The Super Cycle of Memory and Manufacturer Competition
[Relevant Companies Mentioned in This Article] SK Hynix, YMTC, Micron, Samsung, Longsys, Kioxia, Winbond, Macronix, GigaDevice, Puya, Zbit Semiconductor, Nanya Technology.
"Outdated" Process Memory Actually Takes the "Lead"
First, it is essential to clarify a point. Before discussing the landscape of Niche Memory, it is necessary to re-examine its technical positioning: NOR Flash, SLC/MLC, 2D NAND, and niche DRAM have been the market mainstream over the past decade, with capacity concentrated in mature processes of tens of nanometers. The industry's marketing and promotional focus has been entirely on advanced processes like 7nm and 3nm, and mature processes have always been labeled as "outdated processes." However, this definition has been failing in 2026. As Samsung, SK Hynix, and Micron channel over 90% of their capital expenditures into HBM3e/4 and 3D NAND with over 300 layers, niche products remaining on 2D architectures or mature nodes have instead built a unique "process barrier."
Given the shortage of HBM, DDR5, and high-capacity, high-speed NAND driven by AI industry demand, major manufacturers are also adjusting their capacity towards the more profitable and technologically advanced HBM sector. This "mass retreat" from mid-to-low-end capacity has created a new market structure for manufacturers in Mainland China and Taiwan. We can divide global memory original manufacturers into two camps: on one side, the "Big Three" giants betting their future on AI, and on the other, manufacturers from Chinese mainland and Taiwan are unlocking new growth opportunities within mature process nodes.
YOLO Memory Industry Forecast (Image source: YOLO)
Taking SLC NAND products as an example, the capacity of a single chip can no longer keep up with the times. The market demands high-capacity and high-bandwidth products, and overseas original manufacturers like Kioxia and Micron have evolved their process nodes to 14-16nm, which can be said to be a generational lead. However, bringing us to 2026, the demand for HBM forces manufacturers to rapidly switch more capacity to the HBM industry, pushing this industry-leading position aside due to capital returns and demand. In the wafer market, SK Hynix spun off its SLC business to SkyHigh, and subsequently, Puya Semiconductor acquired SkyHigh. On the surface, this looks like a technological reinforcement for Mainland manufacturers, but in reality, SkyHigh's capacity is far from what it used to be, dropping from the previous 9kpcs to the current 2.2kpcs, a massive reduction of nearly 70%. In contrast, manufacturers in Mainland China and Taiwan originally had mainstream processes of 12-24nm. Foundry and original manufacturers like Winbond and GigaDevice are mainly at SMIC and Hua Hong Semiconductor. Although they cannot enjoy the "huge profits" from this HBM wave, the vacated capacity gap ensures the continuity and controllability of their capacity. In other words, in the current era where capacity is king, cooking without rice at 14nm is far inferior to the stable output at 24nm.
The vast market for niche NAND storage requires a stable and controllable supply chain. The withdrawal of overseas original manufacturers like Samsung and SK Hynix is clearly not about reducing capacity, but rather the physical upgrading of equipment to switch to the HBM track. Samsung announced that it will complete the EOL (End of Life) of all 2D MLC NAND by June 2026, reducing monthly capacity by nearly 200kpcs, creating a huge gap in the market. Furthermore, Samsung's outdated equipment needs to be cleared out, and news reports indicate that the vast majority will flow into the Mainland market as second-hand equipment. Originally, Samsung monopolized 45% of the capacity and 55% of the revenue in the MLC market, and almost all well-known large Fabless manufacturers sourced their capacity from Samsung. This wave of EOL not only creates a direct vacuum in a $3 billion market but also completely cuts off the wafer supply for large, medium, and small module manufacturers. Switching suppliers in the short term is a significant challenge for any electronics factory. In Taiwan, Macronix currently has IDM capabilities, while the wafers from XMC are hardly sold externally, leaving module manufacturers in a situation where "a single chip is hard to find." Wafer manufacturers such as Hua Hong, SMIC, and PSMC are accelerating the introduction of MLC. However, MLC's control requirements for noise and charge loss are far higher than those of SLC. It is impossible to get through from trial production to yield ramp-up in less than a year and a half.
SLC NAND Manufacturer Product Comparison (Image source: Minsheng Securities)
Similarly, in the NOR Flash field, the market technology stratification is highly fragmented. Winbond and GigaDevice, in the high-capacity market above 256Mb, currently focus their business mainly on edge AI, AI server control, and automotive ADAS fields; the small and medium storage field of 64-128Mb is in the hands of manufacturers like Puya and Zbit Semiconductor. Due to the different cost boundaries of ETOX architecture and SONOS/floating gate architecture at different capacities for NOR Flash products, the main technical paths pursued by each manufacturer also vary. High-capacity NOR Flash requires complex circuit design to ensure read speeds. Although advanced processes can temporarily solve the problem of insufficient chip area, costs will also rise significantly, and whether end customers will accept it is a considerable issue. Of course, it is undeniable that the current structural price increase due to tight high-capacity capacity is an industry-wide general price increase. The seemingly "outdated" process has instead captured a wave of the market, leading other manufacturers in the industry.
Competition Between Mainland China and Taiwan Manufacturers
The strategies of Mainland China and the Taiwan region in niche memory are now showing a clear trend of divergence: Taiwanese manufacturers entered the industry earlier and have continued to use the early IDM model. The subtle improvements in processes and continuous yield enhancements have indeed earned them their first pot of gold. Moreover, relying on wafer foundry services to continuously improve process capabilities, TSMC has now become the industry hegemon. Mainland manufacturers entered the game later and are more adept at enhancing their technological strength through capital operations, acquisitions, and mergers. At the same time, the strong support from government and industrial funds has created industry giants such as BOE and CXMT.
Drilling down to the segmented market, EEPROM originally belongs to an extremely small field. Currently, the Mainland manufacturer Giantec Semiconductor has become the top three globally, while the market shares of industry giants like STMicroelectronics and onsemi are continuously being eroded. In contrast, Taiwanese manufacturers have currently dropped out. This is because EEPROM has extremely low process requirements but an extremely complex product line, requiring a large amount of manpower to follow up on customer needs and continuously adjust products and production lines to meet terminal demands, which is precisely the strength of Mainland manufacturers.
Returning to the aforementioned niche NAND and DRAM markets, in the NAND market, giants like Samsung, SK Hynix, and Kioxia have made up their minds to clear out "outdated" capacity to focus on HBM. Equipment and production lines are being dismantled one after another, leaving a huge vacuum in the market. Whichever company holds on at this stage and seizes the first wave will have the opportunity to gain a firm foothold. On the DRAM side, industry giants are currently reducing manpower and lowering original capacity to ensure the supply of products like DDR5 and LPDDR5, while nearly 50% of the remaining capacity is still cut. This presents a significant opportunity for manufacturers such as CXMT, GigaDevice, and Ingenic Semiconductor (which acquired ISSI). Mature products such as consumer-grade and automotive-grade DDR3/4 have seen a sharp increase in terminal product prices due to capacity cuts. Along with the successful IPO of CXMT, focusing on the consumer electronics and automotive-grade markets, it is believed that they will be able to outcompete Taiwanese manufacturers once again.
Conclusion
Ultimately, this so-called "niche memory super cycle" is essentially the supply and demand tension in recent years caused by the siphon effect of AI on semiconductor resources. What Mainland and Taiwan manufacturers have caught is not only market share but also the capacity and market that original manufacturers consider "outdated." However, once supply and demand ease and market prices stabilize, whether the new expanded capacity of domestic manufacturers can meet the high-threshold demands of the automotive and industrial levels will be the top priority in determining the reshaping of this round of landscape.
The illustrations in the text are generated by generative AI
References:
In-depth Industry Research "Reshaping the Niche Memory Market Landscape Driven by the AI Super Cycle: Supply and Demand Changes and Opportunities and Challenges for Taiwan and Mainland Manufacturers in Various Sub-sectors Including NOR Flash, SLC, MLC, and DRAM"
Official websites of major companies