CXMT가 Softonic 보도에 따르면 HBM3E의 제한적 양산까지 두 달 남았다고 한다.

Reuters reports that ChangXin Memory Technologies (CXMT), China's largest DRAM maker, plans to begin limited HBM3E production within the next two months. If that happens, China would have a domestic source for the kind of AI memory it needs in growing amounts.

The main obstacle is yield. Reuters puts it at 20% to 30%, which means 70% to 80% of chips fail. Under normal conditions, that is a difficult basis for large-scale production.

HBM3E matters because it sits close to advanced processors and moves huge amounts of data for the accelerators that train and run large models in data centers. Supply is already tight. According to one earlier estimate, a single major U.S. server buildout could require hundreds of thousands of additional DRAM wafers a month, or roughly 40% of global output.

U.S. export controls are also part of the story. The rules, introduced in October 2022 and tightened later, have pushed advanced memory higher on China's priority list by limiting access to top-end chips and some of the manufacturing tools needed to make them.

Analysts have described the effort as "loss-making domestic substitution." Reuters says Beijing is backing it through Made in China 2025, dual circulation, and a reported $47.5 billion chip fund.

For those following AI hardware, the situation is worth watching. CXMT still trails SK Hynix and Samsung, and Reuters says its performance is closer to older HBM3. SK Hynix is reaching yields near 70%, and Samsung is already moving into HBM4 and HBM4E. The resource cost for each working chip also remains high.

The full report is available from Reuters.