CXMT to mass produce HBM3E memory chips in late Q4 2026 despite abysmal yields

Just two months ago, premier Chinese DRAM manufacturer CXMT was entering risk production for HBM3E chips. Even if this technology is two generations behind the current cutting edge HBM4E chips produced by SK Hynix and Samsung, things were looking up for China in the effort to reduce reliance on Western technologies. Now, sources close to Reuters report that CXMT could be ready to enter HBM3E mass production by the end of this year, but the situation is far from actually consolidating China’s domestic chips supply and fueling the AI race domination plans.

Latest reports point out that the CXMT’s HBM3E production yields sit between 20-30%, while SK Hynix achieves ~70% for the same technology. Running commercial fabrication lines at a sub-30% yield means the vast majority of finished silicon is discarded or reworked, driving unit costs to steep premiums. Yet for China’s domestic computing sector, economic efficiency has taken a back seat to sheer availability. However, China still considers this a viable outcome over all the problems caused by the ongoing U.S. export restrictions that left domestic AI hyperscalers including Huawei, Cambricon, MetaX, and Iluvatar CoreX struggling to secure sufficient chips.

With foreign supplies constrained and unofficial channels charging exorbitant markups, memory bottlenecks have forced Chinese domestic AI accelerator vendors to raise current and next-generation chip prices by 20% to 50%. Sourcing CXMT’s domestic HBM3E, even at elevated prices, provides a viable, sanction-insulated supply line to keep local AI accelerator production going.

Select Chinese AI chip designers have already received two variants for the HBM3E chips: a baseline design and an updated revision engineered specifically to salvage yields. Initial tests reveal that CXMT’s HBM3E performs closer to SK Hynix’s previous-generation HBM3 rather than competing on equal footing with current South Korean HBM3E variants. This is most likely caused by the trade-restricted access to cutting-edge fabrication and packaging tools, such as dense through-silicon via (TSV) interconnects and precise vertical stacking. Maintaining signal integrity and thermal dissipation thus remains a major hurdle for CXMT.

CXMT’s compromise to eat massive scrap rates demonstrates that China’s domestic AI hardware ecosystem is prioritizing functional self-reliance over fabrication economics, even as it operates a full architectural generation behind the global frontier.