A denser cell, aimed at servers
Cao Kanyu, president of the Chinese memory maker ChangXin Memory Technologies, told an industry conference in Shanghai on Saturday that the company plans to bring a next-generation DDR5 RDIMM server memory module built on a 4F² DRAM architecture to market by the end of 2026, the state-owned Global Times reported.
The number describes the area of a memory cell in units of the process’s minimum feature size. Mainstream DRAM uses a 6F² layout. Moving to 4F² packs the cells tighter: Global Times put the gain in cell density per chip at 30% to 50%. CXMT says the enabling part is a vertical surrounded gate transistor, which the company says limits current leakage and short-channel effects as the cell shrinks. It has published more than 300 patents related to the architecture.
Why the server part matters
CXMT could have led with a consumer product. It did not. Server DDR5 is the segment where AI build-outs are consuming supply and where margins have moved furthest, and it is also the segment with the hardest qualification bar: a module has to survive a hyperscaler’s validation before it ships in volume.
Tian Feng, president of the Fast Think Institute and a former dean of SenseTime’s Intelligence Industry Research Institute, told Global Times that the world’s three largest memory makers are all working on 4F² or comparable three-dimensional DRAM roadmaps, but that “none has achieved large-scale commercialization yet”. Ma Jihua, a veteran industry analyst quoted in the same report, said a stable architecture would demonstrate server-grade readiness and add pressure on other DRAM producers.
What has not been said
Almost everything a buyer would need. CXMT has not published capacity, speed, voltage or rank configuration for the planned module, has not said where sampling stands, and has not named a customer or a qualification result. Independent coverage has made the same point: igor’sLAB noted that the timeline does not confirm mass production, and that CXMT has not documented how fully the planned product will use 4F².
The comparisons being drawn elsewhere are also framing rather than fact. TweakTown wrote that shipping this year could put CXMT about two years ahead of Samsung on 4F², but no rival has published a schedule that would make that a like-for-like race.
The share picture
Global Times cited an industry report putting second-quarter 2026 DRAM share at 39% for Samsung, 26% for SK Hynix and 25% for Micron, and gave CXMT 10% on Counterpoint Research’s numbers. A tenth of the market is not a challenger position in server memory, where qualification history counts, but it is enough that a working 4F² part would be noticed by the three companies above it.
What to watch
Two things, in order. First, whether a module actually appears before 31 December, or whether “by year-end” becomes “sampling by year-end” — the distinction between a design win and a volume win is the whole story. Second, whether any customer confirms qualification. Until one does, this is a roadmap stated by a company with a reason to state it, at a conference, in a market where everyone’s capacity plans are a form of argument.
