Kioxia Holdings Corp "Peak is not here yet" (Buy) Yasui

研报英文原文证据摘录

Kioxia Holdings Corp "Peak is not here yet" (Buy) Yasui

Kioxia Holdings Corp UBS Research

Pivotal Questions

Q: Is AI inference a temporary driver of NAND

flash memory demand, as in the case of AI

training, or is it a structural growth driver?

UBSVIEW

We view AI inference as a sustainable and structural driver of growth in NAND flash

memory demand. For AI inference, NAND flash memory’s performance and

capacity are design variables that define user experience. This role is unlikely to

change.

EVIDENCE

AI inference has evolved as a real-time service that responds instantly to end-user

actions. With the implementation of RAG and agentic search, architectures are

being explored and introduced to offload and restore KV cache data stored in GPU

HBM using high-performance NAND flash memory/SSDs. More relevant themes are

well explained by a recent APAC focus note by Nicolas Gaudois

WHAT´S PRICED IN?

The market appears to view storage demand for AI inference as sporadic, tied to

the number of servers deployed, as in the case of AI training. Consequently, NAND

flash memory is still perceived as a peripheral storage component. However, NAND

flash memory actually complements HBM for AI inference, functioning as a core

component that constrains response speed and the scale of models that can be

handled. This shift in role is not yet fully priced into the stock, in our view.

AI inference at a point where memory matters more than

computation

Unlike AI training, AI inference does not conclude with a single series of computations

because when users continue a chat, AI models need to retain the earlier context. This

memory is stored in the KV cache.

AI models process numerical data (attention) representing word importance,

relationships, and context.