CXMT Begins Small-Batch HBM3E Production: What It Means for China’s AI Chips

ChangXin Memory Technologies (CXMT) has reportedly begun producing small quantities of HBM3E, an advanced high-bandwidth memory used in AI chipsets. Alibaba’s T-Head and Cambricon are testing the memory with their processors, and CXMT plans to expand production in 2027. 15

That makes the development strategically important for China’s AI supply chain. It does not yet make CXMT a commercially established rival to SK Hynix, Samsung or Micron. The key question is whether the company can turn early production into qualified, reliable memory delivered at meaningful volume.

What HBM3E is used for

High-bandwidth memory, or HBM, is a form of stacked DRAM designed to sit close to an AI processor or accelerator. Its role is to move data quickly between memory and the processor, a requirement for demanding workloads such as AI model training and inference. HBM3E is a newer generation of that technology and is described in reporting as an advanced memory used in AI chipsets. 1

Unlike ordinary memory production, HBM commercialization depends on more than manufacturing functional DRAM dies. The memory must work with an accelerator’s memory controller, advanced packaging, thermal design, firmware and reliability requirements. That makes customer qualification a crucial milestone.

Who is testing CXMT’s HBM3E?

The reported test customers are:

• T-Head, Alibaba’s chip-design subsidiary
• Cambricon Technologies, a Chinese AI-chip designer

Both are reportedly testing CXMT’s HBM3E with their processors. If verification and qualification proceed successfully, the memory could appear in domestic AI-chip products from 2027. That is a reported target, not a confirmed commercial launch date. 15

The distinction matters. Testing demonstrates that a customer is evaluating the technology; it does not prove that the memory has passed all production, reliability or supply requirements.

Why small-batch production is only the beginning

Initial output shows that CXMT has made technical progress, but pilot production is not the same as a competitive HBM business. Before an accelerator maker can depend on a memory supplier, it typically needs confidence in several areas:

• consistent manufacturing yields;
• bandwidth and power performance;
• thermal behavior inside the finished package;
• long-term reliability;
• compatibility with the customer’s processor and software stack; and
• the supplier’s ability to deliver repeat orders.

HBM also requires sophisticated assembly and stacking. A company can produce working samples yet struggle to manufacture enough economically usable parts. For CXMT, customer qualification and repeat volume orders will therefore be more meaningful indicators than the announcement of limited production itself.

How far is CXMT behind the established HBM suppliers?

The HBM market is dominated by SK Hynix, Samsung and Micron. CXMT has become a major DRAM producer—Reuters described it as the world’s No. 4 DRAM maker—but it remains behind the leading companies in advanced memory, particularly HBM.

Industry reporting and commentary have characterized CXMT’s HBM technology gap as several years, while established suppliers are already advancing toward newer HBM generations. That creates a moving target: even if CXMT successfully scales HBM3E, its competitors will continue developing their next products.

CXMT’s potential advantage is not immediate global parity. It is the possibility of giving Chinese AI-chip designers a domestic memory option at a time when export controls limit access to some advanced foreign semiconductor technologies. 1

The manufacturing challenge: yields and lithography

CXMT faces reported low yields in advanced HBM production. Some industry estimates have put early overall HBM yields at roughly 25%, although such estimates are not the same as company-reported results.

The company also lacks access to ASML’s extreme-ultraviolet, or EUV, lithography tools because of U.S.-led export restrictions. Reporting says CXMT must rely more heavily on older deep-ultraviolet, or DUV, equipment and techniques such as multi-patterning. Those methods can increase process complexity, cost and yield pressure. 12

CXMT itself has acknowledged gaps versus Samsung, SK Hynix and Micron in technology accumulation, production capacity and product structure. These constraints do not make HBM3E impossible, but they raise the difficulty of producing it competitively and at scale.

What CXMT’s IPO and earnings can—and cannot—fund

CXMT raised 57.92 billion yuan, approximately $8.6 billion, in its Shanghai STAR Market IPO. Its shares rose roughly 466% on their first trading day. The company said the IPO proceeds would be used mainly for memory-wafer mass production. 28

However, reporting on the IPO prospectus noted that the named projects focused on existing DRAM capacity, technology upgrades and wafer-line improvements rather than a dedicated HBM project. 7 The IPO gives CXMT substantial financial capacity, but it should not be treated as proof that all of the money is already committed to scaling HBM3E.

CXMT’s first-half 2026 results were also unusually strong: revenue reached 150.3 billion yuan and net profit reached 77.6 billion yuan, compared with a loss in the same period a year earlier. 17

Those results reflect a broader DRAM price surge and AI-driven demand and supply constraints across the memory market. They should not be interpreted as evidence that HBM3E is already generating significant revenue for CXMT. 17

Can CXMT become a viable fourth HBM competitor?

In China, plausibly. Globally, not yet on the available evidence.

CXMT has several ingredients for a strategically important domestic HBM business: a large DRAM manufacturing base, strong market access in China, customers testing its memory, new capital and continued demand from AI hardware. Its HBM3E effort could help Chinese accelerator companies reduce reliance on overseas suppliers.

But commercial viability depends on execution. CXMT still needs to improve yields, complete customer qualification, scale packaging and stacking, deliver consistent performance and keep pace with newer HBM products from the established suppliers.

The most defensible conclusion is that CXMT’s HBM3E production is an early commercialization milestone—not yet a disruption of the global HBM oligopoly. A successful 2027 ramp could make CXMT a lower-volume but strategically important supplier for China’s domestic AI-chip ecosystem. Whether it becomes a peer of SK Hynix, Samsung or Micron will depend on production quality and repeatable volume, not simply on the existence of early chips.