CXMT has been teasing about its memory capabilities and manufacturing for a while now, but recent reports show that the company has started to manufacture batches of HBM3E for testing. What exactly is CMXT doing, and what does this say about China’s memory capabilities?
CXMT Starts Production of HBM3E Samples
Recently, it was reported that the Chinese DRAM manufacturer CXMT (ChangXin Memory Technologies) is now producing small batches of HBM3E, a high-bandwidth DRAM technology designed to sit alongside AI processors and feed them data at very high rates.
The new technology, called High Bandwidth Memory 3E, uses vertically stacked DRAM dies that are connected together using high-density interconnects and then integrated into the same advanced package as the AI accelerator. This is important because the performance of HBM does not simply depend on the DRAM itself, but on the entire memory stack and the way it is integrated with the processor.
The hardware is reportedly being tested directly with processors from T-Head and Cambricon, allowing engineers to evaluate the HBM3E as part of an actual AI accelerator system rather than simply testing the memory as a standalone component.
However, CXMT has not disclosed the hardware configuration of these samples. The number of stacked DRAM layers, memory capacity, bandwidth, power consumption, yield and operating characteristics all remain unknown.
HBM3E is still a significant technology for AI accelerators in 2026, even though HBM4 is now entering commercial production. As such, the fact that CXMT is working on HBM3E rather than HBM4 does not by itself mean that its hardware is commercially uncompetitive.
Instead, one of the biggest challenges comes down to the packaging itself. The stacked HBM needs to be tightly integrated with the AI processor while maintaining reliable electrical and thermal operation under sustained high workloads. Getting the memory to function is one thing, but getting the entire package to operate reliably at high bandwidth is another.
If qualification succeeds, CXMT's HBM3E could provide Chinese AI-chip manufacturers with a domestic source of high-bandwidth accelerator memory, reducing their dependence on externally supplied HBM.
However, the hardware has not yet demonstrated production-level performance or reliability. Until CXMT publishes specifications and the memory passes customer qualification at useful yields and volumes, its capabilities cannot be directly compared with those of established HBM suppliers.
What Does this Mean for China’s Memory Capabilities?
Producing HBM3E in batches suggests that CXMT has moved beyond laboratory demonstrations and is capable of manufacturing functioning HBM hardware in a repeatable enough process to support customer testing.
This represents a significant step for China's DRAM industry, because HBM requires substantially more advanced manufacturing, interconnection and packaging capabilities than conventional DDR memory.
CXMT's progress indicates that China's domestic memory industry is no longer limited to catching up with mainstream DRAM generations, but is now attempting to enter the specialised memory market used by high-performance AI accelerators.
The ability to test HBM3E directly with Chinese AI processors is particularly important because it demonstrates the development of a domestic memory-to-accelerator supply chain, rather than CXMT developing HBM in isolation.
If the reported testing is successful, Chinese AI-chip companies could eventually source high-bandwidth memory domestically, reducing one of the major hardware dependencies on Samsung, SK Hynix and Micron.
However, producing small batches for engineering and qualification purposes is very different from achieving the yields, reliability, capacity and cost required for commercial production. The current reports therefore demonstrate capability, but not yet competitiveness.
CXMT is also entering a market where established HBM manufacturers have years of experience, large-scale production and increasingly advanced generations such as HBM4. China's achievement should therefore be viewed as evidence of technological progress rather than proof that it has caught up with the leading suppliers.
The larger significance is that China is demonstrating an increasingly complete domestic semiconductor stack around AI hardware. If CXMT can scale HBM production while Chinese companies continue developing competitive accelerators and advanced packaging, memory could become substantially less of a bottleneck for China's domestic AI hardware industry.
You may also like
