China's largest memory chipmaker, ChangXin Memory Technologies (CXMT), is preparing to begin limited production of HBM3E, an advanced type of high-bandwidth memory used in AI processors, within the next two months, according to Reuters. But the company's current manufacturing yields are only around 20% to 30%, meaning the vast majority of chips coming off its production line do not meet performance specifications.
High-bandwidth memory (HBM) works by stacking multiple memory chips vertically so they can transfer data to an AI processor at extremely high speeds. It has become one of the most constrained parts of the AI supply chain, with demand far outstripping supply. The United States tightened export controls in December 2024 on certain advanced HBM products destined for China, forcing Chinese chip designers to seek supply through unofficial channels where prices can be several times higher.
That backdrop explains why CXMT is pushing ahead despite weak yields. When only one in three to five units is usable, the cost of the defective output gets absorbed by the good chips, making each qualified HBM stack expensive. It also means that "starting production" does not automatically translate into meaningful extra supply, since the limiting factor is qualified volume at scale.
Why low yields matter
Low yields are a cost problem, not a supply fix. If CXMT has to scrap or rework most of its output, its early volumes will not materially loosen the HBM squeeze for AI hardware. That keeps leverage with today's top suppliers—SK Hynix, Samsung Electronics, and Micron—and reinforces a world where scarce memory clears at high prices or gets rationed.
Reuters reported last month that China-focused AI chip designers including Huawei, Cambricon, MetaX, and Iluvatar CoreX raised AI-chip prices by 20% to 50%, with memory costs complicating production ramps. Without a big yield jump toward the roughly 70% that Reuters cited for SK Hynix, that pressure is likely to persist.
Technically, Reuters' sources suggest CXMT's current HBM3E performance is closer to SK Hynix's older HBM3 generation, while the leaders—SK Hynix, Samsung, and Micron—still dominate global output. That gap means even if CXMT ramps production, its chips may not be competitive on performance, further limiting their impact on the market.
What it means for investors
For investors, the key takeaway is that CXMT's entry into HBM3E is unlikely to ease the AI memory shortage anytime soon. The incumbents' pricing power remains intact, which is good news for their margins but a headwind for AI hardware makers that rely on memory.
This dynamic is part of a broader trend where AI chip demand optimism has been driving market moves. In China, factory profits have surged on AI-related demand, but stocks have not always followed, as recent data shows. The memory shortage is also affecting global markets, with copper prices edging up as AI optimism returns and China's import signals tighten.
For everyday investors, this story underscores how supply chain bottlenecks in critical components like HBM can have ripple effects across the tech sector. Companies that produce the memory chips stand to benefit from high prices, while those that consume them face margin pressure. It also highlights the ongoing impact of export controls, which are reshaping global supply chains and creating both risks and opportunities.
As CXMT moves forward, investors will be watching for any signs of yield improvement. A significant jump toward the 70% range could signal a shift in the competitive landscape, but for now, the HBM market remains firmly in the hands of the established players.
