ChangXin Memory Technologies (CXMT), China's largest DRAM manufacturer, plans to begin limited production of HBM 3E within two months despite yields of only 20% to 30%, Reuters reported1,2. The chips' performance is closer to that of SK Hynix's previous-generation HBM 3 rather than the current HBM 3E standard.
Yield economics and supply pressure
At 20–30% yields, only one in three to five units coming off the line meets specifications. The cost of failed output is absorbed into the price of each qualified HBM stack, making CXMT's usable chips expensive on a per-unit basis.
The push forward despite those economics reflects the supply environment created by U.S. export controls tightened in December 2024 on certain advanced HBM products destined for China. Those restrictions have forced Chinese chip designers to source high-bandwidth memory through unofficial channels, where prices can run several times higher than list. Against that backdrop, even costly domestic production offers a path to supply that does not depend on gray-market availability.
Performance gap and competitive context
High-bandwidth memory stacks multiple memory dies so data can move to an AI processor at extreme speed, making HBM one of the most constrained components in the AI supply chain. CXMT's HBM 3E output performing at a level closer to SK Hynix's prior-generation HBM 3 means the Chinese manufacturer is shipping a product that carries the newer label but does not match the throughput or efficiency that leading customers expect from the current generation.
SK Hynix, the dominant HBM supplier, said on October 28 that it expects memory supply shortages to persist through at least 2030 and has locked more than half its output into long-term agreements[1]. Nvidia has separately directed all three major memory makers — Samsung, SK Hynix, and Micron — to fabricate NVHBM base dies on TSMC's 3-nanometer process[3].
ANALYSIS CXMT's decision to begin production at yields that would be commercially unviable under normal market conditions reflects the premium Chinese AI hardware builders are paying for any domestically available HBM supply. The performance shortfall relative to SK Hynix's current-generation HBM 3E, however, limits the chips' utility in systems designed around that specification.