This is Part 2 of a five-part series on semiconductor demand in the age of AI warfare. Part 1 examined the NVIDIA bull-bear debate.

The Signal the GPU Narrative Is Missing

While financial media debated whether NVIDIA’s growth had peaked, Japan’s Kioxia Holdings quietly hit record highs.

Kioxia is not a GPU company. It is one of the world’s largest producers of NAND flash memory — the storage chips used in smartphones, SSDs, enterprise data centers, and AI inference servers. Kioxia’s business should have little to do with the question of whether hyperscalers are over-invested in GPU compute.

And yet there it was: new highs.

This is the market telling us something that the prevailing AI chip narrative has missed. Semiconductor demand in the 2020s is not a single variable labeled “NVIDIA.” It is a complex, multi-layer system, and the layer that may be growing fastest in 2026 is not the one the financial press is focused on.

How AI Infrastructure Actually Works

An AI model like GPT-4 does not live in a GPU. It lives in memory.

During inference — the process of responding to a query — a language model’s weights (the billions of learned parameters that define its behavior) must be loaded from storage into high-bandwidth memory, then fed continuously to the GPU for computation. The GPU itself is fast; the bottleneck is often the memory bandwidth required to keep it fed.

This creates a multi-tier demand structure:

  • HBM (High-Bandwidth Memory) — sits directly beside the GPU die; SK Hynix and Samsung are the dominant suppliers
  • DRAM — system memory that buffers data flowing to and from the GPU
  • NAND flash — long-term storage for model weights, training data, and checkpoint states

As AI models grow larger and AI infrastructure scales, demand grows across all three tiers simultaneously. A new GPU data center requires not just GPUs — it requires HBM, DRAM, and NAND in proportion to the compute being deployed.

Kioxia’s Signal: The Storage Buildout

When Kioxia trades at record highs while GPU-adjacent stocks show uncertainty, the market is pricing something specific: AI data center buildout is continuing, and the storage layer is tightening.

AI training and inference generate enormous data. Training a large language model requires petabytes of training corpora. Running inference at scale requires storing model weights (which can exceed 100 gigabytes for frontier models), caching intermediate computations, and logging outputs for further training. Every GPU rack requires commensurate storage.

Enterprise AI is also creating a new category of demand: retrieval-augmented generation (RAG). RAG systems query large proprietary document stores — corporate knowledge bases, legal archives, medical records — to give AI models access to company-specific information. The storage infrastructure for enterprise RAG is a separate demand driver, largely independent of whether hyperscalers are buying more Blackwell GPUs.

Kioxia’s signal, in other words, is that the storage infrastructure cycle may be running on a different — and later — clock than the GPU infrastructure cycle.

The Broader Japanese Semiconductor Picture

Japan’s semiconductor exposure is predominantly upstream: equipment, materials, and components. This structure has an important property in a multi-layer demand environment: it diversifies across the supply chain.

Tokyo Electron (8035) supplies CVD and etch equipment to fabs producing logic chips (for AI compute), DRAM (for AI memory), and NAND (for AI storage). Its revenue is not dependent on NVIDIA specifically — it is dependent on total wafer production, which is growing across all three categories.

Advantest (6857) tests finished chips, including the memory chips that Kioxia and SK Hynix produce. Higher-bandwidth memory requires more sophisticated testing. Kioxia’s record-high business is Advantest’s record-high backlog.

Shin-Etsu Chemical (4063) and SUMCO (3436) supply silicon wafers to fabs producing chips in all three tiers. As the AI data center stack expands, so does the wafer demand that underlies it.

The point is not that these companies are cheap or that they are obvious buys. It is that the semiconductor value chain is a distributed system, and investors who anchor only to GPU demand are measuring the wrong variable.

The Demand Layers No One Is Discussing

If NAND and DRAM are the second layer of AI semiconductor demand, there is a third layer that has received almost no attention from the financial press.

Military and defense electronics.

The semiconductor content of modern military hardware is increasing rapidly. Electronic warfare systems, guidance computers, secure communications, and autonomous systems all require chips — and they require chips that are often domestically sourced, from fabs with no Chinese supply chain exposure.

In 2024, the United States passed the CHIPS Act specifically to onshore semiconductor production for this reason. South Korea and Japan have followed with their own programs. The implicit argument in all of these policies is that advanced semiconductors have become a national security asset.

The visible manifestation of this trend is the GPU-powered AI labs. But the less visible manifestation — which is just beginning to create investment-relevant demand — is the explosion in armed drone production.

This is the subject of Part 3.

Part 3 — “The Drone Army Revolution: Ukraine’s FPV War and the Hidden Semiconductor Demand” — is available here.

Source: Kioxia Holdings IR materials; company public disclosures | 日本語版

Disclaimer | This article is for informational purposes only and does not constitute investment advice.