CXMT to start limited HBM3E output at 20-30% yield

ChangXin Memory Technologies (CXMT) is moving HBM3E toward the production line. Reuters, citing two people familiar with the matter, reports that the Hefei-based memory maker has already sent HBM3E samples to several domestic AI chip designers and plans to start limited production within the next two months. The same sources put the yield at only 20% to 30%, against roughly 70% for comparable SK Hynix products.

The report adds that in testing, CXMT's samples performed closer to SK Hynix's previous-generation HBM3, and that they are expected to be priced higher than comparable overseas products. Low yield, performance a generation behind and a higher price have not stopped CXMT from pushing ahead, and the reason lies on the demand side: under US export controls, Chinese AI chipmakers cannot get enough high-bandwidth memory.

Illustration: a few intact HBM stacks on a wafer contrast with a tray of rejects, showing low yield

Two accounts that do not line up on timing

Public information on how far CXMT's HBM3E has actually progressed is inconsistent. In early September, The Information reported that CXMT had already begun small-batch production of HBM3E, that Alibaba's T-Head and Cambricon were testing their own processors with it, and that capacity expansion was set for 2027. Reuters' latest account, "limited production within the next two months," runs a step behind that.

The two reports may describe different stages: the former an engineering batch, the latter customer-facing limited supply. CXMT, T-Head and Cambricon have not publicly confirmed any of the trial production or testing, and there have never been official figures for yield, stack height or customer qualification progress, so for now the sources' account is all there is to go on.

The yield estimates, by contrast, are close to one another. Research firm SemiAnalysis put it at about 25% in June; a September report from Korea's ChosunBiz cited a trial-production yield of about 25%, though the product it described was an 8-layer HBM3. Reuters' 20% to 30% falls in the same range, showing no clear improvement and no clear deterioration over the months.

Where it trails the two Korean makers

Laying the timelines side by side makes the gap easier to see. SK Hynix began mass-producing HBM3E in 2024, with Samsung following; if CXMT starts limited production around the end of this year as Reuters describes, it is about two years late. People familiar with CXMT have previously estimated that it trails the leading makers by three to five years technologically.

The difficulty with HBM is stacking. Eight or twelve DRAM dies must be joined vertically with through-silicon vias and then packaged, and a fault in any one layer scraps the whole stack, so overall yield is the product of the yields of each step. SK Hynix's roughly 70% was itself reached only after several product generations. CXMT had a similar start with DDR5: in late 2024, Korean media reported its DDR5 yield might be only 10% to 20%, and it climbed gradually from there.

How tight the demand side is shows up in institutional reports. Morgan Stanley earlier judged that HBM3E supply would be the main bottleneck for China's AI GPU shipments in 2027; it also estimated that, even at a low yield, CXMT could still produce about 3 million to 4 million HBM3E stacks in 2027. How many accelerator cards that capacity could serve depends on how many HBM stacks each chip uses, and the report gives no breakdown.

The capital market has already given it a place

CXMT Technology listed on the STAR Market on July 27 this year and reached its three-month mark on October 27. China Securities Index Co. has now disclosed that it has been added to the SSE Composite Index with a weight of 4.67%, the largest, ahead of Agricultural Bank of China and ICBC; in the SSE STAR Market Composite Index its weight is 15.74%, also first, followed by Cambricon and Hygon Information.

The valuation the market is giving it today comes mainly from the price-rise cycle in commodity DRAM. Whether HBM3E can get from sampling to stable supply depends on how high yield climbs next year: if it stays at two or three in ten, the cost per stack will be hard to bring down to a level where chipmakers are willing to buy in volume, even if no cheaper domestic alternative exists for now.

Sources: Reuters, The Information, CocoLoop, ChosunBiz, Gelonghui; used to check CXMT's HBM3E yield range, the SK Hynix yield comparison and SSE index weight data.