In 1987, Toshiba engineers invented NAND flash memory—storage that keeps its data after power disappears. Nearly four decades later, the business descended from that invention looked like a case study in Japan’s semiconductor decline. It had been sold during its parent’s crisis, postponed an IPO, failed to merge with Western Digital and accumulated enormous losses in a smartphone slump.
In 2026 the story reversed. Artificial intelligence needs more than processors. Model weights, conversations, retrieval databases and inference output must be stored at scale, reached quickly and moved with tolerable electricity. NAND, long treated as a commodity, is becoming a strategic layer of the AI system. Kioxia has begun sampling its tenth-generation BiCS Flash into that change.
HBM is not AI’s only memory
High-bandwidth memory beside GPUs receives the attention. HBM and DRAM are extremely fast, but expensive, power-hungry and capacity-constrained. NAND is slower, yet far cheaper per bit and nonvolatile.
Training requires datasets and checkpoints. Inference accumulates model weights, retrieval documents and vector databases, user context, previous computation called KV cache, plus generated images, video and logs. Keeping it all in HBM is economically impossible.
Kioxia wants SSDs to move from slow repository toward extended memory. Frequently needed information sits on fast flash and reaches GPUs through high bandwidth and many small reads. HBM, DRAM, XL-Flash, TLC, QLC and hard drives form a hierarchy whose placement determines AI cost.
Inside generation 10
The samples announced July 3 are one-terabit TLC devices, storing three bits per cell. Kioxia stacks 332 layers and raises bit density 59% over generation 8. The 4.8 Gb/s interface is 33% faster. Including transfer, write and read power efficiency improve 18% and 30%.
Layer count is only part of the design. CMOS directly Bonded to Array manufactures the memory-cell wafer and control-circuit wafer under separately optimized conditions, then bonds them. On-Pitch Select Gate Drain technology removes unused memory holes, shortening bit lines and reducing word-line capacitance. Scaling happens vertically, laterally and in circuit arrangement.
These are functional samples; mass-production specifications may differ. The chips are intended primarily for enterprise and data-center SSDs, with production planned on new equipment at Kitakami Fab2 in Iwate. Sampling marks technical readiness for customer evaluation. Yield, qualification, timing and order volume remain separate tests.
The inventor of NAND also pioneered the third dimension
Flash memory took its name from the flash-like speed of block erasure. Toshiba invented NAND in 1987, began the world’s first mass production in 1991 and established Yokkaichi operations in 1992. Small size, no moving parts and data retention without power enabled digital cameras, music players, phones, USB drives, smartphones and SSDs.
As shrinking cells on a plane approached physical limits, the company announced the world’s first 3D flash technology in 2007. BiCS—Bit Cost Scalable—stacks electrode plates and drills vertical holes to create cells through many layers. It mass-produced 48 layers in 2016, 64 in 2017 and 96 in 2018. Generation 10 reaches 332.
There is a historical irony. Japan retreated in DRAM, logic and consumer-electronics semiconductors, yet the foundational NAND invention and early 3D breakthrough remained Japanese.
Memory sold to save Toshiba
Toshiba’s 2015 accounting scandal and catastrophic losses at U.S. nuclear subsidiary Westinghouse threatened insolvency. To repair its balance sheet and preserve its listing, Toshiba separated its crown-jewel memory operation. A Bain Capital-led consortium bought Toshiba Memory for roughly ¥2 trillion in 2018.
The company became Kioxia in 2019, combining Japanese kioku—memory—with Greek axia—value. Its Toshiba sign disappeared; Yokkaichi and Kitakami manufacturing and the long joint-development relationship with SanDisk remained.
The transaction was condemned as Japan Inc. selling a technological jewel. Independence also removed memory investment from a conglomerate balancing nuclear, infrastructure and electronics demands, allowing a specialist to judge multibillion-dollar fabrication spending on its own industry cycle.
The postponed IPO and merger that never came
Kioxia planned what could have been Japan’s largest IPO in 2020, then postponed it amid U.S.–China friction, Huawei restrictions and memory conditions. From 2021 it intermittently negotiated a roughly $20 billion combination with Western Digital. The partners shared fabs while competing in sales; talks stalled in 2023, complicated by SK Hynix opposition.
The 2023 NAND collapse was brutal. Smartphone and PC demand fell, inventories rose and prices broke. Kioxia lost ¥100.8 billion at the operating level in the July–September quarter, followed by ¥65 billion the next quarter. Producers cut output and waited for inventories to normalize.
Kioxia finally listed in Tokyo on December 18, 2024. Its roughly ¥750 billion IPO valuation was far below the 2018 purchase price. It was a disappointing exit—but an entry into public markets just before AI began redefining storage demand.
Reading the comeback in numbers
Fiscal 2025 revenue reached ¥2.338 trillion and non-GAAP operating profit ¥876.2 billion, a 37% margin. Reuters reported accounting operating profit of ¥870.4 billion, up 92.7%. Kioxia forecast ¥1.3 trillion of operating profit for April–June 2026 and expected to reach net cash.
AI is not the entire explanation. Higher NAND prices, supply discipline, yen weakness and inventory normalization also drove the reversal. Calling every yen of profit “AI” confuses a cycle with a secular shift. A 12% one-day share decline after reports that OpenAI might delay its IPO showed how elevated—and fragile—expectations had become.
Memory companies repeatedly misjudge demand. When every producer expands during a boom, oversupply and price collapse arrive years later. Strategic value will endure only if long-term agreements and premium SSDs reduce dependence on spot NAND pricing.
Three SSD tiers for the AI memory hierarchy
| Product | Memory | AI role |
|---|---|---|
| CM Series | High-bandwidth TLC / BiCS | KV cache and bandwidth between GPUs and storage |
| GP Series | Low-latency XL-Flash | Ultra-high IOPS, RAG and GPU-memory expansion |
| LC Series | High-density QLC | Models, data lakes and generated-data capacity |
Kioxia aims to lift data-center and enterprise sales above 60% over the medium to long term. CM Series supports Nvidia’s Context Memory Storage approach, moving KV cache to SSDs. GP Series supports Storage-Next and, according to Kioxia, targets more than 100 million IOPS. LC Series includes a 245 TB model.
Capacity, bandwidth, IOPS, latency, endurance, power and price trade against one another. Competitive advantage comes not from one NAND type but from assigning SLC-like XL-Flash, TLC and QLC to the correct tier.
Why generation 9 and 10 run together
Kioxia calls its approach a dual-axis strategy. Generation 9 pursues high performance with relatively low incremental investment; generation 10 uses advanced stacking for capacity and density. The highest generation number is not automatically the best economic product for every customer or fab.
The approach is rational and complex. Parallel processes, controllers and qualifications divide engineering attention. Against Samsung, SK Hynix/Solidigm, Micron and China’s YMTC, victory depends on yield, cost, SSD firmware and customer support—not layer count alone.
What AI turned into a strategic asset
- Opportunity: always-on inference multiplies storage reads.
- Opportunity: expensive GPUs improve the economics of SSD memory extension.
- Risk: NAND oversupply and another price collapse.
- Risk: efficient models reduce capacity below forecasts.
- Risk: fabrication cost, yield, customer concentration and geopolitics.
Kioxia plans annual capital expenditure of about ¥470 billion and R&D of ¥230 billion over three years. Cash from the upcycle must fund the next technology, but missed demand leaves heavy fixed costs. Multi-year agreements improve visibility; without disclosed price and volume terms, outsiders cannot fully measure stability.
The NAND of 1987 looked like a way to miniaturize magnetic storage. BiCS in 2007 was a manufacturing revolution that stacked cells upward. Generation 10 connects both ideas to the inference memory hierarchy. Kioxia’s shares and earnings may fall again—memory remains cyclical. But every AI answer still requires data to be kept and delivered again. The “memory” Japan once sold has returned as a strategic asset beneath the world’s computation.