The storage in your phone, the SSD in your PC, the data centers training AI, they all run on something most people have never heard of: 3D NAND flash memory. Now Japan's Kioxia and US-based Sandisk just unlocked the door to "1,000 layers," a milestone the entire memory industry has been chasing for a decade. On June 14, 2026, the world will see exactly how at the VLSI Symposium in Hawaii. Here's why it matters far beyond the spec sheet.

The 1,000-Layer Wall: Why Memory Has to Grow Skyward

NAND flash memory is the chip that stores your data even when the power is off, from the photos on your smartphone to the SSD in your laptop to the petabyte-scale storage racks inside AI data centers.

For roughly a decade, the industry has expanded NAND capacity by going vertical. Instead of laying memory cells side by side on a silicon wafer like a one-story shopping mall, manufacturers stack them upward like a skyscraper. The footprint stays the same, but the number of cells per chip multiplies.

Samsung kicked this off in 2013 with a 24-layer 3D NAND chip. By 2024, South Korea's SK hynix was mass-producing 321-layer V9 NAND. Samsung is shipping 290-layer V-NAND and chasing 400-layer density. Micron has shipped 276-layer NAND in consumer SSDs and is preparing to leap straight to 400-layer Gen7 chips.

The next milestone is "1,000 layers." SK hynix has publicly committed to passing that threshold by 2030. But beyond about 300 layers, simply adding more floors to the tower stops working. The physics fights back.

The deeper a memory cell's vertical hole gets, the harder it is to etch it cleanly. At the 218-layer node, today's most advanced cryogenic etching tools take roughly an hour to drill the holes through one wafer. Stack any higher and the wafer itself starts to warp, throwing the upper-layer holes out of alignment with the lower layers and crashing the yield rate. New tricks are needed.

Kioxia and Sandisk think they've found one.

What "World First" Actually Means: Enter MSA-CBA

On April 30, 2026, Kioxia and Sandisk announced that they had achieved what both companies call a "world first", a working quad-level cell (QLC) NAND demonstration built on a new structure they're calling MSA-CBA, short for Multi-Stacked Cell Array CMOS, joined together by wafer-to-wafer copper direct bonding. The full technical paper was unveiled at the VLSI Symposium 2026 in Hawaii from June 14 to 18.

To understand why this matters, you need to know the building block underneath: CBA (CMOS directly Bonded to Array). Every NAND chip has two functional layers, the memory cells that store data, and the CMOS logic circuits that control read/write operations. Traditionally these are built on a single wafer. CBA splits them onto two separate wafers, optimizes each for its own job, then bonds them together using microscopic copper pads. Kioxia became the first manufacturer to mass-produce this architecture in its 218-layer 8th-generation BiCS FLASH (BiCS8), and is now ramping up 332-layer BiCS10 production, pulled forward to 2026 from an original 2027 schedule.

MSA-CBA goes one step further. Instead of just bonding one CMOS wafer to one cell array wafer, the new architecture bonds multiple cell array wafers on top of each other. The recent demonstration stacked two 218-layer cell array wafers on top of a CMOS wafer, effectively yielding a 436-layer-equivalent device, and showed that QLC operation works across the bonded stack. Repeat the process with a few more cell wafers and the path to 1,000 layers opens up.

According to the two companies, MSA-CBA solves three of the most stubborn problems in extreme high-layer NAND: degradation of cell current as layers increase, wafer warpage during processing, and the ballooning size of memory blocks (the smallest erasable unit) that hurts random write performance. The companies position the demonstration as "an important milestone toward ultra-high-density 3D flash memory with stack structures of 1,000 layers or more."

The paper title is technical but worth quoting: "A Multi-Stacked Cell Array Architecture with Wafer-to-Wafer Cu Direct Bonding for Ultra-High-Density 3D Flash Memory beyond 1000 Word Lines." The same VLSI Symposium will also feature presentations from SAIMEMORY (a Japanese memory research consortium) and Intel on a different but related concept: a 9-wafer ultra-thin (3 μm-Si per stack) high-bandwidth 3D memory using a "Via-in-one" fusion bonding architecture.

The Three-Way Stacking Race

The global NAND market is essentially a four-player oligopoly: Samsung (Korea), SK hynix (Korea), Micron (US), and the Kioxia-Sandisk alliance (Japan-US). Q3 2025 share rankings put Samsung first at roughly 32%, SK hynix at 19%, Kioxia at 15%, and Micron/Sandisk together at around 12%.

Each player has a slightly different roadmap into the post-300-layer era.

SK hynix is the most aggressive. It has been mass-producing 321-layer QLC NAND since August 2025 and plans full-scale 400-layer NAND production in the first half of 2026. The company has publicly stated a goal of crossing 1,000 layers by 2030 using hybrid bonding (the technical family that includes wafer-to-wafer Cu bonding).

Samsung has been mass-producing its 290-layer TLC NAND (V9) since April 2024, and reportedly reached 400-layer density via a triple-stack architecture in late 2025. Samsung is splitting investment between NAND scaling and high-bandwidth memory (HBM), but it has not slowed its 3D NAND ambitions.

Micron is taking the most contrarian path. Rather than incrementally adding layers, the company plans to leap from 276-layer Gen6 directly to 400-layer Gen7, skipping the 300-layer node entirely. Micron has licensed hybrid bonding IP from Adeia and is expected to be the slowest of the three to deploy it in production NAND.

The Kioxia-Sandisk alliance, meanwhile, is leaning hard on its CBA head start. Kioxia has guided to capital expenditure of around 450 billion yen for fiscal 2026, prioritizing faster development of the tenth-generation BiCS10 and a very high IOPS SSD aimed at NVIDIA's next storage architecture. With MSA-CBA they are now extending that head start into the next decade.

One structural advantage worth highlighting: Kioxia and Sandisk are NAND-only. Samsung, SK hynix, and Micron all also make DRAM, and right now they are pouring most of their capital into HBM, the high-margin DRAM stacked on AI accelerators like NVIDIA's. Pure-play NAND makers are far more exposed to the AI-driven NAND price surge, and if MSA-CBA actually translates into denser, cheaper bits, the underdogs could rewrite the rankings.

How AI Data Centers Rewrote the Priority List

Why is everyone suddenly racing so hard to add layers? The answer is AI.

Generative AI services like ChatGPT, Gemini, and Claude have multiplied the volume of data flowing through data centers, model weights, training corpora, inference caches, user histories, generated images and videos. Even the most powerful NVIDIA or AMD GPU is bottlenecked if the storage feeding it can't keep up.

Research firm TrendForce projects that NAND supply will fall short of demand throughout 2026. SK hynix and Micron have already cut output by roughly 10% to support prices, and Samsung is reportedly weighing a 20-30% price hike in 2026. A 32GB DDR5 memory kit that sold for around 28,000 yen ($178) in late 2025 has nearly doubled to over 50,000 yen ($318) in early 2026. Smartphone and PC price increases are now seen as nearly inevitable.

Kioxia's actual fiscal 2025 results, reported on May 15, 2026, came in well above its own forecast. Revenue rose 37.0% year on year to 2.3376 trillion yen, clearing 2 trillion for the first time, and non-GAAP operating profit nearly doubled to 876.2 billion yen, a 37.5% margin. That operating profit is the highest in eight years, beating even the Toshiba Memory era. The fourth quarter alone produced more operating profit than the whole of the previous year. Guidance for the first quarter of fiscal 2026 is 1.75 trillion yen in revenue and 1.3 trillion yen in non-GAAP operating profit.

Co-development with NVIDIA is also accelerating. Kioxia is building a "Super High IOPS SSD" that targets 100 million IOPS, roughly 100x today's enterprise SSDs, designed to hand data directly to GPUs, bypassing the conventional CPU bottleneck. Evaluation samples are scheduled for late 2026. In this emerging architecture, ultra-high-density MSA-CBA NAND becomes the bottom layer of an AI storage hierarchy: HBM for short-term memory, ultra-fast SSD for working memory, and dense 1,000-layer NAND SSD for long-term memory.

Japan's Quiet Memory Comeback

It would be easy to read this story as just another corporate technology announcement. It isn't.

NAND flash memory was invented at Toshiba in 1987 by engineer Fujio Masuoka, the company that, after multiple corporate restructurings, became Kioxia. Japan invented this product category and then watched Samsung outspend it on production capacity, lose share, and slowly retreat from headline leadership for nearly two decades.

Kioxia spun out from Toshiba in 2018 and rebranded in 2019. After repeated IPO delays, it finally listed on the Tokyo Stock Exchange Prime Market in December 2024. The stock then rose 540% in its first year, the best-performing major stock globally in 2025, and the climb continued: market capitalization reached the 36 trillion yen range by July 2026. Apple and Microsoft are among its biggest customers.

The MSA-CBA announcement layers a technology story on top of that financial story. Combined with Japan's continued dominance in semiconductor production equipment (Tokyo Electron, SCREEN, Lasertec), NAND is one of a shrinking number of semiconductor segments where Japan still leads at the cutting edge.

The caveats matter, of course. SK hynix has its own published path to 1,000 layers. Samsung has more capital. China's YMTC has its own bonded-wafer "Xtacking" technology and is years ahead in some operational metrics, even if export controls complicate its global reach. A "world first" at a conference and "world first to mass production" are very different things.

But measured from the day Toshiba Memory was carved out as a separate company in 2018, the speed of Japan's memory industry rebound has been striking, and quiet enough that most outside observers haven't yet caught up to it.

When Will We Actually Use This?

Kioxia and Sandisk haven't published a production roadmap for MSA-CBA. Based on the typical interval between an announcement at VLSI Symposium and a shipping product, the first commercial chips using this architecture would likely arrive in 2027 or 2028, with serious 1,000-layer-class NAND reaching the market around 2029 or 2030.

By that point, our digital lives will look different. Smartphone storage may become 2TB by default. A consumer-grade PC could ship with 32TB of internal SSD. AI agents that learn from a person's entire digital history, emails, calendar, documents, conversations, would actually have somewhere to store all that context.

Inside data centers, the storage hierarchy will reorganize around three tiers: HBM as short-term memory, ultra-fast SSD as working memory, and ultra-dense NAND SSD as long-term memory. If MSA-CBA delivers on its promise, AI's long-term memory may run on chips designed in Japan.

In Japan, the phrase "semiconductor renaissance" gets thrown around a lot. But this isn't nostalgia, there's genuine technical progress underneath the headlines, and AI's hunger for memory is what's driving it.

How is the semiconductor memory race being covered in your country? Do you notice the storage capacity of your phone or laptop creeping upward year after year? We'd love to hear what this technology shift looks like from where you live.

References