China’s CXMT Targets HBM3E Production Despite Low 20%–30% Yields

China’s memory-chip manufacturer ChangXin Memory Technologies (CXMT) is preparing to begin limited production of HBM3E memory within the next two months, despite reported manufacturing yields of only 20% to 30%. The development highlights China’s efforts to strengthen its domestic artificial intelligence hardware supply chain as access to advanced foreign memory remains constrained.

The production timeline and yield estimates were reported by Reuters on October 28, citing two people familiar with the company’s plans. CXMT did not respond to Reuters’ request for comment, while SK Hynix did not immediately respond to a request for comment on the yield comparison. The figures have not been publicly confirmed by either manufacturer. <Cite refs={["turn917963news22","turn917963search0"]}/>

Why HBM3E matters for AI processors

High-bandwidth memory, or HBM, is a specialised type of memory built by stacking memory dies vertically. Its architecture allows large volumes of data to move rapidly between memory and processors, making it an important component in high-performance AI accelerators used to train and run large language models.

HBM3E is an advanced generation of this technology. Reliable supplies of high-performance memory are essential for manufacturers seeking to produce AI processors at scale. A shortage of suitable memory can constrain accelerator output even when processor designs and other components are available.

For China, developing a domestic HBM supply is part of a wider effort to reduce dependence on overseas semiconductor technology. However, producing a functioning sample is only one stage of the process. Manufacturers must also demonstrate consistent performance, acceptable power consumption, packaging reliability and production yields before a product can be deployed at scale.

Low yields create a significant cost challenge

According to Reuters’ sources, only around 20% to 30% of CXMT’s HBM3E output currently meets specifications, compared with approximately 70% at South Korea’s SK Hynix, described in the report as the market leader.

Yield measures the proportion of manufactured units that meet the required specifications. When yields are low, a manufacturer has fewer usable units to sell from the same production effort. That can raise the effective cost of each usable product and make it harder to expand output economically.

The reported gap is substantial, although the comparison should be treated cautiously: neither company has publicly confirmed the figures in the report, and yield comparisons can depend on the product configuration, production stage and measurement method.

Reuters also reported that CXMT’s HBM3E could cost more than comparable products from overseas competitors. That would create a challenge for a company trying to offer a domestic alternative on both availability and price. The reported figures do not, however, establish the final selling price or the size of any cost disadvantage.

Samples sent as China seeks domestic supply

CXMT has reportedly sent HBM3E samples to several Chinese AI chip designers. Reuters did not identify those recipients in its October 28 report.

Earlier reporting, summarised by semiconductor-industry publication TrendForce, linked testing of CXMT’s HBM3E to Alibaba’s T-Head chip unit and Cambricon. Those names should be understood as previously reported potential testing partners, not as confirmed commercial customers or proof that their processors have completed qualification with CXMT memory. <Cite refs={["turn917963search3","turn917963search7"]}/>

Reuters also reported that CXMT had provided two versions of the memory: a standard design and a revised version incorporating changes intended to improve yields. The report did not establish how much the revised version improves production efficiency or whether either version has completed customer qualification.

That distinction matters. Sample shipments can allow chip designers to test compatibility and performance, but they do not necessarily mean that a supplier has secured a production contract or that the memory is ready for widespread commercial use.

Performance remains another hurdle

Testing described by one of Reuters’ sources indicated that CXMT’s HBM3E performed more like SK Hynix’s earlier-generation HBM3 than the South Korean company’s HBM3E. The comparison was attributed to a source and was not accompanied by a public, independently reproducible benchmark.

If the reported performance gap persists in qualified products, customers may need to weigh the benefits of domestic availability against potential differences in bandwidth, efficiency or overall accelerator performance. The available reporting does not establish how CXMT’s memory would perform across different processors or real-world AI workloads.

The competitive landscape is also moving forward. SK Hynix, Samsung Electronics and Micron Technology are established suppliers in the global HBM market. Reuters cited Counterpoint Research data showing that SK Hynix held 50% of global HBM revenue in the second quarter, followed by Samsung at 33% and Micron at 18%. These figures describe revenue shares for that period, not production capacity or market shares across all memory products. <Cite refs={["turn917963search0","turn917963search4"]}/>

Export controls and domestic demand

The push comes as Chinese AI-chip developers face difficulties obtaining advanced memory. The United States tightened export controls on certain advanced HBM products destined for China in December 2024. Reuters reported that Chinese chipmakers have increasingly sought supplies through unofficial channels, where prices can be several times higher, according to people familiar with procurement.

Domestic production could provide another route to supply if CXMT can improve yields and satisfy customers’ technical requirements. It would not automatically remove every supply-chain constraint: advanced memory depends on complex manufacturing and packaging processes, and restrictions on some semiconductor manufacturing equipment remain an obstacle.

The commercial pressure is broader than memory alone. Reuters previously reported that several Chinese AI-chip companies raised prices for current and next-generation processors by 20% to 50% as higher memory costs complicated production expansion. Those reported price increases indicate cost pressure among the companies covered; they should not be interpreted as a uniform increase across China’s entire AI-chip market. <Cite refs={["turn917963search0","turn917963search6"]}/>

What to watch next

CXMT’s next milestones will be whether its planned limited production begins on schedule, whether yields improve, and whether customers validate the memory for their AI processors. Public evidence of stable shipments, qualified designs and repeatable performance would provide stronger indications of commercial progress than sample deliveries alone.

For now, the development represents a meaningful attempt to build a domestic source of advanced AI memory, but not proof that China has closed the gap with established HBM suppliers. The central question is whether CXMT can turn early technical progress into dependable output at a competitive cost.