Kioxia’s AI-inference bet meets the reality of NAND cyclicality

Hiroo Ota, Kioxia’s representative director, president and chief executive since April 2026, is making that allocation decision. Ota joined Toshiba, Kioxia’s predecessor, in 1985; he worked in semiconductor-memory applied technology and technical marketing, led work on NAND applications from 2001, and became executive vice president in 2024. In the leadership announcement, Kioxia says his recent responsibilities included management work and investor communications around the company’s 2024 initial public offering. That operating and customer background helps explain why the company chose him to lead a shift in product mix and capital spending.

Kioxia projects 46% data-center-demand CAGR and 86% inference-AI CAGR through 2028; company-reported forecast. Source: Kioxia Holdings Investor Day.

Inference is a strategic rationale, not proof of a new market structure

Kioxia’s Investor Day materials distinguish inference from model training. The company argues that agentic and physical-AI systems will run more inferences, creating demand for lower-power, higher-speed storage and moving SSDs closer to the active memory hierarchy.

The materials point to NVIDIA’s CMX and Storage-Next initiatives as examples of approaches that use SSDs for externalized key-value caches and retrieval-augmented-generation data. They also describe Kioxia’s own lineup: CM-series drives for bandwidth, GP-series drives using XL-FLASH for high IOPS and low latency, and LC-series QLC drives for capacity. These are the company’s product and deployment claims; the materials do not independently show that those products have changed the economics of the wider NAND market or that NVIDIA has endorsed Kioxia’s commercial outlook.

Kioxia’s target is to make data-center and enterprise products more than 60% of revenue over the medium to long term, excluding SanDisk sales revenue. It says high-value SSDs could improve profitability while smartphones and PCs remain a revenue base. That strategic focus redistributes rather than removes risk: the same presentation says near-term storage demand in those smartphone and PC markets is in a minor adjustment phase, with some forecasts expecting a slight decline in the year.

The expansion plan depends on a company forecast

Kioxia forecasts a tight NAND supply-demand balance through calendar 2027, attributing it to AI-data-center demand and supply constraints. It plans average annual capital expenditure of about ¥470 billion for the next three years, roughly 60% above fiscal 2025, and annual R&D spending of about ¥230 billion, also about 60% above fiscal 2025. The capital plan includes production equipment, cleanroom interiors and infrastructure intended to preserve capacity flexibility; Kioxia says customer long-term agreements will support its supply framework.

This is an investment plan, not a settled industry fact. The presentation says its industry information and projections are based on information available to Kioxia, are not comprehensive and have not been independently verified. Its quarterly filing likewise warns that forward-looking statements can differ materially because of economic trends, market demand and semiconductor competition. The company does not give full-year plans because it describes the memory industry as highly volatile over short periods.

Kioxia nevertheless expects the next step of the upcycle quickly. For the September quarter, it forecasts ¥2.390 trillion in revenue and ¥1.890 trillion in operating profit—35.2% and 48.8% above the June-quarter results, respectively. These are company outlook figures, not results.

What would make the AI case durable

The immediate test is whether Kioxia meets that September-quarter outlook while converting the data-center portfolio shift from a target into a measurable mix change. The more consequential evidence will be whether multi-year customer agreements, product adoption and disciplined capacity spending hold up through a softer smartphone and PC period.

Kioxia’s current disclosures support a narrower conclusion than a permanent market reset: higher prices and shipments have generated exceptional earnings, and the company is funding a move toward higher-value AI-infrastructure storage. Whether that produces more stable returns than past NAND cycles depends on demand and competitive conditions that Kioxia itself says remain uncertain.