CXMT to start HBM3E production despite low yields amid tight domestic market

CXMT forcing HBM3E into production at 20% yield is a sovereignty play, not a P&L decision. On the tech stack, the 12-to-16-layer TSV stacking bottleneck almost certainly sits at micro-bump alignment and bonding interface defect density. This triggers an order pulse for ASM and BESI bonding tools, while domestic players like AMEC and Tuojing gain critical volume in deposition and etch. More critically, Huawei Ascend and Cambricon now hold a "usable but expensive" domestic HBM option, breaking SK Hynix's scheduling monopoly. On compliance, 20% yield means effective die cost runs roughly 2.5x Hynix's. With EUV locked out and advanced bonding equipment restricted, the BOM is structurally inflated. But China's AI cluster memory gap is rigid demand—this is an availability question, not an ROI question. Competitive dynamics: Hynix and Samsung's optimal move isn't price cuts. It's accelerating HBM4 mass production to stretch the generational gap from "same-gen" to "cross-gen," extending their exclusive window with NVIDIA and AMD. 12-to-24-month outlook: Mirroring BOE's 2016 LCD yield curve (15% to 70%), CXMT's HBM yield should breach 60% within 18 months. The real variable then shifts to whether JCET and Tongfu Microelectronics can ramp 2.5D/3D advanced packaging in sync. The HBM market will formally bifurcate into "alliance-side" and "sovereign-side" pricing architectures, forcing a rewrite of global memory pricing models.