As generative AI (GenAI) shifts from large-model training to large-scale inference, Japanese NAND giant Kioxia forecasts continued price increases for NAND and SSD products. Despite planned capacity expansions by various NAND manufacturers in 2027, Kioxia expects no risk of oversupply imbalance due to the rapid growth in AI demand.
Addressing recent industry concerns over MLC NAND end-of-life (EOL), particularly for legacy 2D NAND processes, Kioxia clarified that discontinuations reflect normal technology transitions rather than reallocating capacity exclusively for specific AI products. The company emphasized it will carefully consider consumer supply needs going forward.
Kioxia's chief technology officer of its SSD division, Koichi Fukuda, highlighted that the long-context demands generated by AI inference have made storage a system bottleneck. He identified four core workload drivers behind the fundamental evolution of SSD technology fueled by GenAI.
AI inference redefines SSD's role
Using its latest 9th- and 10th-generation BiCS Flash technologies, Kioxia aims to redefine the role of SSDs within AI architectures. First, deep inference with extended context exponentially increases KV cache requirements. Given the high cost and limited capacity of DRAM and high-bandwidth memory (HBM), KV cache serves as a short-term storage solution for AI models, where acceleration technologies optimize inference speed and improve memory utilization.
Second, Nvidia's Storage-Next initiative demands ultra-fast SSDs featuring scalable performance and large read capacities. Kioxia's GP series SSDs, based on XL-Flash architecture, position SSDs as capacity expansion vehicles for HBM, meeting stringent AI compute speed requirements. The first generation, delivering 10 million input/output operations per second (IOPS), is slated for release in 2026, followed by a second generation targeting 100 million IOPS.
Third, high-capacity QLC storage solutions are gaining traction not only for big data training and checkpointing, but also for inference and data retrieval. Kioxia introduced the LC9 series with E3.L form factors reaching up to 245.76TB and 2.5-inch variants at 122.88TB, already sampling customers.
The fourth driver is that SSDs are expected to increasingly replace HDDs to reduce the total cost of ownership (TCO) for AI applications.
Beyond layers: Kioxia's dual-axis NAND approach
Additionally, Kioxia launched AiSAQ, an open-source software library optimized for vector databases and retrieval-augmented generation (RAG) scenarios, developed in collaboration with NVIDIA cuVS. This algorithmic toolkit reduces reliance on costly DRAM. Testing showed GPU-accelerated indexing via cuVS speeds index construction by 20.4 times while supporting searches across up to 4.8 billion vector parameters.
Kioxia described its product's unique evolutionary logic, diverging from traditional NAND strategies focused solely on increasing layer counts. Instead, it integrates multiple technologies such as density enhancement and power control through a dual-axis approach combining BiCS 9 and BiCS 10. BiCS 10 emphasizes capacity and performance gains via additional layers, whereas BiCS 9 merges existing CMOS Bonded to Array (CBA) technology with advanced packaging to deliver high performance at lower costs.
Balancing data centers and edge AI
Regarding rising storage prices and capacity allocation, Kioxia stressed it will not concentrate all production on data centers. Recognizing the growing demand for running large language models (LLMs) on smartphones, the company plans to launch UFS 5.0 products with 10.8GB/s transfer rates in 2026 to alleviate edge AI read/write bottlenecks. Its supply strategy aims for balanced market coverage, adjusting dynamically by application sector and market trends, especially maintaining close customer engagement in edge AI consumer electronics.
Fab expansion, stable outlook
Kioxia is expanding investments at its Kitakami and Yokkaichi fabs. While some worry about a potential oversupply in 2027, Kioxia remains optimistic that the industry's supply-demand structure will stay stable and robust.
On other NAND makers' moves to discontinue MLC NAND, Kioxia reiterated that phasing out older products like 2D NAND is a standard lifecycle process driven by high production costs and aging equipment — not displacement by AI workloads. The company commits to facilitating customer transitions toward next-generation 3D NAND products.
Article translated by Lily Hess and edited by Jerry Chen
