Morgan Stanley Asia Pacific Summit, November 18, 2026: CEO David Goeckeler and CFO Luis Visoso on durability-driven NAND demand, the NBM contract structure and the Kioxia JV

Sandisk executives used a fireside chat with Morgan Stanley's Joseph Moore to describe a shift in how hyperscalers specify NAND: away from density alone and toward durability, power efficiency and bandwidth. CEO David Goeckeler said customers are now asking to speak with Sandisk's NAND engineers about durability, and that the request is opening a new category of demand in single-bit-per-cell (SLC) flash and a new product called high bandwidth flash (HBF).

NAND Becomes a Multi-Attribute Technology

Goeckeler said inference is pulling NAND into parts of the memory architecture that DRAM previously occupied. His reasoning was economic: "If I'm going to scale it, it needs to be the most economic answer, and that's going to drive you to NAND." He noted that his team estimates total global DRAM production at less than 100 exabytes, while NAND is a zettabyte market. He put DRAM at 20x to 40x the cost of NAND. Use cases he cited include KV cache, staging and model post-training.

The change in customer requests is what he called most notable. "NAND was always deliver me density, right? TLC, QLC, more layers, more density. Now it's we need durability. If you can deliver me durability, I can replace DRAM in this part of the architecture," he said, adding that this does not mean replacing all DRAM. He described SLC as offering more than an order of magnitude more durability than TLC. In agentic workloads, he said, a CPU's scratch pad is written far more often, making DRAM overspecified and scarce, while SLC NAND could be good enough. "People are like, SLC NAND is back, really?" he said.

Goeckeler said these use cases could drive tens of exabytes of demand over the next couple of years. Any Sandisk product can be configured to run in SLC, so no new node or die is needed. He offered a supply illustration: if the industry planned 20% supply growth a couple of years out and 1% to 1.5% of that became SLC demand, effective supply growth would fall to roughly 16%. "The impact is quite significant," he said. He called it "completely orthogonal to this idea NAND is just a commodity," and noted that durability differs widely across suppliers because it depends on the electrical properties of the cell. "This is not a commodity," he said.

When Moore suggested everything is in short supply at once because NVIDIA reduced rack memory content, pushing demand to fast local storage, Goeckeler partly agreed but said the primary driver is system-level cost optimization at scale, with scarcity secondary. "You have to be relentlessly focused on cost" when a technology is being deployed to billions of people, he said.

High Bandwidth Flash: Die Taped Out, Samples Next Year

Goeckeler gave the most specific update yet on HBF. Sandisk has taped out a die, which is now in the fab and will return next year. A controller is being built to go with it. The aim is to put the product in customers' hands to learn the use cases and the scale of demand. He said the design targets inference-appropriate bandwidth, improved durability and 8x to 16x the density of the alternative. "We're not trying to replace DRAM," he said. "What we're trying to do is get the attribute of our technology good enough to solve the use case."

He cautioned that HBF is not plug compatible. "You've got to change the rest of the system a little bit to accommodate our technology." Moore noted that SK hynix devoted about a third of its memory presentation at the Open Compute Project summit a month earlier to the same concept. Visoso said HBF is not included in current customer contracts, but the relationships are tight enough that follow-on discussions come naturally through deeper quarterly business reviews.

Data Center Mix and Client Weakness

Goeckeler said Sandisk's data center business grew 1,700% year over year, rising from a small portion of revenue to close to half. He joked that repeating that growth next year would be difficult. He said 2026 will see more than triple-digit growth in data center exabytes and that data center will exit the year as more than half of the NAND market. He described Sandisk's history in three arcs: a strong consumer brand sold through 350,000 points of presence, a client era in which it held roughly 25% client share against 13% to 14% bit share, and now data center, supported by an enterprise portfolio built over many years.

He was direct on the pressure in PCs and smartphones. "If it's dependent on basically very inexpensive NAND, that's going to be very challenged," he said, adding that Sandisk sees it in its own business and is adjusting. Visoso explained the mechanism: older nodes mostly serve PCs and mobile, while newer nodes can go anywhere. Sandisk prioritizes newer nodes for data center and attractive edge segments and lowers production on older nodes that are not selling at attractive prices. Goeckeler said automotive exabyte growth is twice the market rate over the next several years and that customers are discussing physical AI....