--- title: "CXMT Launches HBM3E Production for AI with Only 20–30% Yield: China Attempts to Close Accelerator Memory Deficit" description: "China's CXMT is launching small-scale production of HBM3E memory for AI accelerators, despite a yield of only 20–30% and lagging behind SK hynix and Samsung by at least one generation." date: 2026-10-28T16:07:00.450Z lang: en url: https://xab.info/en/posts/cxmt-hbm3e-memory-mass-production-ai tags: [cxmt, hbm3e, ai-memory, china-chips, dram, export-controls] author: "Alexander Buran" publisher: "XAB.info" --- # CXMT Launches HBM3E Production for AI with Only 20–30% Yield: China Attempts to Close Accelerator Memory Deficit ![CXMT semiconductor AI chip interface with glowing elements on a tech board](https://xab.info/media/2026/10/28/cxmt-hbm3e-massovyy-vypusk-pamyati-dlya-ii/cxmt-hbm3e-massovyy-vypusk-pamyati-dlya-ii-1.webp) ## Key Points - CXMT will begin small-scale production of HBM3E for AI accelerators within the next two months and has already sent samples to local developers. - The yield of usable chips is only 20–30% compared to about 70% at SK hynix, making production expensive. - In terms of characteristics, CXMT's HBM3E is closer to SK hynix's basic HBM3; the company lags behind leaders by at least one generation. - The launch is driven by the AI memory shortage in China amid US export restrictions introduced in December 2024. Chinese memory manufacturer ChangXin Memory Technologies (CXMT) plans to begin small-scale production of HBM3E within the next two months—an advanced type of high-speed memory used in tandem with artificial intelligence processors. This is reported by Reuters, citing industry sources. The company has already sent chip samples to several local AI accelerator developers, making the launch one of China's most noticeable steps in a segment previously controlled effectively by just three players: SK hynix, Samsung, and Micron. ### Memory Shortage and Sanctions Background CXMT's decision is directly tied to the acute shortage of AI memory in China. These chips are critical for implementing the national program to create domestic alternatives to Nvidia AI accelerators, shipments of which advanced models to the PRC Washington restricts. In December 2024, the US tightened export controls on certain types of HBM, after which Chinese manufacturers increasingly try to procure such memory through unofficial channels, where prices can be several times higher than market rates. Under these conditions, domestic production—even if imperfect—becomes a strategic necessity. ### The Economics of Low Yields The key challenge of the launch is the extremely low yield of usable dies: according to Reuters, it stands at around 20–30%, whereas industry leader SK hynix achieves roughly 70%. Due to this high defect rate, production is costly, requiring a significant portion of output to be discarded or sent back for reworking. According to unofficial information, CXMT began its own HBM3E development back in 2024, and the finished chips will be more expensive than those of foreign competitors. Cost pressures have already affected the market: Huawei, Cambricon, MetaX, and Iluvatar CoreX previously raised prices on their AI accelerators by 20–50% due to rising memory costs, complicating volume expansion plans. ### Characteristics and Lag Behind Leaders In terms of technical parameters, CXMT's HBM3E is, according to Reuters sources, closer to SK hynix's basic HBM3 than to the same manufacturer's more advanced HBM3E. The company offered clients two chip variants—standard and updated, featuring modifications to improve yields. HBM is created by vertically stacking chips on top of each other, providing high-speed data exchange with processors, which is critical for training and inference of large language models. Meanwhile, in July Samsung announced the expansion of HBM sales and the delivery of HBM4E samples to key clients, highlighting that CXMT lags behind recognized leaders by at least one generation. ### Market Positions and Broader Context At the end of Q2, SK hynix accounted for 50% of global HBM sales revenue, Samsung about 33%, and Micron approximately 18%. In the same quarter, CXMT captured 10% of the global DRAM market, becoming the fourth-largest supplier of traditional memory chips. Due to US sanctions, the company cannot acquire advanced manufacturing equipment, making it difficult to close the technological gap. Nevertheless, CXMT demonstrates momentum: less than a year after mastering LPDDR5, it announced the start of mass production of LPDDR6, beating recognized leaders on timelines. The launch of HBM3E, even with a high defect rate, cements China's status as a participant in the race within the memory industry's most complex segment. ## Fact-Check & Sources - [CXMT запустит массовый выпуск HBM3E, несмотря на катастрофический уровень брака window-new](https://3dnews.ru/1150544/cxmt-zapustit-massoviy-vipusk-hbm3e-nesmotrya-na-katastroficheskiy-uroven-braka) - Primary source: small-scale HBM3E launch in the next two months, 20–30% yield vs ~70% at SK hynix, market shares, sanctions, accelerator price increases. - [Китайская CXMT впервые в мире начала массовый выпуск LPDDR6](https://overclockers.ru/blog/Global_Chronicles/show/262723/Kitajskaya-CXMT-vpervye-v-mire-nachala-massovyj-vypusk-LPDDR6) - Context: On 29.08.2026, CXMT started mass production of LPDDR6 (12.8 Gbps) for the Xiaomi 18 Fold—used to characterize the company's momentum. - [CXMT запустила пробное производство памяти HBM3E](https://3dnews.ru/1147730/cxmt-zapustila-probnoe-proizvodstvo-pamyati-hbm3e) - Confirms the fact of trial HBM3E production and JEDEC standard parameters (1024-bit width, 9.2–12.4 Gbps per pin). - [Китайская память выходит на новый уровень: CXMT приблизилась к созданию HBM3E для ИИ](https://www.ixbt.com/live/offtopic/kitayskaya-pamyat-vyhodit-na-novyy-uroven-cxmt-priblizilas-k-sozdaniyu-hbm3e-dlya-ii.html) - Reference to The Information: start of small-scale HBM3E production, market control by three players, scaling plans by 2027. ## FAQ ### What is HBM3E and why is it needed? HBM3E is high-speed memory created by vertically stacking chips on top of each other. It provides fast data exchange with processors and is critical for the operation of AI accelerators used to train and run large language models. ### What is CXMT's chip yield rate? According to Reuters, the share of usable products is about 20–30%, whereas at SK hynix this figure reaches approximately 70%. Due to high defect rates, production is expensive. ### Why is China producing HBM domestically? An acute shortage of AI memory arose in China after the US tightened export controls on certain types of HBM in December 2024. This forces CXMT to launch its own production despite technological lag. ### How far behind the leaders is CXMT? In terms of characteristics, CXMT's HBM3E is closer to basic HBM3 from SK hynix than to advanced HBM3E. Following Samsung's announcement of HBM4E samples, the lag is estimated at least one generation.