AI Demand for NAND Surges as Focus Shifts from 'Training' to 'Inference'

Samsung and SK hynix Go All-In on Next-Generation Process Transition

SK hynix Achieves Over 10 Trillion Won in Quarterly NAND Sales for the First Time

Emerging as the New Profit Driver After DRAM

NAND flash, which had long been mired in deficits and was once dubbed the "troublemaker" of the semiconductor industry, is now poised for a dramatic comeback. As it establishes itself as a core component in artificial intelligence (AI) "inference," it is emerging as a major profit driver rivaling DRAM. In particular, with both Samsung Electronics and SK hynix accelerating their transition to next-generation NAND processes, the competition for leadership in the memory market is rapidly shifting from DRAM to NAND.


Samsung: 'Low Latency, Ultra-High Capacity' vs. SK hynix: 'Solidigm Synergy'

According to industry sources on May 2, Samsung Electronics has completed preparations for mass production of its next-generation NAND solution, the PCIe 6th generation (Gen6) solid-state drive (SSD), and has received positive feedback from major clients during initial sample evaluations. PCIe Gen6 offers roughly double the data transfer speed of its predecessor, making it an optimized interface for AI inference servers that must process massive data in real time. Samsung Electronics plans to leverage this to secure demand for next-generation AI servers and data centers, which is expected to surge in the second half of the year.

Samsung Electronics' 9th generation Quad-Level Cell (QLC) V-NAND. Samsung Electronics

Samsung Electronics' 9th generation Quad-Level Cell (QLC) V-NAND. Samsung Electronics

View original image

The company is also accelerating the transition to its 9th generation (V9) quad-level cell (QLC) technology. With the completion of its "1-terabit (Tb) QLC" in March, Samsung plans to expand its ultra-high-capacity lineup, including "256TB server SSDs." This strategy aims to minimize data transfer delays and maximize storage utilization for AI workloads that require frequent data access.


SK hynix is also pushing ahead with its transition to cutting-edge NAND processes to reshape itself as a provider of "AI-tailored solutions." This month, Solidigm, SK hynix’s NAND subsidiary, established an "AI SSD Platform Development Center" in Vancouver, Canada. This move is aimed at strengthening research and development (R&D) capabilities and converting legacy processes to the most advanced technologies. SK hynix reportedly plans to transition 50% of its domestic NAND production to the latest 321-layer (9th generation) process within the year. The 321-layer NAND, currently the world's highest stack, is considered a key milestone for maximizing data density and enhancing cost competitiveness.

'321-layer 2Tb QLC NAND' developed by SK hynix. Photo by SK hynix

'321-layer 2Tb QLC NAND' developed by SK hynix. Photo by SK hynix

View original image

In particular, SK hynix aims to build a full lineup for AI data centers through a two-track strategy: combining the parent company's "high-performance, high-speed products" with Solidigm's "high-capacity QLC-based enterprise SSDs (eSSD)." The company has already completed customer certification for the world’s first QLC product based on 321-layer NAND, laying the groundwork for market leadership.


AI 'Inference' Drives 7-Fold Surge in NAND Prices

This NAND rebound is closely tied to a paradigm shift in the AI industry, as the focus moves from "training" data to "inference" using fully developed models. Major AI companies such as NVIDIA are proposing architectures that store data generated during inference not only on high-bandwidth memory (HBM) but also on NAND-based storage, increasing the importance of high-performance storage solutions.


In particular, "KV cache" technology—temporarily storing and rapidly accessing massive intermediate data during inference—has become a key focus area. To handle this, high-performance enterprise SSDs, which offer superior power efficiency and data processing speeds, are now essential components, fueling a sharp rise in NAND prices.


Actual price indicators are showing a steep upward curve. According to market research firm TrendForce, contract prices for enterprise SSDs are expected to rise by 48–53% quarter-on-quarter in Q2 this year. This follows a nearly 50% surge in Q1, marking two consecutive quarters of rapid increases. The prices of some standard products have soared nearly sevenfold in just one year. Additionally, major manufacturers allocating DRAM production lines to HBM output has led to a "balloon effect," further reducing NAND production capacity and intensifying supply shortages.


Market experts predict that the current tight supply-demand situation will persist at least through 2027. While Micron has announced a $24 billion investment in a NAND fab in Singapore and SK hynix is expanding its Dalian Plant 2, there is a significant lead time before these facilities reach mass production. Surim Lee, a researcher at DS Investment & Securities, explained, "Chinese companies remain focused mainly on domestic and general-purpose products, so their impact on global data center supply and demand is expected to be limited."


The earnings indicators of major memory manufacturers also reflect NAND’s growth. In the first quarter, SK hynix’s NAND sales reached 11.041 trillion won, surpassing the 10 trillion won mark for the first time. This represents a 248% surge from a year earlier. The securities industry estimates that Samsung Electronics’ NAND operating margin for Q1 is around 53%, and forecasts that NAND will become the new "cash cow" of the memory semiconductor industry.



An industry insider commented, "While the market is still centered on DRAM, such as HBM, the profitability of NAND has improved to an incomparable degree. The competition among companies is no longer just about increasing volume, but about who can deliver more AI-optimized, tailor-made solutions—a true 'transformation' battle."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing