CXMT plans HBM3E production even as most of its chips fail quality checks

China’s ChangXin Memory Technologies, better known as CXMT, plans to begin making HBM3E. That’s one of the most advanced memory chips paired with AI processors, and CXMT is reportedly going ahead even though most of what comes off its line doesn’t work.

According to two people familiar with the matter who spoke to Reuters, as little as 20% of the chips CXMT produces are usable. In most industries, a 70% to 80% failure rate would end the conversation. In China’s AI chip sector right now, it apparently starts one.

A factory that mostly makes scrap, on purpose

Reuters’ sources put CXMT’s HBM3E yields at 20% to 30%. Yield is the share of chips from each silicon wafer that pass testing and can actually be sold.

The research findings say CXMT intends to start limited commercial production within the next two months. This is not a full-scale ramp. It’s a controlled launch while the company keeps working on its manufacturing problems.

The gap with the market leader is wide. SK Hynix reportedly achieves yields of around 70% on comparable products. CXMT’s own HBM3E also trails leading South Korean competitors in testing, according to Reuters.

There’s a further complication about what CXMT is actually building. The research findings indicate that its production capabilities align more closely with HBM3, the previous generation, than with the current HBM3E standard. So the chip carries a next-generation label while its performance profile may sit a step behind.

Why anyone would buy a chip with those numbers

High-bandwidth memory works by stacking DRAM chips vertically and placing them right next to an AI processor. That setup lets data move far faster than conventional memory allows. Training and running large AI models depends heavily on that speed, which is why HBM has become one of the tightest bottlenecks in AI hardware.

In December 2024, the US imposed export controls on HBM products. Those rules cut Chinese buyers off from the advanced memory produced by SK Hynix, Samsung and other foreign suppliers. Reuters reports that the result has been an acute shortage of AI memory inside China, and that CXMT’s decision reflects that shortage.

Customers are already testing the product. CXMT has shipped HBM3E samples to domestic clients, including Alibaba’s chip unit T-Head and AI chip designer Cambricon, for preliminary evaluation, according to the research findings.

Background: China’s memory champion

CXMT is China’s largest maker of dynamic random-access memory, or DRAM, the working memory found in everything from phones to data center servers. The research findings rank it as the fourth-largest DRAM producer globally by revenue.

The company runs fabrication plants in Hefei and Beijing. It is targeting a monthly DRAM capacity of approximately 350,000 wafers by the end of 2026, according to the research findings.

What this means for the AI memory market

For the global memory industry, the immediate competitive threat looks small. A 20% to 30% yield means CXMT’s HBM3E will be costly to produce and limited in volume. Its chips also reportedly lag South Korean rivals in testing. SK Hynix and Samsung are unlikely to lose sleep over the first batch.

Export controls have effectively given CXMT a protected home market where buyers have few other options. That gives the company paying customers and real-world feedback while it works on its yields.

Customer validation is the second signal. Results from testing at T-Head and Cambricon will show whether CXMT’s memory performs well enough for production AI hardware, or only well enough to fill a shortage.