A memory chip that most people never see is quietly deciding who can build the next generation of AI hardware. For China, that chip has a name: DRAM made by ChangXin Memory Technologies, better known as CXMT (长鑫存储).
For three decades the global DRAM market was a closed club. Samsung, SK hynix and Micron controlled more than 90% of supply, and no new manufacturer of meaningful scale had broken in since the 1990s. CXMT is the attempt to end that monopoly — and the attempt is no longer theoretical.
The DRAM gap China closed
CXMT was founded in 2016 in Hefei, Anhui province, and within nine years it became the world's fourth-largest DRAM producer. It now sits behind only the three incumbents, with industry estimates putting its global share near 7–8% by late 2025, up from under 4% early that year.
The product story moved quickly. CXMT shipped China's first domestically developed DDR4 die in 2019 — the moment the country's DRAM industry went "from zero to one." By early 2025 it had pushed DDR5 into commercial shipments, with data rates up to 8000 megatransfers per second and dies up to 24 gigabits. Its mobile line, LPDDR5X, entered mass production in 2025 at speeds up to 10,667 MT/s, roughly two-thirds faster than the prior generation. Those are JEDEC-compliant parts, not lab curiosities: Corsair has sold retail DDR5 modules built on CXMT silicon under its Vengeance line, a proof point that the chips meet international standards in the open market.
Why HBM is the real prize
The strategic prize, though, is not PC memory. It is high-bandwidth memory (HBM), the stacked DRAM that sits next to AI training and inference chips from Nvidia, AMD and Huawei. HBM carries a fat price premium and is the single most supply-constrained component in the AI buildout.
Here CXMT is late but moving. According to reports from Taiwanese industry outlets citing mainland media, CXMT delivered 16nm-class HBM3 samples to customers including Huawei in late 2025 and is targeting full mass production in 2026. It plans to dedicate roughly 20% of its total DRAM capacity to HBM, which analysts estimate could reach around 60,000 wafers a month once ramped. HBM3E development is targeted for 2027. The company is also reported to be building a dedicated HBM back-end packaging plant in Shanghai, with production slated to come online around the end of 2026.
The honest benchmark: CXMT trails SK hynix and Samsung by what multiple reports describe as roughly three years in HBM. The incumbents are already shipping 12-layer HBM3E and moving toward 16-layer stacks. CXMT's first HBM is a credible entry, not a lead.
Building capacity without EUV
The hardest part of the CXMT story is the machine it cannot buy. Since 2019 the United States has pressured the Netherlands to keep ASML's EUV lithography out of Chinese hands, and restrictions on deep-ultraviolet (DUV) tools have tightened since. CXMT's answer has been to push DUV immersion lithography with multi-patterning to its physical limit, reaching 10nm-class features — its latest "G4" platform is built on an approximately 16nm-class process — without EUV.
That works, but it costs more and yields less than an EUV-based node. It is also why CXMT's leading products remain one to two process generations behind the Korean and American leaders, who use EUV for their cutting-edge DRAM. Capacity, however, is real: TrendForce-cited figures put CXMT's total wafer output around 280,000 wafers a month in 2025, heading toward 290,000–300,000 in 2026, with a goal of roughly 15% global DRAM share.
The IPO that could reset the memory map
CXMT has been preparing a STAR Market (科创板) listing. It began IPO tutoring (辅导) filings in mid-2024, and reports from Taiwanese financial media say it could raise between RMB 20 billion and 40 billion (roughly USD 2.8 billion to 5.6 billion) and is targeting a post-money valuation around RMB 300 billion (about USD 42 billion). If it lists, it would be one of China's largest semiconductor offerings and give the company a war chest to build HBM capacity and upgrade fabs.
Inside China, CXMT is already the default supplier for Huawei, Xiaomi, Oppo, Vivo and Honor, and cloud giants such as Alibaba, Tencent and Baidu are sourcing more of its DRAM for data centers, pushed by both cost and a national preference for domestic supply. Internationally, reports say Apple has shown interest as it diversifies beyond the big three — a sign the silicon is being taken seriously outside China.
What readers can do now
- Track HBM3 mass-production timing (2026) and HBM3E (2027) as the real signal of CXMT's AI relevance, not its DDR5 shipment headlines.
- Watch the STAR Market filing once it goes public: the prospectus will be the first audited look at capacity, margins and customer concentration.
- For supply-chain watchers, monitor CXMT's packaging partnerships (e.g., with Hefei-based foundry collaborators) as a leading indicator of HBM scale-up.
Honest limitations
CXMT is still privately held and has not published audited financials, so most revenue, profit and capacity figures circulating in media are estimates, not filings. The IPO timing and size remain reported plans, not confirmed events. The HBM3 sample and 2026 mass-production claims come from Taiwanese and mainland trade media rather than a CXMT primary disclosure, so treat the timeline as indicative. Process-node and yield comparisons to Samsung and SK hynix are based on industry analysis and should be read as directional. None of this is investment advice.