Kioxia $KXIAY: Investment Thesis

You missed Sandisk? There's still another one - SanDisk of Japan

While Kioxia is not a small company by any means, it’s surprising how little coverage there is on Substack and Twitter, so I decided to check it out and share my findings.

When I first started monitoring memory companies, it was in May 2025. When I seriously started looking at them, it was in Q4 2025–January 2026. During this time, it was mostly Micron and Sandisk making the headlines. There wasn't a lot of talk about SK Hynix or Kioxia. While I invested in SK Hynix at that time, I didn't pull the trigger on Kioxia until May.

In May, Kioxia had the real inflection point.

May 15, 2026 earnings release.

They blew past guidance.

They had guided FY2025 revenue growth of 27.7%–33.0% and operating profit growth of 57.1%–77.0% and delivered 37% revenue growth and nearly doubled operating profit.

The forward guidance brought even more surprises. Shares jumped after the company said it expects operating profit of ¥1.3 trillion for Q1 FY2026 alone. This is more than the entire prior fiscal year’s record profit.

Let’s dive deeper to understand what’s driving Kioxia’s growth, and, most importantly, whether there’s more upside left in this stock.

For better experience and image quality, consider reading on substack (link in my profile).

Table of Contents

• How it all started?
• What is Kioxia’s business about?
• What is NVM/SSD memory?
• Why NVM and SSDs are critical today
• TAM growth and future demand
• Investor’s Day Recap
• Challenges faced by growing inference systems
• Sandisk relationship
• Financial picture
• Capital allocation and shareholder returns strategy
• Valuation
• Final thoughts

How it started?

Kioxia's roots go back to Toshiba Corporation which invented NAND flash memory in 1987. Following heavy losses at their US nuclear subsidiary Westinghouse, the memory division was spun off in 2017 as Toshiba Memory and officially rebranded as Kioxia in October 2019. In 2018, a Bain Capital-led consortium including Korean, US, and Japanese investors acquired a controlling stake in the company for approximately $18 billion. After several delays, Kioxia completed its highly anticipated initial public offering on the Tokyo Stock Exchange Prime Market in December 2024 under the ticker 285A.

What is Kioxia’s business about?

Kioxia is often referred to as the Sandisk of Japan, which already gives a hint what their business is about.

Kioxia is a pure play memory manufacturer and one of the top three NAND flash producers in the world.

The company develops, produces, and sells non-volatile semiconductor memory (NVM) and solid-state drives (SSDs). Its storage components are critical infrastructure found across multiple end-markets, feeding data into smartphones, personal computers, automotive systems, and most importantly (and focus of our story) massive enterprise AI data centers.

Kioxia operates a massive and capital heavy manufacturing primarily in Japan. As their fabrication requires billions of dollars in equipment, Kioxia shares the costs of factory construction and R&D through a 25 years joint venture partnership with SanDisk.

At the moment Kioxia's business is undergoing a major structural shift due to the generative AI boom. AI data centers and their inference servers require a lot more storage to hold large language models and AI generated outputs. To get the best out of this new business reality, Kioxia is transitioning away from relying on cyclical consumer electronics (like smartphones) and is targeting AI market with intent to increase sales for data centers and enterprises.

What is NVM / SSD memory?

**Non-Volatile Memory (NVM / NAND Flash)

**NVM, specifically NAND flash memory is a storage medium whose contents are permanently retained even when a device’s power is turned off.

The main value of 3D NAND flash memory is its ability to deliver a huge data storage capacity and efficient data transfer. Compared to older 2D planar NAND, modern 3D NAND significantly reduces the energy consumption required for data processing.

2D vs 3D? What’s the difference

2D NAND

: All its memory cells sit in one flat layer on the chip. In order to add more storage engineers had to reduce the size of each cell. They became so small that they started to electrically interfere with each other, which led to data corruption and faster wear. That’s the physical wall 2D hit.

3D NAND

: Instead of shrinking cells, engineers started to stack them vertically in layers. Think of it like adding floors to a building instead of building more rooms onto one floor. This allowed them to get more storage without reducing anything and because the cells aren’t squeezed to their physical limit anymore they also got better durability and lower power use.

What is BiCS FLASH?

It’s Kioxia’s proprietary 3D flash memory technology (Kioxia originally invented 3D flash memory stacking technology in 2007).

Rather than just mindlessly stacking more layers (which becomes expensive and degrades cell reliability), Kioxia optimizes BiCS FLASH using a “Dual-Axis Strategy”.

They combine vertical stacking with lateral scaling (shrinking the area horizontally) and innovative architectures like CBA (CMOS directly Bonded to Array), where the memory cells and the logic circuits are manufactured on separate wafers and bonded together.

This approach allows their latest 8th-generation (218 layers) and 10th-generation (332 layers) BiCS FLASH to deliver massive capacity, industry-leading speeds (up to 3.6 - 4.8 Gbps), and incredible power efficiency while keeping manufacturing costs low.

Solid-State Drives (SSDs)

While raw NAND chips physically store the data, an SSD represents the completely finished storage device built around them. An SSD functions by packaging these NAND flash memory chips together with a controller (System on Chip) that manages all data flow, a DRAM cache for temporary data storage and mapping tables, and a high-speed interface (like PCIe) to connect to the motherboard.

Unlike HDDs that rely on spinning magnetic platters, SSDs have no moving parts which makes them a lot more durable, quieter and energy efficient. Due to advanced controllers and high speed PCIe interfaces, SSD is able to deliver a lot faster input/output operations per second (IOPS), massive bandwidth throughput and a very low latency which eliminates the data loading wait times that usually come with mechanical drives.

The NAND Flash Memory Packages:

These are the black chips on the SSD circuit board where your data is permanently stored. Inside these packages is Kioxia’s

BiCS FLASH

technology.

The Controller (SoC):

This is the “brain” of the SSD. It manages all the data flowing in and out of the stacked NAND chips.

The DRAM Cache:

This acts as a temporary, ultra-fast workspace. It stores the mapping tables (which tell the controller exactly where data is located in the massive 3D NAND stacks).

Why NVM and SSDs are critical today

Both were historically used in PCs and smartphones for fast, reliable storage, but SSDs have evolved into a core component which dictates the performance of entire AI systems alongside GPUs.

As the AI market is moving from training to inference (Agentic/Physical AI) the volume of data processing is increasing exponentially. This creates a massive system bottleneck. GPUs and their built-in High Bandwidth Memory (HBM) lack the capacity to hold large datasets needed for these continuous computations. To help with this problem, data centers today are redefining SSD and are using it not only as a storage, but also as an extended memory layer for GPUs.

How they are used in AI Data Centers:

• Extended GPU Memory: Frameworks like NVIDIA’s Context Memory Storage (CMX) utilize high-bandwidth SSDs to temporarily expand the GPU’s cache. This allows the system to efficiently store past computational results on the SSD, bypassing strict GPU memory limits and keeping expensive GPU clusters fully utilized without waiting for data reloads.
• Retrieval-Augmented Generation (RAG) Servers: RAG servers must instantly search massive external databases to improve AI response accuracy. Because DRAM capacity is too limited and HDDs are far too slow, data centers are utilizing super-high IOPS SSDs to act as extended memory and achieve fast search speeds.
• Massive Output Storage: AI models generate a large amount of output data. High-capacity SSDs are being deployed as the foundational infrastructure to store this rapid surge of AI-generated content efficiently in data centers.

TAM growth and future demand

The insatiable hardware requirements of hyperscalers are currently driving AI data center build-outs faster than the NAND supply chain can keep up.

Two categories of data exist for NAND demand growth: unit/bit demand forecasts and dollar-value market size forecasts. They diverge significantly and should be reported separately.

Bit demand

The company guides for flash memory bit demand to grow from 997 exabytes in 2025 to 1,807 exabytes in 2028, a 22% CAGR. Within that total, Kioxia forecasts data center demand rising from 295 exabytes (30% of total demand) in 2025 to 909 exabytes (50% of total demand) in 2028.

Kioxia’s Integrated Report 2025 separately states an expected ~20% CAGR in flash memory demand from 2025 to 2029, citing TechInsights’ NAND Market Report Q2 2025 as the source. The two figures are consistent with each.

Dollar-value market size

These four estimates span a $19B range for essentially the same base year, and CAGR estimates range from 5.3% to 11.0%.

For comparison, Counterpoint Research reported global NAND market revenue of $46B for Q1 2026 alone, up 246% year-over-year. Annualized, that quarter alone approaches or exceeds several of the full-year 2026 estimates above.

Management’s estimation

1. Overall NAND Market Growth 22% CAGR
2. Data Center 46% CAGR
3. Inference as the Primary Catalyst 86% CAGR

We covered what is Kioxia about. We explained their products. We talked about the importance of NAND today, and estimates for its TAM expansion.

Now, let’s talk about the Investor Day. The company shared a lot about its strategy, capex, growth plans, shareholder value creation, new products, current issues and so on.

I believe there’s a great value in learning more about Investor Day topics to understand more about Kioxia’s story.

On Kioxia’s investor relations page you can find the full presentation, the transcript and investor day Q&A. I focused on what I thought was the most relevant for this investment thesis, and I did not include all the slides.

Investor’s Day Recap

Transition from a cyclical hardware manufacturer to a structural backbone of global AI infrastructure

On June 2, 2026, Kioxia hosted its Investor Day under the theme "Flash Memory Scales AI Inference". This event marked a definitive "phase change" for the company, where it outlined its transition from a cyclical consumer hardware manufacturer into a structurally stable and highly profitable backbone of global AI infrastructure.

What did they talk about during the presentation?

Kioxia is doing strategic shift in the revenue mix away from commodity smartphone and PC memory and toward Data Center and Enterprise.

The target is to have over 60% of revenue from Data Center & Enterprise, medium to long term which is quite high from where the business is today.

Why they’re focusing Data Centers and Enterprises? See on the next slide.

Total flash memory demand: 997 exabytes (2025) → 1,807 exabytes (2028), a 22% CAGR.

The growth isn’t evenly spread. Data center demand 295 EB in 2025, growing to 909 EB by 2028 is compounding at 46%, more than double the market-wide rate.

Smartphone and PC demand is described in the deck as flat or slightly declining near-term.

Data centers are roughly 30% of total flash demand in 2025 to exactly half by 2028.

Within data center demand, inference workloads are growing at an 86% CAGR, against just 16% for training. That’s the most accurate way to confirm the training-to-inference shift. Kioxia is saying that the inference alone is growing more than 5x faster than training within their own demand model.

Supply demand (left)

Kioxia’s own supply-to-demand ratio starts above 100% in Q1 2025, then falls around 93-95% through late 2025 into early 2026, and only gradually recovers back toward the 100% line by Q4 2027, without fully closing the gap in the window shown. A ratio below 100% means demand is running ahead of supply.

Management’s own label on the slide is direct: “Tight supply-demand to persist throughout FY27.”

Revenue base (right):

The bars showing 2024 and 2025 sit flat and very low. 2026E jumps sharply, and 2027E jumps even more. Chief Strategy Officer Junichiro Yaguchi said several research firms estimate NAND market revenue will be roughly four times larger in 2026 than 2025, with that trend continuing into 2027.

The two charts tell a single story from different angles. We have the shortage (left chart) as the cause, and the revenue explosion (right chart) is the effect.

Here are two slides about Kioxia’s strategy.

• Kioxia plans to spend ~¥470B ($2.89B) annually on capex over the next three years, a 60% jump from FY2025, driven by AI/datacenter NAND demand.
• Growth is being locked in via customer LTAs for stable demand, while management commits to keeping capital efficiency above the industry average.

“We are targeting an average annual reduction in front-end cost per gigabyte (GB) in the 10% range. Through the migration of BiCS FLASH™, we have factored in bit density improvements of 50% or more per each new generation.During transitions between product generations, we don’t base our key performance indicators solely on the layer count; rather, we plan to achieve competitive cost reductions by pursuing the optimal mix of 2D shrink (lateral scaling) and layer count.”

Key info here is roughly 10% annual reduction in front-end cost per gigabyte. That’s the number that drives margin expansion regardless of NAND pricing, meaning if NAND price stays flat, it would push their profitability higher.

Each BiCS FLASH transition (as mentioned on the slide: Gen.6 to Gen.8 to Gen.10) comes with a density jump of 1.5x to 1.6x per step. More bits per die with the same manufacturing footprint and lower cost per gigabyte.

The real target is the optimal combination of layer count and 2D shrink (shrinking the cell horizontally, the older 2D-NAND lever, still in play alongside vertical stacking). That’s a deliberate move by Kioxia as they’re trying everything they can to optimize cost per bit.

Challenges faced by growing inference systems

We’ve established that data center demand is what’s driving this NAND cycle and we also covered that inference is growing faster than training. Now we’re going one level deeper. What does inference actually do to a storage system, and why does it need SSDs specifically to solve it?

Kioxia’s SSD department talked about this question during the investor presentation. Here is their answer in three parts:

• the problem inference creates
• the new SSD use cases that problem unlocks
• and the specific products Kioxia built to solve these problems

A basic AI query used to be simple. GPU ran one inference, then looks something up if needed and returns an answer. The storage didn’t stress a lot, even old HDDs were fine for the data sitting there.

Agentic AI changed things. Today, it’s not one single inference. There’s a chain of them, thinking, checking, repeating. Due to this, two problems show up.

• KV cache problem - every time a model processes information, it builds a memory of everything in the current session - this is KV cache. In a single query it’s usually small enough to sit inside GPU’s own memory. In the Agentic AI chain of events the cache keeps growing and at some point it can’t fit the GPU memory anymore. When this happens, the system has to “throw” the old cache and reload it later, and that reloading becomes a bottleneck slowing everything down.
• RAG problem - DRAM memory used for this simply doesn’t have the necessary capacity.

RAG (Retrieval-Augmented Generation) is an AI framework that grounds Large Language Models (LLMs) in external, proprietary, or real-time data before generating a response. Instead of relying solely on its static training data, the AI searches a custom knowledge base for relevant facts and uses them to answer your questions.

Kioxia has an SSD based fix for inference problems we mentioned earlier.

KV cache overflow - instead of losing that cached data or being limited by GPU memory size, move the data on an SSDs located close to the GPU. There are two technologies enable this:

• Context Memory Storage (CMX), which treats SSDs as an extension of the KV cache.
• NVIDIA’s Storage-Next, which lets the GPU talk to the SSD directly (skipping the CPU to avoid slowing down the procesS).

RAG at scale - instead of trying to force a huge knowledge base into limited DRAM, Kioxia uses SSDs that are large and fast to be able to search quickly.

This means that the SSD became an active memory system which sits between the GPUs’ tiny and fast HBM, and slow old school bulk storage. It does job that neither of those two can do at the same time - to be fast enough to keep up, but cheap enough to hold enough data.

Kioxia’s has three purpose-built SSD lines which aim to solve problems we covered on the previous slide.

• CM Series handles the KV cache overflow problem. It is high-bandwidth, built for the constant swap traffic between GPU and storage as agentic AI’s context grows.
• GP Series handles the RAG problem. It is ultra-fast, low-latency, positioned to sit right alongside HBM as a cost-effective extension of it.
• LC Series handles the data volume problem. It is built to economically absorb the flood of data AI systems generate and consume.

If you enjoyed the read so far, please like and share!

Sandisk relationship

Kioxia and Sandisk have a 25+ strategic partnership centered around a massive manufacturing and R&D JV. While they collaborate closely on production, they also operate as direct competitors in the market for flash memory and storage products.

Here are the key details of their relationship:

The manufacturing JV

• Kioxia and SanDisk jointly invest in manufacturing equipment installed at Kioxia’s Yokkaichi and Kitakami plants in Japan. They are splitting capex for semiconductor fabrication which is giving both of them ability to manufacture at a scale, something that would be difficult to achieve alone.
• R&D and its costs for NAND flash memory are shared equally, while Kioxia actually owns the physical facilities and controls 100% of the manufacturing operations.
• The JV entities sell 50% of their processed wafers to Kioxia and 50% to SanDisk. However, because Kioxia also possesses its own independent capacity, the overall production capacity at the these facilities is split 60% for Kioxia and 40% for SanDisk. The two companies account for 30% of global NAND production.

Agreement Extension

• On January 29, 2026, the companies announced a formal five-year extension of their joint venture agreements for the Yokkaichi and Kitakami plants, which will now run through December 31, 2034.
• As part of the contract extension, SanDisk is paying Kioxia $1.165 billion in cash. This compensates Kioxia for manufacturing services and guarantees long-term product supply allocation for SanDisk.

Financial picture

Kioxia indeed reminds me of SanDisk. Not just due to its business, but also due to its hyper-growth, both on the business side and in its stock price.

Stock info as of today.

• Price ¥76,260
• Shares outstanding 546.3 million
• Market cap ¥41.66 trillion ($256B)

Revenue

FY2025 saw a lot of growth acceleration. The year ended with revenue up 37% YoY.

Now let’s look at operating profit. It ran way faster than revenue, and this happened in Q2, Q3, and Q4, with Q4 being the strongest at no less than 317% QoQ growth. Full year operating profit was ¥870.4B, up 92.7% YoY from ¥451.7B in FY2024 which is nearly double.

Operating margin expanded from 13.1% in Q1 to 59.5% in Q4, with the full-year margin at 37.2%, up from 26.5% in FY2024.

For Q1 FY2026 (April–June 2026), Kioxia’s own guidance calls for revenue of ¥1,750.0B (+74.5% QoQ) and Non-GAAP operating profit of ¥1,300.0B (+117.0% QoQ).

Kioxia’s Q1 FY2026 results are scheduled for release July 31, 2026.

Q4 Results

Guidance vs results - passed with flying colors.

But here’s the most interesting part.

Q4 was inflectionary. Sales almost doubled, led primarily by SSD & Storage segment.

Maybe you’re thinking, okay after this much growth, it has to slow down?

Here’s the guidance for the next quarter.

Operating profit, again, triple digit growth. Net income, same story.

Capital allocation and shareholder returns strategy

This is what management presented during the investor day:

• Driven by an unprecedented projected 74% operating profit margin for Q1 FY2026, Kioxia expects to eliminate its net debt and achieve a net cash positive position by the end of that quarter.
• To capture AI-driven growth while avoiding oversupply, Kioxia plans to allocate approximately 470 billion yen annually to capex and 230 billion yen to R&D over the next three years. In their long-term financial model, they intend to cap capex at around 20% of revenue and maintain R&D spending at 8-9% of revenue.
• The company is targeting the highest ROIC among manufacturing companies, noting that their trailing 12-month ROIC has already exceeded 60%. Management emphasized strict investment discipline and committing to fund only projects that surpass their internal hurdle rates.
• Kioxia is transitioning from 1y spot contracts to LTAs. This stabilized model targets an average operating margin in the mid 20% range and an annual per-gigabyte cost reduction in the mid 10% range.
• Kioxia plans to initiate progressive dividends and potential share buybacks starting in FY2027. Management provided a benchmark of returning up to 50% of their excess cumulative free cash flow to shareholders, though they noted this ratio will remain flexible depending on future M&A opportunities and growth investments.

What remains is to talk about the valuation and my own summary.

You can find those on my substack. Link can be found on my profile.

If you found the article useful, please consider giving it a like & share.

Thank you!