Reuters reported on October 28, 2026, that Chinese memory maker ChangXin Memory Technologies (CXMT) planned to begin limited production of HBM3E within the following two months. HBM, or high-bandwidth memory, stacks memory chips vertically to speed data exchange with AI processors. The reported plan comes with a low estimated manufacturing yield: 20% to 30% of CXMT’s HBM3E chips would meet required specifications.
What the yield estimates say
Reuters attributed the yield estimates to unnamed sources. They put CXMT’s HBM3E yield at 20% to 30%, compared with close to 70% for SK Hynix. Yield measures the share of manufactured chips that meet specifications. These are reported estimates, not a standardized head-to-head result.
The yield figure describes manufacturing output, not memory speed. Separately, one source said CXMT’s HBM3E performed closer in testing to SK Hynix’s earlier HBM3 than to SK Hynix’s HBM3E. That comparison was qualitative.
Samples and designs
CXMT reportedly sent HBM3E samples to several Chinese AI-chip designers in two versions: a conventional design and an updated design intended to improve manufacturing yield. The report did not name the recipients. It also said CXMT began internal HBM3E development in early 2024.
Why CXMT is pursuing production
Reuters’ sources linked the plan to Chinese AI-chip companies’ difficulty securing enough advanced memory. The report also described US export controls as a constraint on access to certain advanced HBM products and semiconductor-manufacturing equipment. The account places CXMT’s plan in that supply context without quantifying its costs or financial returns.
Online reactions offer competing interpretations, not a measure of CXMT’s business prospects. Some Reddit commenters argued that high memory prices or the value of refining a manufacturing process could make low yields tolerable; others questioned the economics or the report’s framing. Yield alone does not determine whether a project is profitable.
