DigiTimes 综述：AI 内存需求重塑台湾存储业，南亚科/华邦电转战略资产（无新增催化）

The AI wave is triggering a surge in memory demand that extends from cloud data centers to edge applications. This surge has not only delivered a substantial dividend to relevant supply chains but is also driving a fundamental structural shift across the broader memory ecosystem.

Memory is evolving from a cyclical, highly standardized commodity into a strategic asset for the AI era. Having survived multiple severe downturns, Taiwan's memory industry is moving beyond its role as a mere beneficiary of redirected AI orders to become an indispensable pillar of the global AI supply chain.
Capacity crunch redefines mature nodes
Memory was historically viewed as a standardized bulk commodity whose pricing closely tracked the boom-and-bust cycles of consumer electronics. However, the AI boom has created intense capacity crowding-out effects, triggering unprecedented supply shortages and price hikes across the board. Global integrated device manufacturers have responded by phasing out legacy production lines to reallocate capacity toward advanced processes and high-end products.

Credit: Jerry Chen
Yet mature memory technologies remain far from obsolete. Beyond their pivotal roles in industrial control, networking, automotive, and edge AI sectors, niche memory types such as DDR4, NOR Flash, and SLC or MLC NAND are also essential components in enterprise SSDs and server systems, serving as vital building blocks for AI infrastructure. Consequently, Taiwanese memory makers are shedding their peripheral status to become critical anchors for global supply stability and system performance, fundamentally redefining their core value proposition.
Hard lessons from a near wipeout
Taiwan's memory journey has been marked by painful historical lessons. The island's DRAM manufacturing sector took root in the early-to-mid 1990s, initially relying heavily on technology licensing and capacity reservation agreements. By the early 2000s, Taiwan had captured roughly 20% of the global DRAM market, ranking second worldwide behind South Korea. Yet despite building a respectable manufacturing scale, the local industry remained deeply dependent on foreign patents and core technology transfers.
That dependence left the sector exposed when the 2008 global financial crisis struck. South Korea's Samsung Electronics executed an aggressive countercyclical expansion strategy, slashing prices and flooding the market at the depths of the downturn, severely crushing competitors. Global DRAM giant Elpida was driven into bankruptcy, while Taiwan's DRAM sector was nearly wiped out.
Powerchip was delisted and forced to pivot its business model, ProMOS Technologies underwent corporate restructuring and exited the market, and Inotera was eventually acquired by Micron Technology. The sector was widely written off at the time as a complete disaster area.
A pivot to niches and controllers
Although the global DRAM market consolidated into an oligopoly dominated by Samsung, SK Hynix, and Micron, end-market demand cycles continued to drive steep margin volatility. To avoid competing head-on with the top three giants in commodity DRAM, Taiwanese manufacturers strategically pivoted toward niche markets.
Nanya Technology incrementally built its proprietary 10nm-class manufacturing process, progressing from DDR3 and DDR4. Winbond Electronics carved out deep positions in NOR Flash, SLC NAND, and niche DRAM. Macronix focused on high-reliability NOR Flash, SLC or MLC NAND, and ROM, becoming a core vendor for Japanese gaming giant Nintendo as well as leading automotive tier-1 suppliers.
While Taiwan did not establish a domestic NAND Flash wafer manufacturing ecosystem, fabless vendors such as Silicon Motion and Phison Electronics emerged as dominant global leaders in NAND Flash controller ICs. Meanwhile, memory module makers including Transcend, ADATA, Apacer, Innodisk, and Team Group anchored themselves in consumer and industrial markets, consistently ranking among the top 10 memory module suppliers globally.
Nanya's long road to tech autonomy
To ensure it would never again be caught without foundational R&D capabilities, Nanya Technology embarked on a nearly 30-year journey toward technology self-reliance. Beginning with joint R&D alongside IBM in 1998, Nanya later partnered with Germany's Infineon and its spun-off memory unit Qimonda in 2002, and worked with Micron from 2008 until terminating joint development in 2012.
By early 2020, Nanya had successfully rolled out its fully in-house 10nm-class DRAM technology node. Today, its 1b process has advanced to mass production of 16Gb DDR5. Customer validation for the 1c process is scheduled for the first quarter of 2027, while development of the 1d and 1e processes is underway alongside plans to introduce EUV lithography equipment.
In a landmark strategic move, Nanya Technology recently closed an NT$78.72 billion private placement. Beyond funding new fab construction and capacity expansion, the round brought in four heavyweights as strategic shareholders: SanDisk, Kioxia, SK Hynix's Solidigm, and Cisco. By cementing deeper ties with key global clients through equity participation, Nanya is locking in long-term supply agreements, further proving that memory is no longer just a commodity component but a strategic asset requiring guaranteed supply.
Winbond's CUBE and silicon capacitor push
Winbond President Chen Pei-ming noted that Taiwan's traditional vulnerability lay in producing standardized commodities that remained exposed to macro cyclical swings, even within niche markets. Winbond's strategic pivot focuses on stabilizing supply-demand dynamics through structural product mix optimization to build a defensive business moat, representing precisely this shift toward a cycle-resilient business model.
Winbond's growth engine centers on its 16nm process node and its proprietary CUBE customized high-bandwidth memory architecture. Mass production yields for its 16nm process have hit 80%, with a target of nearly 90% by late 2026. Monthly capacity at Module A of its Kaohsiung fab will reach 24,000 wafers by year-end, while Module B is designed to support next-generation 14nm and 12nm DRAM nodes with EUV capability.
Alongside its CUBE platform, Winbond has successfully expanded into silicon capacitors for advanced AI packaging, operating on a foundry model to supply basic capacitor units and process design kits. Mass-produced units currently achieve a capacitance density of 3,300 nF per square millimeter, while the latest samples reach 5,400 nF per square millimeter. Backed by a dedicated NT$4 billion expansion investment, Winbond's specialized capacitor line has rapidly reached full capacity.
Weighing CXMT's 4F² challenge
Industry attention recently turned to Chinese DRAM maker ChangXin Memory Technologies (CXMT), which announced plans to launch server DDR5 RDIMMs using a 4F² cell architecture by late 2026. Compared with the standard 6F² architecture used by global DRAM leaders, a 4F² design shrinks the memory cell area by roughly one-third at equivalent process feature sizes.
However, industry analysts point out that CXMT's 1z process parameters remain relatively relaxed, making 4F² implementation less technically complex. Applying 4F² to the leading trio's ultra-dense fifth- or sixth-generation 10nm-class nodes raises serious reliability and yield concerns, which makes direct performance comparisons misleading. Whether CXMT's 4F² architecture can achieve stable, yield-viable mass production remains to be proven.
Custom memory for edge AI
While mainland Chinese producers are aggressively targeting advanced memory nodes, their immediate competitive threat to Taiwanese firms remains manageable. Operating pragmatically within capital and resource constraints, Taiwanese suppliers are focusing heavily on HBM-like applications that prioritize high bandwidth, low latency, and low power consumption, delivering customized, high-value-added solutions for real-time edge AI and inference workloads.
Winbond's CUBE platform is initially targeting edge AI devices and is projected to account for a double-digit share of the company's total DRAM revenue by 2027. AP Memory's VHM solution utilizes true 3D vertical stacking to integrate DRAM directly over the logic die, with customer tape-outs for AI and HPC projects slated for between late 2027 and 2028.
Meanwhile, Nanya Technology is building out its customized Ultra-Wide I/O memory platform using 3D IC technologies such as through-silicon vias and wafer-on-wafer stacking. To reinforce this strategy, Nanya acquired a strategic stake in PieceMakers Technology and announced plans for a new advanced facility in Pingtung.
Taiwan's memory sector is successfully transcending the narrow margins of commodity chip markets to establish a highly differentiated value proposition. By executing a dual-track strategy pairing long-term supply stability agreements with bespoke memory architectures, Taiwanese players are connecting directly with global AI chip giants and tier-1 system architects, securing a complementary, highly strategic stronghold within the global AI memory ecosystem.

Credit: Jerry Chen