ChangXin Memory Technologies (CXMT) has achieved only 25% production yields on its HBM3E memory, according to a TechPowerUp report citing industry sources. The figure signals a wide gap between China's leading DRAM manufacturer and the established HBM suppliers that serve the AI accelerator market.

HBM3E is the current high-end tier of high-bandwidth memory, the stacked DRAM used on datacenter GPUs and AI accelerators. SK Hynix, Samsung and Micron ship HBM3E in volume today. A 25% yield means three of every four die stacks fail final test, making the economics of mass production unworkable at current levels.

Why does a 25% yield matter?

High-bandwidth memory stacks multiple DRAM dies on top of a base logic die and connects them through vertical vias. Every added layer multiplies the chance of a defect. At 25% yield, CXMT would discard most of its finished wafers — a cost structure that cannot compete with suppliers reporting mature yields on comparable products.

The report does not specify which HBM3E stack height or capacity CXMT is attempting, or when the company expects yields to improve. It also does not state current output volumes.

Where does CXMT stand in HBM?

CXMT, based in Hefei, is China's most advanced DRAM maker and has concentrated on mainstream DDR5 and LPDDR5 products. Reports over the past year indicated the company was developing HBM capabilities to enter the AI memory segment. The 25% figure suggests the program remains in an early, pre-qualification stage rather than near commercial shipment.

For AI GPU vendors, HBM supply is effectively limited to the Korean and US duopoly-plus-one of SK Hynix, Samsung and Micron. Chinese accelerator designers, constrained by export controls on US memory technology, have no domestic high-bandwidth memory source qualified at HBM3E level.

What are the market implications?

If CXMT cannot raise yields substantially, the timeline for a domestic Chinese HBM supply chain extends further out. Achieving HBM3E-grade production typically requires not only process maturity but also TSV packaging capacity and stack-level test capability — infrastructure the incumbent suppliers built over several product generations.

The 25% figure, if accurate, places CXMT at roughly a generation behind in manufacturing readiness. Closing that gap would require both process fixes and packaging scale-up before the company could bid for meaningful volumes.

No statement from CXMT, and no customer qualification data, accompanied the report. The company does not publicly disclose yield figures.