China's CXMT (ChangXin Memory Technologies), the country's largest DRAM maker, plans to begin limited production of HBM3E memory in the next two months, even with yields (usable chip rate) of only 20% to 30%. The exclusive Reuters report highlights the pressure for self-sufficiency in AI chips amid US export restrictions.
What happened?
According to people familiar with the matter, CXMT has already sent HBM3E samples to several Chinese AI chip designers. The company expects to start small-batch production within two months. Two versions were provided: a standard design and a revised one with changes aimed at improving yield.
Current yields of 20-30% contrast sharply with around 70% at leader SK hynix. Tests show performance closer to the previous HBM3 than to SK hynix's HBM3E. The low usable rate is expected to make the product more expensive to manufacture.
Why does this matter?
HBM (High Bandwidth Memory) is essential for AI processors, enabling rapid high-volume data transfer. With US restrictions limiting China's access to advanced HBM, local companies face acute shortages. CXMT is taking the risk of inefficient production to supply domestic demand and reduce reliance on expensive imports (often through unofficial channels).
This fits Beijing's strategy to boost the domestic semiconductor industry. Companies such as Huawei and Cambricon have already raised AI chip prices by 20-50% due to memory costs.
What changes in practice?
- Products using CXMT's HBM3E could reach the market as early as 2027 if development proceeds as planned.
- Higher costs may further pressure Chinese AI hardware prices.
- The move strengthens the push for technological autonomy, even with short-term efficiency disadvantages.
- Competition with SK hynix and Samsung remains unequal in quality and volume.
CXMT has been discreet about progress. At an investor briefing in September, management avoided specific HBM details and reaffirmed the existing roadmap.
By GeekikiBot
