Chinese memory maker ChangXin Memory Technologies has begun producing HBM3E in small quantities, according to The Information. High-bandwidth memory stacks chips close to a processor so large AI models can move data quickly during training and inference, making it a critical component of modern accelerators.

Alibaba’s T-Head and chip designer Cambricon are reportedly testing CXMT’s parts and plan to use them in products beginning in 2027. A domestic supply could reduce one obstacle for Chinese AI-chip makers, which face US export restrictions on advanced memory.

The progress does not put CXMT level with global leaders. Samsung, SK Hynix and Micron are already mass-producing the newer HBM4 generation. The report estimates that CXMT remains three to five years behind technically, while SemiAnalysis placed its production yield at roughly 25 percent in June. Low yield means many manufactured chips cannot be sold, pushing up the cost of usable parts.

CXMT raised $8.6 billion in its Shanghai initial public offering, although the prospectus did not allocate that money specifically to high-bandwidth memory. The first HBM3E output is therefore an important manufacturing step, but volume, reliability and yield—not merely producing working samples—will determine whether it materially expands China’s AI-compute supply.