On October 28, Reuters reported that Chinese memory maker CXMT has shipped HBM3E samples to domestic AI chip designers and expects to begin limited production within two months. According to people familiar with the matter, yields are only 20–30%, far below SK hynix's roughly 70%. Assuming the cost of each unit put into production is the same, the manufacturing cost per good unit would be 2.3 to 3.5 times that of SK hynix. CXMT is pressing ahead because Chinese AI chipmakers cannot secure enough advanced memory to expand production. Even at a higher manufacturing cost, being able to source memory domestically is likely to carry substantial value.
How much does a 20–30% yield affect manufacturing costs?
According to Reuters, CXMT offered customers two types of product: a standard-specification version and a modified version redesigned to improve yields. However, performance observed in testing was reportedly closer to the older-generation HBM3 than to SK hynix's HBM3E. Because of the low yields, manufacturing costs are expected to exceed those of comparable products from overseas makers, and many defective units will have to be scrapped or reworked. CXMT nonetheless plans to move ahead with production, aiming to supply Chinese AI chipmakers that are struggling to obtain advanced memory.
Yield is the share of products put into the manufacturing process that can be shipped as good units. For example, if 100 units are put into production and SK hynix's yield is about 70%, it gets about 70 good units, while CXMT at 20–30% gets only 20 to 30. If the cost per unit put into production is the same, CXMT's cost per good unit is about 2.3 to 3.5 times SK hynix's. These multiples come from 70÷30 and 70÷20, and are a simple calculation that assumes both companies have the same per-unit input cost.
It is worth noting that the yield figures themselves carry uncertainty. The numbers reported by Reuters came from people familiar with the matter and were not officially disclosed by CXMT. It is also unclear which product or how many stacked layers SK hynix's roughly 70% refers to, or whether the two companies' yields cover the same range of process steps. Actual manufacturing costs are also affected by differences in per-unit input costs and by the share of products that can be recovered through rework, so the 2.3–3.5x figure cannot be treated as the actual difference in the two companies' production costs.
HBM is made by vertically stacking multiple DRAM chips, so losses from defective units tend to grow as the process advances. In a technical briefing from November 2023, Japan's Advantest explained that HBM requires additional inspection steps after stacking, which increases demand for testers. A product that fails inspection after stacking and cannot be reworked wastes all the materials and processing costs invested up to that point. A manufacturer with low yields must either pass these losses on in its selling prices or absorb them itself.
In addition, in ChosunBiz's report that we covered on September 10, the yield for CXMT's 8-layer HBM3 was put at about 25%. The newly reported 20–30% relates to HBM3E, a different product, so it cannot be concluded that yields have not improved.
Why CXMT is pressing ahead despite higher costs: China's sourcing difficulties
In September, Reuters reported that four companies—Huawei, Cambricon, MetaX, and Iluvatar CoreX—had raised prices on their current and next-generation AI chip products by 20–50%. Rising memory prices were reportedly an obstacle to expanding production. According to the report, Chinese chipmakers have become dependent on expensive gray-market HBM since U.S. export controls were tightened.
In this situation, what matters to AI chipmakers is not only the purchase price of memory but also whether they can secure the volumes they need. Without HBM, they cannot complete and ship AI accelerators, and the companies have already raised product prices citing higher memory costs. Domestic HBM that can be obtained reliably could therefore be an attractive option even if it costs more to make than overseas products. However, CXMT's selling prices have not been reported, so it is unclear whether it will pass its higher manufacturing costs on to customers or absorb them. What can be explained at this stage is CXMT's economic incentive to proceed despite low yields; whether it becomes a genuinely advantageous supplier for customers is a separate question.
The fact that CXMT is also offering customers a modified version aimed at improving yields suggests it intends to refine its manufacturing process through limited production while building a track record of supply. According to Reuters, CXMT began in-house HBM development in 2024, which means it took about two years from the start of development to shipping samples.
SK hynix, meanwhile, announced the start of mass production of 12-layer HBM3E in September 2024. The number of layers in CXMT's HBM3E has not been reported, so the products' specifications and manufacturing difficulty cannot be compared directly. Even so, there is a gap of about two years between SK hynix's mass-production announcement and the expected start of CXMT's limited production. For Chinese manufacturers to narrow the gap with the leading players, they will need to keep improving manufacturing technology and yields.
Performance challenges and the indicators to watch through year-end
The biggest concern in considering adoption of CXMT's HBM3E is performance. According to Reuters, in tests by customers, CXMT's product performed closer to the older-generation HBM3 than to SK hynix's HBM3E. However, CXMT has not disclosed specific performance figures such as the HBM3E's memory bandwidth, so the size of the gap is unknown.
A useful benchmark is the HBM3 standard that JEDEC established in 2022. Under the standard, HBM3's maximum data rate is 6.4Gb/s per pin, which gives a maximum bandwidth of 819GB/s per device. SK hynix's 12-layer HBM3E, which entered mass production in 2024, runs at 9.6Gb/s per pin. The ratio between the two speeds is 1.5x, and assuming the same 1,024-bit-wide interface, 9.6Gb/s corresponds to a bandwidth of about 1.23TB/s.
It is unknown what data rate CXMT's product actually achieves. In AI accelerators, insufficient memory bandwidth can mean data cannot be fed to the compute units fast enough, preventing them from delivering their full performance. Having a product that can be sourced and having it meet customers' performance requirements and actually be adopted are separate matters.
Still, in a Chinese market where advanced memory is hard to obtain, there is considerable significance in evaluating HBM that can be sourced domestically and building up a supply track record, even if it falls short of overseas products in performance. This assessment applies first to why CXMT is moving to limited production; it does not guarantee full-scale mass production or adoption in commercial products.
It also remains unconfirmed which AI chips will actually carry the memory. According to The Information's August 31 report, which we covered on September 1, T-Head and Cambricon are conducting adoption trials pairing CXMT's HBM3E with their own processors. The plan is conditional: if the trials go smoothly, the chips would carry the memory from 2027, and adoption in mass-produced products has not been decided. Cambricon is also one of the four companies whose AI chip price increases were reported in September.
There are two conditions under which this outlook would fall apart. One is if the start of limited production cannot be confirmed within the "two months" Reuters cited, that is, by around the end of 2026. The other is if it is reported that the design companies that received samples have declined to adopt them because of performance, quality, or other reasons.
Improvement in yields will also be an important indicator of progress. Using SK hynix's roughly 70% yield as the baseline and assuming the same per-unit input cost, CXMT's yield would need to exceed 35% for its cost per good unit to fall below twice SK hynix's (70÷2=35). If the modified version were confirmed to exceed that level, the disadvantage on manufacturing cost could be judged to have narrowed. However, actual adoption will also depend on performance, quality, supply volumes, and selling prices. Whether CXMT can start limited production as planned, and whether it can reliably supply products that meet customers' requirements, will determine whether domestic HBM can grow from a stopgap response to sourcing difficulties into a sustained source of supply.