SanDisk Patent Stacks AI Processors on NAND to Break Memory Wall

The rapid growth of generative AI model parameters has allowed GPU computing power to outpace data-transfer capacity, creating a “memory wall.” HBM offers high bandwidth but is constrained by capacity and cost, while NAND provides greater capacity at lower cost but has higher latency. SanDisk is therefore using CBA technology to bring computing and storage closer together, targeting the AI inference market.

SanDisk Technologies filed for U.S. patent US 12,430,274 B2 on November 27, 2023, and received approval on September 30, 2025. The architecture places an AI processor or GPU directly on CBA NAND dies with capacities of up to about 2TB, with HBM positioned on both sides and roughly 200,000 interconnects carrying data between them. The company has not announced a mass-production timetable or investment amount.