DRAM, the chip that serves as your device's short-term memory, is hitting the physical limits of its flat design. The way out is to stack memory cells vertically. And an unexpected challenger has entered the ring: Kioxia, a company known only for flash storage, is now gunning for the DRAM market.

Why DRAM Is Going 3D: The Physics Problem Every Device Faces

DRAM is the high-speed workbench where a processor parks the data it's actively using. Its speed and capacity set the ceiling on system performance, which is why every smartphone, PC, and AI data center depends on it.

For decades, DRAM makers shrank transistors on flat, two-dimensional chips. The leading-edge process has now reached around 12 nanometers and the physical limits are closing in. As transistors get smaller, interference between neighboring cells rises and the tiny charges holding the data leak away faster.

Enter 3D DRAM: stack the memory cells vertically and capacity grows without the chip footprint growing with it. NAND flash proved the approach works, with manufacturers now stacking over 300 layers in volume production. The DRAM industry is racing to apply the same principle.

Kioxia's Bold Move: Using a TV Screen Material to Build Next-Gen DRAM

The most surprising entrant in this race is Kioxia (formerly Toshiba Memory), a pure-play NAND flash manufacturer with no DRAM products at all.

On December 10, 2025, at IEDM in San Francisco, Kioxia presented new advances in its "OCTRAM" (Oxide-Semiconductor Channel Transistor DRAM) technology, which uses IGZO (Indium Gallium Zinc Oxide) as the transistor channel material. OCTRAM itself debuted a year earlier at IEDM 2024, developed jointly with Taiwan's Nanya Technology; this year's work takes it into three dimensions.

If IGZO sounds familiar, it's because the same material sits behind LCD and OLED display panels. So why would a display material work for memory chips?

The key is IGZO's extremely low leakage current. DRAM stores data as electrical charges that gradually leak away, requiring constant "refresh" cycles, and the bigger the array, the heavier that refresh power gets. IGZO transistors leak orders of magnitude less than conventional silicon, which cuts that overhead directly.

In this latest demonstration, Kioxia formed lateral IGZO transistors eight layers at a time by replacing silicon nitride regions with oxide semiconductor, and confirmed they worked. It reported an on-current above 30μA and an off-current below 1 attoampere (10^-18 amperes).

The method matters as much as the numbers: the alternating oxide/nitride stack it relies on is the same one Kioxia has spent years perfecting in 3D NAND, where the company routinely builds hundreds of stacked layers. It has its sights set on commercialization in the 2030s.

The Competitive Landscape: A Three-Way Race Across East Asia

Samsung (South Korea): First to Publish a 3D DRAM Roadmap

Samsung Electronics, the world's largest memory maker, was the first major company to publicly outline a 3D DRAM development roadmap, presenting it at Memcon 2024. The Korean giant plans to introduce an early version based on vertical channel transistor technology and deliver a fully stacked DRAM by 2030.

According to Korean outlet KED Global, Samsung projects that 3D DRAM will achieve a base capacity of 100GB per chip, roughly three times the current DRAM maximum of about 36GB. The company has also established a dedicated 3D DRAM R&D lab in Silicon Valley, signaling serious commitment.

SK hynix (South Korea): The AI Memory King, Eyeing 3D DRAM

SK hynix dominates the High Bandwidth Memory (HBM) market. Counterpoint Research put its Q3 2025 revenue share at 57%, ahead of Samsung (22%) and Micron (21%), and it remains NVIDIA's primary supplier. That lead carried into the broader market: SK hynix passed Samsung in DRAM revenue for the first time in Q1 2025 and held the top spot for three straight quarters before Samsung reclaimed it in Q4. For full-year 2025, SK hynix posted a record operating profit of 47.2 trillion won against Samsung's 43.6 trillion, beating its rival on profit for the first time.

HBM stacks multiple DRAM chips vertically with ultra-fast connections, and it's essential for training and running large AI models. SK hynix announced it had completed HBM4 development in September 2025 and began volume supply in the first half of 2026, aimed at NVIDIA's next-generation "Rubin" platform.

On the 3D DRAM front, SK hynix has disclosed that it's exploring IGZO as a channel material for future memory, putting it on a similar trajectory to Kioxia.

CXMT (China): The Sanctions-Defying Rising Star

China's largest DRAM manufacturer, CXMT (ChangXin Memory Technologies), keeps gaining ground despite U.S. export restrictions. Counterpoint Research puts its DRAM revenue share at 5% in Q3 2025, up from 3% a year earlier, and it has begun mass-producing DDR5 and LPDDR5X chips.

In the 3D DRAM arena, CXMT also presented IGZO-based transistor technology at IEDM 2025. Their approach uses back-end-of-line (BEOL) processes to form IGZO transistors, achieving an impressive on-current of 60.9μA/μm. Additionally, CXMT showcased a 3D FeRAM (ferroelectric memory) technology derived from 3D DRAM structures, targeting a "near-nonvolatile" memory that could blur the line between DRAM and storage.

While U.S. semiconductor equipment export controls limit CXMT's access to cutting-edge manufacturing tools, the company's ability to innovate within these constraints has alarmed Korean competitors.

YMTC (China): Closing the Gap in NAND Flash

On the NAND flash side, YMTC (Yangtze Memory Technologies Corporation) has emerged with advanced 3D NAND chips exceeding 232 layers, approaching the technology levels of Samsung and Micron. While not a direct 3D DRAM player, YMTC's progress illustrates the broader strengthening of China's semiconductor memory capabilities.

The AI Demand Explosion: Why Memory Is the Next Bottleneck

The urgency behind 3D DRAM development stems from a single word: AI.

Training and running large language models (LLMs) like ChatGPT requires enormous memory capacity and ultra-fast data transfer speeds. NVIDIA's latest Blackwell GPUs consume massive amounts of HBM, and demand for AI-grade memory has surged to unprecedented levels.

The bill has landed on conventional memory. TrendForce reports that conventional DRAM contract prices rose 90–95% quarter over quarter in Q1 2026 and forecasts a further 58–63% jump in Q2, as Samsung, SK hynix, and Micron reallocate advanced nodes and new capacity toward server and HBM products. PC and smartphone makers are the ones squeezed out, and meaningful new fab capacity isn't expected before late 2027.

But even HBM will eventually hit its own capacity and bandwidth limits. That's where 3D DRAM comes in: lift the capacity per unit area far enough and the next bottleneck moves somewhere else.

Can Japan's Memory Industry Make a Comeback?

In the 1980s, Japanese companies held roughly 80% of the global DRAM market. Trade friction with the U.S., the rise of Korean competitors, and slow management decisions ended that. In Q3 2025, Counterpoint put DRAM revenue share at SK hynix 34%, Samsung 33%, and Micron 26%: over 90% between the three of them, with no Japanese player on the board.

Kioxia's 3D DRAM development challenges this status quo. A NAND-only company entering the DRAM market is unprecedented, but the paradigm shift to 3D DRAM represents a once-in-decades window of opportunity for new entrants.

It isn't the only Japanese attempt. SAIMEMORY, a SoftBank subsidiary founded in December 2024, is working with Intel on ZAM (Z-Angle Memory), a stacked DRAM architecture aiming to cut power consumption by up to half versus HBM. The plan calls for roughly 8 billion yen through fiscal 2027 to build a prototype, with practical deployment targeted for fiscal 2029.

The challenges are real. Kioxia's technology is still in the research phase, and mass production means clearing reliability, cost, and yield hurdles. Intel and Micron's Optane was technically impressive and still failed to find a market.

What the eight-layer result does mark is the shift from proposing a structure to demonstrating a stack. Using IGZO, a material with deep roots in Japanese display technology, to flank the AI memory market is one of the few plays Japan's semiconductor industry has left on this board.

What's happening in the memory industry in your country? Are there efforts to build homegrown memory technology or reclaim market share?

References