Bernstein: Entire AI Memory Chain Worth Watching, Kioxia the Exception

Investment bank Bernstein has published a note telling investors that the AI-driven memory boom is no longer a uniform bet on a single technology. Instead, analysts recommend looking at the entire memory chain, from HBM through DRAM to SSDs and HDDs, since the type of AI workload now determines which segment grows fastest.

Four Kinds of Memory Hunger

Bernstein's note breaks down memory demand by the stage of an AI model's workload rather than by customer type. Training large models requires bandwidth above all, which limits how large a model can be, so what matters most is HBM memory connected directly to GPUs, along with system DRAM and local SSDs and storage resources supporting the whole process.

Inference, meaning the actual generation of responses by a deployed model, splits into two stages with entirely different characteristics. The initial stage, known as prefill, is compute-bound, so HBM and the GPUs themselves are what matter most. The response-generation stage, decode, is instead memory-capacity-bound, since this is when the system has to store the growing KV cache.

KV Cache Bigger Than the Model Itself

The most practical takeaway from the note concerns the KV cache itself. Bernstein points out that at large deployment scale, this cache, which stores the conversation state and context processed by the model, can take up more space than the model weights themselves. The effect hits agentic systems especially hard, since they save intermediate results and communicate with external tools, sharply driving up memory demand.

Retrieval-augmented generation, or RAG, adds another front of demand: substantial SSD and HDD capacity to store knowledge bases, plus a growing role for system DRAM in indexing and searching data. In practice, this means AI-generated demand now extends far beyond the HBM memory that has so far been the market's main focus.

New Technology Layers

In response to this spread-out demand, manufacturers are developing new layers of memory technology. Bernstein cites CXL memory, Nvidia's initiative known as Storage Next, CMX context memory, and high-bandwidth flash memory. The goal of these solutions is to strike a balance between performance, capacity and cost, since no single technology today satisfies all four workload types at once.

Technical barriers remain high - Bernstein analysts, on high-bandwidth flash memory

Winners and One Loser

Bernstein draws concrete investment recommendations from this analysis. Samsung Electronics, SK hynix, Micron, SanDisk, Seagate and Western Digital all received an Outperform rating, a buy recommendation with an expectation of beating the market. The only exception is Japan's Kioxia, which received an Underperform rating because, unlike its competitors, it has no HBM business of its own and also faces growing competitive pressure from Chinese NAND memory makers.

This distinction matters because it shows that simply being a memory maker is no longer enough to win analysts' favor. Exposure to HBM and the ability to support the new memory layers needed for agentic AI deployments are becoming the main criteria separating winners from losers in this cycle.

What It Means for Portfolios

For investors tracking companies like Samsung, SK hynix or Micron, Bernstein's note is a signal not to treat the memory sector as a single, uniform bet on the AI boom. Exposure to HBM, DRAM, NAND and hard drives responds to different phases of the model deployment cycle, and the growing importance of the KV cache in agentic AI could, in coming quarters, shift demand toward segments that haven't been the market's focus so far, such as system DRAM or high-capacity storage.

The note doesn't include new price targets, but it organizes a framework for thinking about the entire memory chain across different AI use cases, making it a reference point for assessing memory makers' upcoming earnings in the months ahead.