China's rapidly expanding memory manufacturer ChangXin Memory Technologies (CXMT) could significantly increase global DRAM supply by 2028, potentially disrupting the market and putting pressure on memory prices. The warning comes from Charles Kau, widely recognized as the Godfather of Taiwan's DRAM industry, who believes China's aggressive semiconductor expansion could eventually create a supply surplus. Meanwhile, speculation surrounding TSMC's potential entry into advanced memory manufacturing has sparked debate over the future of traditional DRAM and emerging 3D memory technologies.
According to a November 10 report from Taiwan's TechNews, Kau, a former president of Nanya Technology, warned that CXMT is expanding its manufacturing infrastructure at an extraordinary pace, with plans involving approximately 10 semiconductor fabrication facilities across China. He believes much of this additional production capacity could become operational within 2 years, introducing enough DRAM into the market to challenge established manufacturers, including Samsung, SK hynix, and Micron.
Industry projections support the possibility of substantial growth. According to SemiAnalysis, CXMT could reach approximately 420,000 monthly wafer starts during 2027, representing around 17% of global DRAM manufacturing capacity. Additional forecasts suggest production capacity could approach 600,000 wafers per month by the end of 2028, nearly doubling its estimated current output.
| Period | Estimated Monthly DRAM Wafer Capacity |
|---|---|
| 2026 | 320,000 wafers |
| 2027 | 420,000 wafers |
| 2028 | 500,000 to 600,000 wafers |
Figures represent industry capacity forecasts, not confirmed production output or completed DRAM shipments.
Such expansion could provide relief to a memory industry currently experiencing substantial pressure from AI infrastructure demand. As highlighted in earlier industry warnings about memory shortages potentially lasting until 2030, manufacturers are increasingly prioritizing high bandwidth memory (HBM) and profitable server products, restricting the availability of conventional DRAM used in gaming PCs, laptops, and consumer electronics.
However, Kau believes CXMT could disrupt this imbalance even without matching the most advanced manufacturing technologies. While established manufacturers increasingly rely on sophisticated production processes, CXMT's growing capacity could introduce substantial quantities of mainstream memory into the market. Nevertheless, increased Chinese production does not guarantee lower international prices, as domestic demand, trade restrictions, manufacturing yields, and customer qualification will influence how much additional memory reaches global customers.
Kau also questioned CXMT's ambitious 4F² DRAM cell architecture, which promises greater memory density by reducing the physical footprint of individual cells. Although CXMT has outlined plans to introduce 4F² DDR5 server memory, Kau believes the design still faces substantial engineering obstacles involving capacitor structures, electrical leakage, and production complexity. A successful demonstration would not necessarily establish commercially competitive manufacturing yields.
The discussion also extends to TSMC, following reports that NVIDIA had encouraged the Taiwanese semiconductor manufacturer to explore 3D DRAM production. Kau dismissed the possibility of TSMC becoming a conventional DRAM manufacturer, arguing that advanced logic fabrication and high volume memory manufacturing require fundamentally different expertise. TSMC has similarly emphasized that its business remains focused on dedicated semiconductor foundry services.
However, semiconductor researcher Dr. Frederick Chen believes TSMC could pursue an alternative through Back End of Line (BEOL) memory integration. Unlike conventional DRAM manufactured on dedicated memory dies, BEOL technology embeds memory structures within the interconnect layers of a logic processor, potentially improving density while reducing the distance between computing elements and stored data.
TSMC has already demonstrated a 1T1C oxide semiconductor memory architecture at the 2025 Symposium on VLSI Technology and Circuits. Its experimental design combines 1 transistor and 1 capacitor per memory cell, with a 0.75V operating voltage, retention exceeding 128 milliseconds, and random access cycle times below 8 nanoseconds.
The technology could eventually provide a denser alternative to conventional SRAM for certain on chip cache applications. However, BEOL memory remains a research technology and is not equivalent to commercially manufacturing standalone DDR5 or HBM products. There is currently no confirmed production schedule for TSMC to introduce such memory into consumer or server processors.
The possibility of CXMT significantly expanding DRAM production by 2028 is encouraging for PC enthusiasts who have experienced rising memory prices and increasingly limited upgrade options. Greater manufacturing competition could help reduce the industry's dependence on a handful of suppliers while improving access to affordable DDR5 memory.
However, additional production capacity must translate into actual consumer availability before it can influence prices. If most new manufacturing output remains concentrated within China, international markets could experience considerably less relief than the headline figures suggest.
TSMC's BEOL memory research is equally interesting, but it addresses a different technological challenge. Bringing memory closer to processing cores could eventually improve cache density and efficiency, although it is unlikely to replace conventional system RAM or solve today's DRAM shortages.
The industry should prioritize expanding accessible memory production rather than concentrating resources exclusively on high margin AI infrastructure. Affordable, upgradeable memory remains essential for PC gaming, independent development, and consumer hardware ownership.
Do you believe CXMT's expansion could finally bring DDR5 prices down by 2028, or will demand from AI data centers continue keeping memory expensive despite increased production?
