CXMT is set to start HBM3E AI chip production even though its yields are reportedly as low as 20%.

China's largest maker of conventional memory chips, ChangXin Memory Technologies (CXMT), plans to begin limited production of advanced high-bandwidth memory (HBM3E) despite manufacturing yields as low as 20%, highlighting the cost and technological barriers facing Beijing's effort to build a domestic artificial intelligence chip industry.

CXMT expects to start limited production within the next two months after sending samples of its HBM3E chips to several Chinese AI chip designers, two people familiar with the matter told Reuters. The company is pressing ahead despite producing a relatively small proportion of chips that meet specifications, as domestic AI processor makers struggle to secure enough advanced memory to expand output.

The decision underscores a central weakness in China's push to reduce its reliance on foreign AI technology. While domestic chipmakers are developing processors intended to compete with products from Nvidia, their ability to manufacture those processors at scale also depends on access to advanced memory, an area in which South Korean and US suppliers retain a substantial technological lead.

CXMT's manufacturing yields are estimated at 20% to 30%, compared with about 70% at South Korea's SK Hynix, the market leader in high-bandwidth memory, the sources said.

The Information first reported CXMT's HBM3E plans, while Reuters reported the production timeline and yield details.

High-bandwidth memory is a critical component of advanced AI processors. Manufacturers stack memory chips vertically to enable rapid data transfers between memory and processors, helping systems handle the large volumes of information required to train and run sophisticated AI models.

The technology has become necessary as AI developers build larger models and deploy computing systems that require high memory bandwidth. A shortage of suitable HBM can constrain the production and performance of AI accelerators even when processors themselves are available.

For China, that constraint has become more acute because of US export controls. Washington tightened restrictions on certain advanced HBM products destined for China in December 2024, limiting domestic AI chipmakers' access to foreign supplies. Chinese companies have increasingly sought memory through unofficial channels, where the chips can cost several times more, according to the sources.

CXMT's decision to proceed with HBM3E production reflects the pressure created by that supply gap. Even if its chips are expensive and less efficient to manufacture, a domestic source could offer Chinese AI processor designers an alternative to increasingly difficult or costly overseas procurement.

However, low yields impose a substantial economic penalty. If only 20% to 30% of chips produced meet specifications, manufacturers must absorb the cost of materials, processing, and equipment used to make a much larger number of chips than they can ultimately sell as usable products. Some defective chips may be reworked, but others must be discarded.

That means CXMT could face higher production costs at precisely the point when it is trying to supply customers that need large quantities of memory at commercially viable prices. One source said CXMT's HBM3E is expected to cost more than comparable products from overseas competitors.

The company began developing the technology internally as early as 2024, the sources said. It has since provided customers with two versions: a standard design and an updated version incorporating changes intended to improve manufacturing yields.

The updated design suggests CXMT is attempting to address the production bottleneck rather than simply maximize output regardless of quality. Whether those changes can materially improve yields, however, remains uncertain.

The immediate priority appears to be securing a usable domestic supply for Chinese AI chipmakers. That could help them increase production, but it does not necessarily mean the resulting systems will match the cost, performance or reliability of those built around leading foreign memory products.

The strain is already visible across China's AI chip supply chain. Reuters reported last month that Huawei, Cambricon, MetaX and Iluvatar CoreX had raised prices for current and next-generation AI chips by 20% to 50%, as memory costs complicated efforts to ramp up production.

CXMT's entry into HBM3E could eventually ease some of that pressure. In the near term, however, low yields and higher expected prices mean its products may provide only limited relief. Domestic availability does not automatically translate into affordable supply when manufacturing inefficiencies consume a large share of production.

CXMT's production plans also illustrate the distance China must cover to compete with established HBM suppliers. SK Hynix, Samsung Electronics and Micron Technology dominate the global market, benefiting from manufacturing expertise and experience accumulated across successive generations of advanced memory.

Counterpoint Research data cited in the report show SK Hynix held 50% of global HBM revenue in the second quarter, followed by Samsung with 33% and Micron with 18%. Most of that revenue came from HBM3E products.

CXMT, by contrast, held 10% of global DRAM revenue in the same quarter, making it the world's fourth-largest supplier of conventional memory chips. Its position in standard DRAM demonstrates an established manufacturing base, but producing conventional memory at scale is not equivalent to mastering the more demanding process of stacking and integrating advanced HBM.

One source said CXMT's HBM3E performance in testing was closer to SK Hynix's previous-generation HBM3 than to the South Korean company's own HBM3E. If that assessment holds, CXMT would face a disadvantage not only in manufacturing efficiency but also in product performance.

The gap is widening as competitors move to newer technologies. Samsung said in July that it had expanded HBM4 sales and shipped HBM4E samples to major customers, putting CXMT at least one generation behind in commercial HBM development.

US export controls further complicate the effort by restricting CXMT's access to some advanced chipmaking equipment. That limits the tools available to improve manufacturing processes and close the gap with international rivals.

The broader implication is that China's AI ambitions depend on more than developing domestically designed processors. The country also needs a reliable supply of advanced memory that can be produced at sufficient volumes, acceptable yields, and competitive prices.

CXMT's willingness to begin HBM3E production despite high defect rates shows how urgently Chinese chipmakers need alternatives to restricted foreign supplies. But the move also exposes the trade-off facing Beijing: prioritizing domestic availability may help keep AI hardware development moving, while low yields and inferior performance risk making that hardware more expensive and less competitive.