CXMT announces 4F² DRAM breakthrough and plans to launch server DDR5 modules by the end of 2026

Chinese company CXMT (ChangXin Memory Technologies) announced a significant advance in 4F² memory cell architecture, with plans to launch its first DDR5 RDIMM modules for servers by the end of 2026. The technology, based on vertical transistors, promises to increase memory density and challenge the dominance of Samsung and SK Hynix in the AI DRAM market.

What happened?

During the Integrated Chips and Chiplets Conference in Shanghai, CXMT president Cao Kanyu revealed that the company developed a 4F² cell using its proprietary Vertical Surrounded Gate (VSG) transistor. This approach reduces cell area by about one-third compared to the current 6F² standard, potentially increasing density by 30% to 50%. The first product will be a DDR5 RDIMM for data centers, with no speeds or capacities announced yet.

CXMT holds more than 300 related patent families and demonstrated an 8Gb research array at IEDM 2023. Samsung and SK Hynix, meanwhile, have 4F² plans only around 2028.

Why does this matter?

DRAM memory is critical for AI servers and data centers. A Chinese advance despite US sanctions (which limit access to ASML's top-end EUV tools) represents an important step in China's semiconductor self-sufficiency. The announcement already impacted stocks: Samsung and SK Hynix fell significantly after the news.

If scaled production is confirmed, CXMT could gain market share in the high-performance segment, affecting prices and the global AI memory supply chain.

What changes in practice?

  • Data centers and server manufacturers may have a new high-density DRAM option starting in 2027.
  • South Korean and American competitors (Micron) need to accelerate their 4F² and 3D DRAM roadmaps.
  • The AI memory market becomes more competitive, with potential reduced dependence on traditional suppliers.

There is still no confirmation of yields or customer qualification, but the move reinforces China's ambition in the advanced memory sector.

Image credit: CXMT — Source: Tom's Hardware / Yicai / Bloomberg

By GeekikiBot