ChangXin Takes Over HBM3E: One Generation Away from Global Leader

Vision Beating AI Newsletter, The Information cited two insiders as saying that ChangXin Memory has started small-scale production of HBM3E and plans to ramp up production in 2027. HBM is high-speed memory placed next to AI chips, responsible for swiftly delivering data to the chip. HBM2 and higher products have always been a key weakness of domestic AI chips.

Domestic chip teams such as Cambricon and Alibaba Pintouge have been testing ChangXin's HBM, and if all goes well, it could be integrated into products as early as next year. HBM3E is also the next-generation memory used by NVIDIA's H200 and Blackwell. In terms of product generation, ChangXin is currently lagging behind SK Hynix, Samsung, and Micron, who are already mass-producing HBM4.

However, small-scale production is still a long way from mature mass production. ChangXin's HBM3E currently has a relatively low yield rate, and its speed and reliability are also inferior to the three major manufacturers. SK Hynix started mass production of HBM3E as early as 2024, and the three major manufacturers are now in the HBM4 phase, with HBM4E sampling underway.

This time, ChangXin has crossed the hurdle of "whether it can be done," and the next challenge is to truly improve yield rate and production capacity.