China's top memory maker ChangXin Memory Technologies, known as CXMT, has reached a new milestone in advanced AI hardware: it is producing HBM3E in small quantities for the first time. The development matters because HBM3E is a fast memory type used in many current AI processors, and high bandwidth memory is a key part of the systems that train and run large AI models.
The advance does not put CXMT at the front of the global market. According to the source article, the company remains a generation behind Samsung, SK Hynix, and Micron, which are mass-producing HBM4. Still, small-batch HBM3E production gives China's AI chip ecosystem a domestic component to test as it works around limits on access to advanced HBM chips.
Why HBM3E matters for AI chips
HBM stands for High Bandwidth Memory. The basic design stacks memory chips and places them right next to the processor. That close physical placement is important because AI processors need to move large amounts of data quickly while training and running large AI models.
In plain terms, a powerful AI processor is not useful on its own if it cannot get data fast enough. HBM helps reduce that bottleneck by putting fast memory near the processor rather than treating memory as a distant supporting part. That is why HBM appears in many current AI processors.
CXMT's move into HBM3E therefore touches a central issue in AI hardware: not just whether China can design AI chips, but whether those chips can be paired with the advanced memory they need. The source article frames this as one hurdle for China's own AI chips, especially because US export rules limit the purchase of advanced HBM chips.
Who is testing CXMT's memory
The Information, citing two insiders, reported that CXMT is producing HBM3E for the first time in small quantities. The article says Alibaba's T-Head and Cambricon are testing the memory and plan to use it in products starting in 2027.
That testing step is important because it is where a memory product moves from technical achievement toward practical use. Producing small quantities shows that CXMT has a working path to HBM3E. Testing by chip developers shows that the memory is being evaluated for future products rather than existing only as a lab milestone.
The timeline also matters. The products mentioned in the source are planned to start in 2027, which means this is not described as an immediate broad rollout. It is a signal about the direction of China's AI hardware supply chain, not proof that CXMT has already matched the leading memory suppliers at scale.
The gap with global leaders remains
CXMT's progress comes with clear limits. The source article says the company is a generation behind Samsung, SK Hynix, and Micron, which are mass-producing HBM4. That comparison is significant because HBM3E production in small quantities is not the same as mass production of the next generation.
The Information also reports that CXMT is three to five years behind technically and struggles with low yields. SemiAnalysis estimated those yields at around 25 percent in June. Low yield means production is less efficient, which can make it harder to supply customers consistently and compete with companies already manufacturing at greater maturity.
The result is a mixed picture. CXMT has taken a meaningful step by producing HBM3E, but the company still faces technical and manufacturing challenges. The leading suppliers are already mass-producing HBM4, while CXMT is beginning with small quantities of HBM3E.
What this changes for China's AI supply chain
The main implication is strategic. If Chinese AI chipmakers can use domestically produced HBM3E, they may have one fewer dependency in a supply chain shaped by export restrictions. The source article says this could lower one hurdle for China's own AI chips because US export rules limit the purchase of advanced HBM chips.
That does not mean the hurdle disappears. The source points to several constraints at the same time:
- CXMT is producing HBM3E only in small quantities.
- Samsung, SK Hynix, and Micron are already mass-producing HBM4.
- CXMT is reported to be three to five years behind technically.
- Yield remains a challenge, with SemiAnalysis estimating around 25 percent in June.
- CXMT's Shanghai IPO raised $8.6 billion, but none of that is earmarked for HBM according to the prospectus.
Those details suggest that CXMT's first HBM3E chips are best understood as an early domestic foothold rather than a completed catch-up. The company has entered a critical part of the AI processor stack, and Chinese chip developers are testing the result. But the distance to the global leaders remains visible in both generation and manufacturing scale.
For China's AI ambitions, the significance is not only the memory chip itself. It is the possibility of pairing local AI processors with local high bandwidth memory in future products. Whether that becomes a practical advantage will depend on testing, yields, and the ability to move beyond small quantities.