SanDisk shares jumped nearly 10% on August 12 after the company and Kioxia unveiled a new generation of QLC 3D flash memory designed for AI and other data-intensive workloads.
The technology delivers higher storage density, faster data transfer and improved power efficiency, addressing a growing challenge for data centers as AI systems generate and process larger volumes of data. The development also highlights the growing role of high-capacity flash storage alongside GPU memory in AI infrastructure.
SanDisk and Kioxia unveil denser QLC flash for AI workloads
SanDisk and Kioxia have unveiled their 10th-generation QLC 3D flash memory, designed to deliver more storage capacity, higher bandwidth and better power efficiency for AI and data-intensive workloads.
The technology uses a 332-layer architecture and achieves more than 37 Gb/mm² of bit density, up to 60% higher than the companies’ eighth-generation technology. Kioxia’s technical data puts the density at 37.6 Gb/mm², the highest reported among QLC products.
A key part of the design is CMOS directly Bonded to Array, or CBA. Instead of building the CMOS logic and memory array together, the two are manufactured separately and then bonded using wafer-to-wafer alignment. This approach allows the companies to optimize each part independently while supporting higher density and more efficient manufacturing.
The new QLC flash also reaches a 4.8 Gb/s NAND interface speed, a 33% improvement over the eighth generation. It uses Toggle DDR6.0 and the Separate Command Address protocol to support the higher transfer rate. The technology also incorporates Power-Isolated Low-Tapped Termination, which helps reduce power consumption during high-speed data transfers.
The bigger improvement is not simply about fitting more storage into the same space. Higher bit density allows more data to be stored on each chip, which can help improve storage capacity and economics in large data centers. For AI infrastructure, where data volumes are growing quickly, that can mean more capacity without requiring a proportional increase in physical space and power consumption.
Power efficiency could be the bigger breakthrough
AI data centers are dealing with more than a storage capacity problem. As AI workloads grow, the power needed to move and process data is also becoming a major challenge. SanDisk and Kioxia have addressed this with interface technologies designed to reduce I/O power consumption while supporting higher NAND speeds.
A key technology is Power-Isolated Low-Tapped Termination (PI-LTT). It uses lower-voltage power alongside the standard 1.2V supply for the NAND interface, helping reduce the energy required during data transfers. According to Kioxia and SanDisk, the technology cuts data input power by 10% and output power by 34% compared with their eighth-generation 3D flash technology.
This is important because the new flash technology also raises the NAND interface speed to 4.8 Gb/s, a 33% improvement over the previous generation. Higher speeds can increase power consumption and heat, so improving efficiency at the same time helps limit the additional thermal and energy load on data-center systems.
TAL’s insight: The bigger story is the combination of higher density, higher bandwidth and lower I/O power. For hyperscale AI data centers, where thousands of storage devices can operate together, even modest efficiency gains at the chip level can become meaningful at system scale. That makes this more than a routine NAND upgrade.
