ChangXin Memory Technologies, China’s most prominent domestic DRAM manufacturer, is making its most aggressive play yet in the AI hardware supply chain. CXMT has begun producing high-bandwidth memory chips, starting with HBM3 samples shipped to major clients including Huawei, with eyes on the even more advanced HBM3E standard.
The move matters because high-bandwidth memory is the bottleneck component in AI accelerators. Every major AI chip, from Nvidia’s H100 to AMD’s Instinct series, depends on HBM stacks to feed data to processors fast enough to keep up with large model training. Until now, that market has been dominated by three players: SK hynix, Samsung, and Micron.
What CXMT has actually built
CXMT began shipping HBM3 samples to clients, including Huawei, during the second half of 2025. The chips use a 16 nm process and 8-high stacks, meaning eight individual DRAM dies layered vertically and connected with through-silicon vias.
Mass production of HBM3 is targeted for the end of 2026, though analysts have flagged slippage risks. The more advanced HBM3E variant, which offers higher bandwidth and improved power efficiency, is anticipated for 2027.
The technology gap is real. CXMT’s current capabilities lag behind the industry leaders by roughly one to two generations. SK hynix is already shipping HBM3E in volume and developing HBM4. Samsung and Micron are in similar positions.
Early yield rates tell the story of that gap. CXMT’s HBM3 production is reportedly achieving yields of around 25% for its 8-high stacks. For context, mature HBM production lines at established vendors typically run at significantly higher yields. A 25% yield means three out of every four chip stacks don’t make it through quality testing, which is expensive and slow.
The IPO that broke records
CXMT’s July 2026 debut on the Shanghai STAR Market was nothing short of spectacular. The company’s IPO raised approximately 57.9 billion yuan, roughly $8.6 billion. On its first day of trading, the stock surged about 466%, pushing the company’s valuation to approximately 3.3 trillion yuan.
For the first half of 2026, CXMT reported revenue of 150.3 billion yuan, an 874% year-over-year increase. Net profit hit 77.6 billion yuan.
Those gains are largely driven by strong DRAM pricing and AI-related supply constraints rather than HBM revenue specifically. Traditional DRAM, the kind that goes into servers, PCs, and phones, still accounts for the bulk of CXMT’s output. The company’s DRAM capacity is expected to reach roughly 300,000 to 350,000 wafers per month by the end of 2026.
The EUV problem
CXMT’s biggest structural challenge isn’t ambition or capital. It’s access to equipment.
The company lacks access to extreme ultraviolet lithography machines, the cutting-edge tools made exclusively by Dutch firm ASML that enable the smallest, most efficient chip designs. US-led export controls have effectively blocked Chinese chipmakers from acquiring EUV systems, forcing companies like CXMT to rely on older deep ultraviolet (DUV) lithography.
DUV can still produce functional chips, but it requires more processing steps, results in larger feature sizes, and generally caps how competitive the end product can be against chips made with EUV. For HBM specifically, this constraint affects both the density and power efficiency of the memory stacks CXMT can produce.
What this means for the AI memory market
Where CXMT could matter is in the domestic Chinese market. With Huawei developing its own AI accelerators, including the Ascend series, there’s a captive customer base that needs HBM and has limited options for sourcing it from Western-aligned suppliers. Even at 25% yields and older process nodes, CXMT’s HBM3 could fill a critical gap for Chinese AI hardware makers who are building systems under export restrictions.
If CXMT does achieve HBM3 mass production by late 2026 or early 2027, even at modest volumes, it could ease pricing pressure in China’s domestic market while having limited direct impact on global HBM pricing. The more interesting scenario comes later: if CXMT can close the technology gap and push into HBM3E at meaningful scale by 2027, it becomes a genuine fourth competitor in a market that has operated as a three-player oligopoly among SK hynix, Samsung, and Micron.