CXMT: China's Largest A-Share IPO This Year Is a Memory Test
A memory shortage is testing how much usable supply China can source at home after years of industrial investment.
This piece which was produced in collaboration with TP Huang, who has been writing about semiconductors and advanced Chinese technology on Twitter and Substack for years. We hope to add more semiconductor content to our pipeline and already have several related research projects in the works.
Things You Might Have Missed
A few recent notes from the broader Tech Buzz China universe before the CXMT piece:
CATL’s Q1 2026 profit: CATL earned RMB 20.74 billion in Q1 2026, more than Chery, Geely, BYD, SAIC, Great Wall, Seres, and Changan combined.
Embodied intelligence funding: China’s embodied intelligence sector raised RMB 46 billion across 288 deals in H1 2026, with the top five companies taking 37% of the total.
From Weijin Research: Zhipu’s GLM-5.2: A Usability Breakthrough for Chinese Open-Source Models? With open weights and strong benchmarks, GLM-5.2 may mark a turning point for putting China’s open-source AI to practical use.
You may have seen CXMT in the headlines recently. It is legitimately a big deal. Earlier this week, SemiAnalysis published its own in-depth report on the company. This is our take.
ChangXin Memory Technologies (CXMT), China’s leading domestic DRAM maker, is coming to market just as the memory industry is enjoying its strongest tailwinds in years. The Hefei-based company is seeking RMB 29.5 billion ($4.1 billion) on Shanghai’s STAR Market in what would be the largest mainland A-share IPO so far this year.
Memory has become one of the defining AI investment themes. While GPUs capture most of the headlines, AI servers also require enormous amounts of DRAM, the fast working memory used while servers, PCs, smartphones, and AI systems are running. Investors have rewarded memory manufacturers accordingly: shares of Micron and SK hynix (both around $1.3T in market capitalization as of this writing) have roughly quadrupled over the past 18 months as AI demand reshaped expectations for the sector.
Much of that excitement has centered on HBM, the stacked high-bandwidth DRAM used alongside AI accelerators. But CXMT’s business today is built on conventional DRAM. In the first quarter of 2026, the company reported revenue of RMB 50.800 billion ($7.0 billion), up 719.13% year over year, and net profit of RMB 33.012 billion ($4.56 billion). Those results suggest China’s domestic DRAM market has already reached meaningful scale, even before HBM becomes a material contributor.
That is why “China’s Micron” is probably the more useful shorthand than “China’s SK hynix.” SK hynix remains the benchmark for HBM leadership. CXMT’s opportunity today lies in mainstream DRAM: DDR5, the current standard for servers and PCs; LPDDR5X, the lower-power memory used in mobile devices; and server DRAM produced at sufficient scale for China’s cloud providers, server manufacturers, smartphone brands, PC OEMs, and AI infrastructure builders to treat it as a viable supply option.
That alone would represent a meaningful shift. China can reduce its reliance on imported DRAM by sourcing a growing share of the memory it buys in volume from CXMT rather than Samsung, SK hynix, or Micron. As AI drives tighter memory markets, that may be enough to reshape China’s memory supply chain.
Hefei’s Long DRAM Bet
CXMT traces its roots to a group of entities established in Hefei in 2016, including Innotron Memory, Hefei Rui-li Integrated Circuit Manufacturing (睿力集成), and Hefei Zhiju (合肥智聚). Today, the company is formally Changxin Keji Jituan Gufen Youxian Gongsi (长鑫科技集团股份有限公司), or CXMT Corporation. Its founder and chairman, Tsinghua alumnus Zhu Yiming (b. 1972), is best known for founding GigaDevice (兆易创新), China’s leading flash memory and microcontroller (MCU) company. GigaDevice gave Zhu credibility in memory-adjacent semiconductors, but not a ready-made DRAM manufacturing business.
China was one of the world’s largest buyers of memory chips but not a meaningful producer of DRAM. Samsung, SK hynix, and Micron dominated the market, while mainland China lacked a globally relevant DRAM manufacturer.
The project emerged after China’s failed 2015 attempt to acquire Micron underscored its dependence on foreign DRAM. Hefei’s municipal government, already known for deploying state capital with a venture-capital-like investment model, partnered with Zhu to build a domestic DRAM champion. When private investors balked, Hefei state capital reportedly took an 80% stake in an initial RMB 150 billion 12-inch DRAM fab project. Code-named the “506 Project,” it was established in Hefei’s Airport Economic Demonstration Zone.
By 2019, trade publications described CXMT as a company with “no DRAM heritage” but a functioning fab. It had built Fab 1 and an R&D center in Hefei, reached production capacity of 20,000 wafers per month, and was producing 8Gbit LPDDR4 and DDR4 chips on a 19nm process. Like many latecomers, it recruited engineers from South Korea and Taiwan, hired former Qimonda technical staff including Karl-Heinz Kuesters, who spent 24 years at Qimonda and Infineon, and reportedly licensed Qimonda’s buried-wordline DRAM technology through WiLAN/Polaris.
CXMT is less a founder story than an industrial policy project. DRAM rewards accumulated capabilities: capital, manufacturing know-how, intellectual property, experienced engineers, yield learning, process documentation, customer relationships, and a government willing to finance the difficult middle years. China did not suddenly develop a DRAM industry. It took CXMT nearly a decade to move domestic DRAM from aspiration to commercially viable supply.
That industrial bet is now seeking RMB 29.5 billion on Shanghai’s STAR Market. As noted earlier, if completed, it would be the largest mainland A-share IPO so far this year. Eastmoney data covering completed mainland A-share IPOs between January 1 and June 23 shows the largest completed offering raised about RMB 5.03 billion ($694 million). Across 67 completed IPOs, total proceeds reached roughly RMB 60.0 billion ($8.3 billion). CXMT’s proposed raise would be almost six times larger than this year’s biggest completed A-share IPO and roughly half of all mainland A-share IPO proceeds raised year to date.
The proceeds are targeted rather than general corporate funding. The prospectus allocates RMB 13.0 billion ($1.8 billion) to DRAM technology upgrades, RMB 7.5 billion ($1.0 billion) to memory wafer manufacturing line upgrades, and RMB 9.0 billion ($1.2 billion) to next-generation DRAM research and development, spanning node upgrades, manufacturing capacity, and future memory technologies.
Investor appetite has already been evident. J.P. Morgan noted that China Merchants Securities-H rose 6.5% on May 19, versus a 1.2% gain for the average H-share broker, after CXMT updated its prospectus. JPM attributed the move primarily to renewed interest in China Merchants Securities’ estimated 0.54% pre-IPO stake in CXMT through a subsidiary. Even indirect exposure proved enough to move the stock.
The Cycle Is Real, But So Is The Output
CXMT is coming public amid a sharp memory upcycle. With memory among the semiconductor industry’s most cyclical businesses, today’s financials deserve some caution.
The company’s main-business gross margin improved from -2.19% in 2023 to 5.00% in 2024 and 41.02% in 2025. The prospectus cautions against extrapolating recent performance, noting that first-half 2026 results may not be sustainable — a reflection of an industry where DRAM prices swung from a high of $7.89 per GB to a low of $1.78 per GB in the first half of 2023.
The pattern is familiar: the industry boomed between 2016 and 2018 before pricing and profits rolled over, then again in 2021 and 2022 before the 2022-23 downturn forced production cuts. The current cycle has different drivers — AI demand, HBM allocation, and tight conventional DRAM supply — but memory remains an industry where favorable conditions can reverse quickly once supply catches demand.
Omdia data cited in CXMT’s prospectus show Samsung, SK hynix, and Micron (the DRAM industry’s “trigopoly”) controlled more than 90% of global DRAM revenue in 2025, with shares of 33.96%, 34.48%, and 23.41%, respectively. CXMT ranked fourth globally and first in China, increasing its share from 3.97% in the second quarter of 2025 to 7.67% in the fourth. The memory upcycle undoubtedly helped, but the pace of that gain is still quite remarkable.
Capacity is one reason investors are looking beyond CXMT’s current market share. As of June 30, 2025, the company operated three 12-inch DRAM wafer fabs across Hefei and Beijing. Industry estimates put Micron at around 320,000 wafer starts per month and CXMT at roughly 250,000-300,000 by the end of 2025, putting the two companies in roughly the same league by manufacturing capacity.
That does not mean they are economically comparable. Similar wafer capacity can produce very different financial results depending on yield, die density, customer and product mix, qualification status, and pricing power. Customers can still view CXMT as a second source rather than the market benchmark, limiting its ability to command incumbent-level margins.
More broadly, the focus on EUV as the defining constraint on China’s semiconductor industry can be misleading. DRAM is not leading-edge logic and depends less on the latest lithography tools. Micron reached its 1-beta DRAM node without EUV, and the available evidence suggests CXMT’s current expansion is not primarily constrained by access to the newest lithography equipment or by chemical restrictions that would halt production.
CXMT has also built a deep domestic supply chain, compiled below:
Nonetheless, the technology gap remains real. Trade-press reporting places CXMT’s G4 DDR5, the latest mainstream DRAM standard for servers and PCs, on roughly a 16nm, or D1z-like, process, with a G5 roadmap moving toward D1a- and D1b-like territory. By feature size, CXMT appears to trail the Big 3 by roughly three years.
At the product level, however, the gap is narrower. CXMT has announced mass production of LPDDR5X at 8533 Mbps and 9600 Mbps, with 10667 Mbps sampling, and has showcased DDR5 products for servers, workstations, and PCs with peak speeds of 8000 Mbps and die densities of up to 24Gb.
China can build competitive DDR5. The harder part is manufacturing it consistently at the yield, cost, reliability, and scale demanded by major customers.
At CXMT’s current scale, relatively small changes in wafer starts, yield, die size, or DDR5 pricing can have an outsized impact on revenue. One rough model begins with quarterly wafer starts and DDR5 pricing:
Even under conservative assumptions, the revenue outlook remains highly sensitive to commodity DRAM prices, yield, and usable output. If conventional DRAM prices remain elevated, CXMT can generate substantial profits during a memory shortage without leading in HBM.
Over time, investors are likely to care less about wafer capacity than qualified bit output. Producing and selling DRAM at something approaching Micron’s scale would say far more about CXMT’s competitive position than the number of fabs it operates.
AI, HBM, and Why Ordinary DRAM Matters
HBM remains the defining memory for frontier AI accelerators. But it also reshapes the conventional DRAM market because Samsung, SK hynix, and Micron have finite manufacturing capacity, engineering resources, and customer allocation. As more of those resources are directed toward HBM and leading AI customers, conventional DRAM tightens. AI has made ordinary memory more valuable just as China has, for the first time, a domestic supplier with meaningful scale.
Despite the AI excitement, CXMT remains primarily a conventional DRAM company. Mobile devices accounted for 60.40% of DRAM revenue in 2025, with servers contributing 26.51% and PCs 3.87%. Revenue from DRAM used in AI-compute servers remained limited, although server DRAM revenue overall grew from RMB 395 million ($55 million) in 2023 to RMB 1.927 billion ($266 million) in 2024 and RMB 15.970 billion ($2.2 billion) in 2025, helped by the rapid adoption of DDR5.
Customer qualification is becoming just as important as technology. The prospectus lists Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, Honor, OPPO, and vivo among CXMT’s end customers, although direct sales often flow through distributors. Qualification is beginning to spread into the module ecosystem as well. Gloway and KingBank have introduced DDR5 modules using CXMT chips, enabling 48GB and 96GB kits; Corsair has used CXMT DDR5 in certain Vengeance products; and HP and Dell have begun qualifying CXMT memory.
The current shortage could accelerate that process. In a well-supplied market, customers have little incentive to qualify another DRAM supplier. During a shortage, however, that calculation changes. That does not guarantee lasting demand once the cycle turns, but it can accelerate commercial adoption.
HBM nevertheless remains an important source of upside. Industry reporting in 2024 said CXMT had begun mass-producing HBM2, while noting yield uncertainty. HBM2 is an earlier generation than the HBM3 and HBM3E used in today’s leading AI accelerators, where Samsung, SK hynix, and Micron have already established leadership and are moving toward HBM4. Other reporting has discussed CXMT roadmaps for HBM3 and HBM3E, leaving room for further catch-up.
That assumes China must follow the same technology path as the global leaders. Huawei’s HiZQ roadmap suggests another possibility. Rather than asking whether CXMT can match SK hynix generation for generation, the question becomes whether Chinese accelerator makers can build competitive AI systems around a domestic memory ecosystem. That does not mean CXMT is qualified into Huawei accelerators today. It does suggest that the value of domestic DRAM may depend as much on how it is integrated into Chinese AI systems as on matching the latest HBM generation.
Packaging remains one of the biggest bottlenecks. Advanced HBM packaging is technically demanding, and domestic packaging capacity remains a key bear-case risk. If Chinese accelerator makers can adopt packaging approaches that are easier to manufacture and scale, the competitive gap may be wider — or narrower — than a simple HBM generation comparison suggests.
For now, the investment case is centered on mainstream DRAM. CXMT’s business is built on DDR5, LPDDR5X, and growing customer adoption across servers, smartphones, and PCs. HBM would be an additional growth driver, not a prerequisite.
An Overlooked Constraint: Tungsten
Memory is often discussed in terms of fabs, process nodes, and packaging. But as our earlier diagram shows, the supply chain begins much earlier.
Semiconductor manufacturing depends on gases, specialty chemicals, powders, deposition materials, lithography tools, spare parts, and qualified suppliers. Final chipmakers are often described as independent, but that independence rests on long and sometimes fragile upstream supply chains.
One example is tungsten hexafluoride (WF6), a deposition gas used in both logic and memory manufacturing, including DRAM, NAND, and 3D NAND. Supplier materials describe semiconductor-grade WF6 as a critical process material.
Market research has suggested Japanese suppliers warned Korean semiconductor manufacturers of potential WF6 shortages, with inventories expected to last only until mid-2026. Chinese suppliers are estimated to account for more than half of global WF6 production.
A more aggressive version of the argument goes further. Partner-side trade data suggest China’s tungsten powder exports have fallen 85%. If sustained, that could leave Japanese and Korean producers of tungsten hexafluoride (WF6) with less feedstock just as Chinese WF6 suppliers prioritize domestic customers. For now, this remains conjecture. As of June 2026, we have not seen evidence that WF6 supply has constrained Samsung, SK hynix, or Micron.
If that were to change, the discussion around CXMT would shift as well. The question would no longer be only whether China can catch the global memory leaders technologically. It would also be whether control over key upstream materials gives China’s domestic memory industry an advantage that is not obvious from process nodes or product roadmaps alone.
What Would Change Our Mind
 | A guest post by
|