China’s CXMT Makes Breakthrough in Advanced Memory Chips

China’s leading memory-chip maker, ChangXin Memory Technologies, has begun producing advanced high-bandwidth memory in small quantities, a milestone that could ease a key constraint on the country’s development of homegrown AI processors.

CXMT is making HBM3E, an advanced type of high-bandwidth memory used in many of today’s leading AI processors, according to two people familiar with the matter. That means its product is now only one generation behind the memory entering mass production at the world’s leading manufacturers, Samsung, SK Hynix and Micron Technologies. CXMT plans to expand production in 2027, the two people added.

Several Chinese chip designers are testing the memory with their processors, including Alibaba Group’s T-Head and Beijing-based Cambricon Technologies. If all goes to plan, these companies will use CXMT’s memory in products as early as next year, according to the two people. Their orders could give CXMT the production experience and customer feedback it needs to make HBM on a larger scale.

CXMT and Cambricon didn’t respond to requests for comment. Alibaba’s T-Head didn’t have a comment.

CXMT’s progress with HBM is a major advance for China’s chip industry, coming as demand for high-bandwidth memory has never been more intense, thanks to the AI boom. U.S. export restrictions limit Chinese companies from buying the most advanced HBM from Western companies, one of the major factors holding back Chinese AI chip designers.

News of CXMT’s advances comes just a few weeks after the company went public in Shanghai, raising $8.6 billion to expand production and improve its technology. The initial public offering was a big hit with Chinese investors, with CXMT shares soaring 472% in the debut.

The company said on Friday that its first-half revenue this year increased almost tenfold to 150.3 billion yuan ($22.3 billion), while it reported a net income of 77.6 billion yuan ($11.5 billion), reversing a loss year earlier. CXMT did not mention its progress in HBM development in the earnings release.

HBM is made by stacking memory chips and placing them close to a processor, allowing data to move rapidly between the two components, which is essential for training and running large AI models. The AI boom, which has made advanced AI chips highly sought after, has also strained demand for HBM. The three leading manufacturers are racing to fill orders from Nvidia, Google and other large developers.

Washington restricted shipments of advanced HBM to China in 2024, and Chinese companies need U.S. approval to buy less advanced memory.

CXMT’s entry into the market could therefore have a big impact on local tech firms. But much remains uncertain, as its ability to manufacture the chips is also constrained by U.S. restrictions that block Chinese companies from buying some of the most sophisticated foreign equipment used by leading memory manufacturers, the two people said.

Founded with local government backing in 2016, CXMT has become a centerpiece of the Chinese government’s plan to make China more self-reliant in advanced chips. It has grown into the world’s fourth-largest producer of DRAM, the standard memory used in computers and servers.

Advanced HBM is harder to make than conventional memory such as DRAM. HBM requires manufacturing, stacking and packaging multiple DRAM chips together, and each step creates another chance of failure. Products with more layers get progressively harder to produce.

Despite the progress CXMT has made in producing HBM3E, it is three to five years behind the leading memory chipmakers in HBM technology. SK Hynix, for instance, Nvidia’s main supplier of memory, started mass producing HBM3E in 2024. CXMT continues to struggle with low yields, one of the two people said. Yield refers to the proportion of products leaving a production line that work as intended.

Cambricon and T-Head have reason to use CXMT’s memory anyway, even though it is less reliable and slower than what the three leaders sell. The companies are working on how CXMT’s HBM can function with their processors, the people said.

The state absorbs much of the cost of going first. National and regional authorities offer tax benefits, subsidies and other incentives to companies that develop, buy or operate domestic semiconductor technology.

The version of HBM CXMT is now producing, HBM3E, works with leading AI processors including Nvidia’s Blackwell graphics processing units and Google’s latest tensor processing units. Numbered versions such as HBM3 and HBM4 mark major generations, while the “E” denotes a faster version of the preceding generation. Nvidia’s H100 uses HBM3, its H200 and Blackwell processors use HBM3E, and the newer Rubin uses HBM4.

Meanwhile, SK Hynix, Samsung and Micron are mass-producing HBM4 and have begun supplying samples of HBM4E, a faster version intended for processors, expected next year.