Hardware

CXMT Debuts China's First HBM3E AI Memory Chips

Chinese memory manufacturer ChangXin Memory Technologies has begun small-scale production of HBM3E chips, a crucial step toward domestic self-sufficiency amid strict US export controls.

The Decoder1 day agoHardware
Image: The Decoder

ChangXin Memory Technologies, China's leading memory manufacturer, has initiated small-scale production of its first HBM3E chips. This high-bandwidth memory is essential for training and running large-scale artificial intelligence models, as it sits directly alongside processors to accelerate data transfer. The development represents a significant milestone for China's domestic semiconductor industry as it navigates tightening trade restrictions.

Despite this breakthrough, CXMT remains technically positioned behind global market leaders like Samsung, SK Hynix, and Micron, which are already mass-producing next-generation HBM4 memory. Industry analysts estimate that CXMT lags three to five years behind these competitors. Furthermore, the Chinese manufacturer faces severe manufacturing hurdles, with market research firm SemiAnalysis estimating CXMT's production yields at just 25 percent as of June.

Domestic chip designers are already preparing to integrate the new memory. Alibaba's T-Head division and Cambricon are currently testing CXMT's HBM3E chips, with plans to incorporate them into commercial products starting in 2027. This timeline offers a potential lifeline for Chinese AI hardware developers who face severe restrictions on purchasing advanced American-designed or foreign-manufactured HBM components due to United States export controls.

Financially, CXMT recently raised $8.6 billion through its Shanghai initial public offering, though its prospectus did not explicitly earmark any of these funds for high-bandwidth memory development. For AI practitioners and hardware engineers in China, the emergence of a domestic HBM3E supplier, even with low initial yields, signals a path toward building viable AI accelerators locally. While performance may initially lag behind Western equivalents, it establishes a baseline of technological independence.

This is our own summary of reporting by The Decoder

More in Hardware