ChangXin Memory Technologies has started producing HBM3E in small quantities, according to a report from The Information citing two insiders. It’s the first time China’s top memory maker has manufactured the high-bandwidth memory found in many current AI processors, the stacked-chip technology that sits next to the processor and handles the massive data throughput required to train and run large models.
Alibaba’s T-Head and Cambricon are already testing the chips, with plans to integrate them into products starting in 2027. That timeline matters more than the announcement itself. US export restrictions have cut off Chinese chip designers from advanced HBM, so a domestic source—even one that’s a generation behind—removes a bottleneck that has nothing to do with logic design and everything to do with supply chain politics.
But the gap is real. Samsung, SK Hynix, and Micron are already mass-producing HBM4, putting CXMT roughly three to five years behind on technology. The bigger problem is yields. SemiAnalysis pegged them at around 25 percent as of June, which means three out of every four chips coming off the line are scrap. That’s not a manufacturing hiccup; that’s the fundamental challenge of stacking DRAM layers with the precision HBM demands. CXMT raised $8.6 billion in its Shanghai IPO, and notably, none of that money is earmarked for HBM in the prospectus.
For Cambricon and Alibaba, the existence of domestic HBM3E is a hedge. They can design chips knowing memory won’t be the thing that kills the project. But with those yields and a 2027 integration timeline, nobody in Beijing is closing the gap with Nvidia’s supply chain anytime soon. The question is whether HBM3E is good enough for China’s domestic AI ambitions—and at 25 percent yields, that question is still very much open.
