAND interface speed, while improving write and read power efficiency. Kioxia is using CMOS directly Bonded to Array technology and On-Pitch Select Gate Drain technology, both introduced in earlier BiCS generations, to support the performance and density improvements.
AI infrastructure is reshaping memory demand beyond high-bandwidth memory. Training accelerators depend heavily on HBM, but inference, retrieval-augmented generation, vector databases, checkpointing, model storage, and data staging all increase pressure on SSD capacity, bandwidth, latency, power efficiency, and endurance.
NAND has already moved closer to the centre of the AI infrastructure discussion, with AI inference demand pushing flash supply into constraint. Kioxia’s 10th-generation sampling gives that demand a product-level example, linking higher-density 3D NAND directly with enterprise storage systems built for AI workloads.
The company has also been updating storage products below the data-centre tier, including PCIe 5.0 BG-series SSDs for OEM systems. The 332-layer BiCS FLASH development targets a different class of workload, but the direction is consistent: storage performance, density, and efficiency are becoming design variables across compact systems, client platforms, and AI data centres.
NAND performance is only one part of the enterprise SSD equation. Controller architecture, firmware, error correction, power-loss protection, thermal throttling, endurance management, interface support, and qualification all determine whether a memory device becomes a dependable product in high-duty-cycle use. Higher-density NAND enables larger drives, but it also raises the importance of data integrity and thermal behaviour.
The 4.8Gb/s interface speed addresses the growing cost of data movement. Large AI systems need fast access to embeddings, datasets, model weights, logs, and intermediate data. If the storage tier cannot feed the compute layer efficiently, processors and accelerators spend more time waiting on memory and storage traffic.
Power efficiency also carries more weight as data centres become constrained by electrical availability and cooling capacity. Storage devices occupy rack space, consume energy continuously, and add to facility heat. Improvements in read and write power can support higher storage density without increasing power and thermal load at the same rate.
Controller vendors and SSD makers will still have to qualify the devices inside complete drives, where error correction strength, flash translation layer behaviour, garbage collection, wear levelling, and temperature management determine real deployment performance. Advanced NAND only becomes useful infrastructure when firmware and system design can preserve endurance under sustained workload pressure.
Kioxia’s technology strategy shows a split between cost-conscious and advanced-performance NAND generations. Ninth-generation solutions can serve applications where investment efficiency and cost are dominant, while 10th-generation technology supports more aggressive density and performance requirements. That separation fits a market where not every SSD needs the most advanced NAND, but AI infrastructure can justify it where capacity, efficiency, and bandwidth affect system economics.
The sample phase begins a longer qualification path for enterprise customers. Endurance, firmware behaviour, thermal stability, reliability, and supply availability will shape adoption before volume deployment. Even at this stage, the direction is clear: AI workloads are pulling NAND into a more strategic role, and 332-layer BiCS FLASH is Kioxia’s next step in meeting that demand.
