Designed for InferenceNarrow moat

Kioxia Holdings (285A) — moat facet

A 245-terabyte drive and a line built to hold a model's key-value cache are the only things here that could convert flash from a gigabyte price into an architectural specification.

The most differentiated thing Kioxia has ever built is not a memory chip. It is a set of drives designed around what an AI model actually does.

The three lines announced on 2 June 2026CM series - TLC, high bandwidthHolds the model's key-value cacheGP series - XL-FLASHOver 100 million IOPS for RAG servingLC series - high capacityTopping out at a 245TB driveTwo of the threeBuilt to Nvidia platform specificationsInference is memory-bound: an accelerator waiting on storage is the costliest idle asset.
The only thing here that could turn flash from a gigabyte price into a specification.

At its investor day on 2 June 2026, Kioxia set out three lines. A high-bandwidth series using conventional three-bit flash, aimed at holding the key-value cache a language model reads from during generation, built to Nvidia's platform specification. A high-performance series using the company's low-latency XL-FLASH, exceeding 100 million input-output operations a second, aimed at retrieval-augmented generation servers. And a high-capacity series topping out at a 245-terabyte drive1.

Each of those targets a distinct bottleneck. Inference is memory-bound rather than compute-bound: the model has to re-read enormous quantities of context and retrieved data, and the cost of doing that from the wrong tier of storage is measured in idle accelerators.

If any of it works, it converts flash from a component priced by the gigabyte into a part specified by an architecture. That is the whole argument for Kioxia being worth more than a commodity multiple.

None of it is in the results yet. The June 2026 quarter was explained by average selling prices rising about 70%2, not by any of these products.

The number to watch is whether these lines are ever broken out as revenue. A product that stays inside the SSD and storage segment is a product that has not changed the company.

Moat trajectory: Widening

The June 2026 investor day set out three product lines built for inference workloads specifically, including a 245-terabyte drive and parts aligned to Nvidia platforms. A year ago none of it existed.

The number that tests this moat
Reported
Revenue guidance, September 2026 quarter
¥2,390.0bn

Drives designed for AI inference are Kioxia's attempt to sell storage rather than flash. Revenue at or above guidance while bit growth stays low would show the premium products carrying price.

Source: Kioxia Holdings first-quarter FY2026 results ↗
⚠ Threats to the moat
References
  1. ReportedThe CM series holds the key-value cache and supports Nvidia's CMX platform; the GP series uses XL-FLASH and exceeds 100 million IOPS for retrieval-augmented generation; the LC series tops out at a 245-terabyte drive.
    Kioxia Holdings — 'Kioxia Announces Growth Strategy for the AI Inference Era at Investor Day', 2 June 2026. The company targets data centre and enterprise sales above 60% of the total over the medium to long term, with annual capital expenditure of approximately ¥470 billion and research and development of ¥230 billion across a three-year plan. The product portfolio comprises the CM Series — high-bandwidth SSDs with TLC flash optimised for key-value cache storage and supporting NVIDIA's CMX platform; the GP Series — high-performance SSDs with XL-FLASH exceeding 100 million IOPS and compatible with NVIDIA Storage-Next for retrieval-augmented generation servers; and the LC Series of high-capacity SSDs including a 245-terabyte model. Tenth-generation BiCS FLASH sample shipments were to begin in summer 2026. The company is securing multi-year long-term agreements to improve revenue visibility and the quality of profit, and will evaluate shareholder returns on the basis of cumulative free cash flow in excess of requirements over multiple years. — medium-term plan · publ. 2026-06-02 · source ↗
  2. ReportedThe June 2026 quarter was explained by average selling prices rising about 70%.
    Kioxia Holdings, first-quarter FY2026 results for the three months to 30 June 2026, as reported from the company's results presentation and earnings call — revenue ¥1,767.1bn, up 76.2% sequentially and 415.5% year on year; non-GAAP operating profit ¥1,326.2bn at a 75% margin; non-GAAP net income ¥887.0bn; non-GAAP EBITDA ¥1,402.1bn; adjusted non-GAAP gross profit ¥1,405.5bn, an 80% margin. Average selling prices rose about 70% with bit shipments growing a low single-digit percentage; eighth-generation BiCS FLASH exceeded 50% of total output. SSD & Storage ¥1,174.7bn, 66% of sales, with data centre and enterprise more than 60% of that and PC-related drives slightly below 40%; Smart Devices ¥525.7bn; Other ¥66.7bn. Operating cash flow ¥866.3bn, days inventory outstanding 102. Cash ¥791.0bn from ¥470.7bn, ¥407.5bn of senior loans repaid, a net cash position of ¥186.7bn against a net debt-to-equity ratio of 120% a year earlier, and an equity ratio of 51%. Guidance for the September quarter: revenue ¥2,390.0bn and non-GAAP operating profit ¥1,900.0bn, again roughly 70% of the growth from price. Capital expenditure of about ¥470bn a year is planned across the three years from FY2026, with roughly ¥200bn of research spending for the year and concentrated investment in the tenth and eleventh BiCS FLASH generations; the company targets roughly 50% long-term agreement coverage for calendar 2028. A buyback of up to 30 million shares (5.5% of those outstanding) for up to ¥800.0bn was authorised for 3 August to 30 October 2026, alongside a three-for-one share split with a 30 September 2026 record date. Calendar 2026 NAND bit growth is put in the high teens, with demand expected to exceed supply in 2027. — quarter to 30 June 2026 · publ. 2026-08-03 · source ↗
Sources
Generated September 23, 2026