China's largest memory chipmaker, CXMT, is preparing to begin limited production of HBM3E, an advanced high-bandwidth memory used in artificial intelligence chips, even though only 20% to 30% of the chips it currently produces meet specifications. The company has sent HBM3E samples to several Chinese AI chip designers and expects to start limited output within the next two months, according to people familiar with the matter. The figures underscore both the ambition and the difficulty of China's push into the most demanding segment of the memory market.

The HBM3E plan is significant because high-bandwidth memory has become a bottleneck for AI accelerators. HBM stacks multiple memory chips vertically and connects them through high-speed links, allowing processors to move data far faster than conventional DRAM. That bandwidth is essential for training and running large language models, which repeatedly move enormous quantities of data between processors and memory. Without sufficient HBM, even a powerful AI chip can be starved of data and underperform.

Key facts

• CXMT plans to start making HBM3E despite yields of about 20% to 30%.
• The company has sent HBM3E samples to several Chinese AI chip designers.
• Limited production could begin within two months.
• CXMT's HBM3E performed closer to an older HBM3 product than to current HBM3E in testing, one person said.
• CXMT has supplied two versions: a standard design and an updated design intended to improve yields.
• The chip is likely to cost more than comparable foreign products.
• Chinese AI chipmakers are short of advanced memory, partly because of tighter US export controls introduced in December 2024.
• Huawei, Cambricon, MetaX, and Iluvatar CoreX raised AI chip prices by 20% to 50% as memory costs rose.
• SK Hynix held about 50% of global HBM revenue in the second quarter, Samsung about 33%, and Micron about 18%, according to industry estimates.
• CXMT held about 10% of global DRAM revenue, making it the fourth-largest supplier of conventional memory.
• SK Hynix says HBM4 has beaten its target yield and will pass HBM3E in shipments and revenue in the fourth quarter.
• Samsung says it has expanded HBM4 sales and shipped HBM4E samples.

Why HBM is so hard to make

High-bandwidth memory is not simply a faster version of ordinary DRAM. It is a complex stack of dies that must be thinned, aligned, bonded, and connected with thousands of through-silicon vias. Each layer must work reliably, and the finished stack must survive heat, stress, and the demands of continuous operation inside an AI system. A single weak die can ruin an entire stack, which is one reason yields are lower and production costs are higher than for conventional memory.

Yield measures the share of chips coming off a production line that meet specifications. In HBM, yield is especially important because a finished stack contains multiple chips and advanced packaging steps. A low yield means many stacks must be discarded or reworked, raising the effective cost of every usable unit. CXMT's yield of roughly 20% to 30% is far below the level achieved by the market leader, which gets around 70%, according to the people familiar with the matter. That gap matters commercially: it limits how many chips CXMT can sell and makes each good chip more expensive to produce.

The technical challenge extends beyond raw manufacturing. HBM products must be qualified by AI chip designers and system builders, a process that can take months. Customers need to know that memory will work reliably at scale, that supply will be consistent, and that performance meets the needs of their accelerators. CXMT's samples are therefore an important first step, but they do not guarantee mass adoption. The company must demonstrate that its HBM3E can meet customer requirements while improving yields and controlling costs.

The performance gap and the cost penalty

CXMT's HBM3E is also facing a performance gap. In testing, its HBM3E performed closer to SK Hynix's older HBM3 than to the current HBM3E, one person familiar with the matter said. That does not mean the product is unusable. Many AI workloads can run on older or lower-performance memory if the supply is available and the price is right. But it does mean CXMT may struggle to compete at the high end of the market, where customers are willing to pay a premium for the fastest and most reliable parts.

The company has sent customers two versions of the product: a standard design and an updated design meant to improve yields. The existence of two versions suggests CXMT is trying to balance performance and manufacturability. A yield-improved design may be easier to produce, but it could sacrifice some performance or require different customer validation. The standard design may target customers that need higher performance, while the updated version may be aimed at improving the economics of production.

Cost is another concern. CXMT's chip will likely cost more than comparable foreign products, one person said. That may seem counterintuitive for a Chinese supplier seeking to gain market share, but low yields can offset any cost advantage. If CXMT must throw away or rework most of its output, the cost per usable chip rises. Customers may still buy the product if imported alternatives are unavailable or far more expensive through unofficial channels, but they will do so with clear trade-offs in performance, reliability, and price.

China's memory squeeze

The push into HBM3E is driven by a severe supply crunch. Chinese AI chipmakers cannot get enough advanced memory to raise output, according to the people familiar with the matter. The United States tightened export controls on some advanced HBM products bound for China in December 2024, restricting access to supplies from leading foreign manufacturers. Chinese firms have since sought the memory through unofficial channels, where it can cost several times more, the people said.

That scarcity is already affecting the Chinese AI chip market. Huawei, Cambricon, MetaX, and Iluvatar CoreX raised prices for their AI chips by 20% to 50% as memory costs rose. Cambricon's latest sales missed expectations as chip supply tightened. The price increases show how a shortage of one critical component can ripple through the entire stack, raising costs for chip designers, server makers, and cloud operators. They also explain why Chinese customers may be willing to accept a lower-yield domestic HBM product: the alternative may be no product at all, or a product that costs far more through grey-market channels.

For CXMT, the opportunity is clear. Domestic demand for HBM is large and growing, and the company has an established position in conventional memory. It held about 10% of global DRAM revenue in the second quarter, making it the fourth-largest supplier of conventional memory, according to industry estimates. It said in September that it had caught up with Samsung and Micron in conventional memory. It also swung to a profit of nearly $20bn this year as memory prices soared. That financial and technical base gives CXMT resources to invest in HBM, but it does not erase the gap with established HBM leaders.

Market share and the HBM4 race

The HBM market is highly concentrated. SK Hynix held about 50% of global HBM revenue in the second quarter, Samsung had about 33%, and Micron had about 18%, according to industry estimates. These shares reflect years of investment, customer qualification, and manufacturing expertise. HBM is not a commodity that can be turned on quickly. It requires deep integration between memory design, advanced packaging, and AI chip design. The leading suppliers have spent years building those capabilities and relationships.

CXMT is entering the market at a difficult moment. Its rivals have already moved on to the next generation. SK Hynix said on Wednesday that HBM4 had beaten its target yield and would pass HBM3E in shipments and revenue in the fourth quarter. It posted record profit in the same results. Samsung said in July that it had expanded HBM4 sales and shipped HBM4E samples. In other words, while CXMT is preparing limited HBM3E production, the global leaders are already ramping HBM4 and sampling HBM4E. That puts CXMT at least a generation behind in the most advanced segment.

The generational gap does not make CXMT's effort irrelevant. China's AI chip industry needs domestic HBM supply, even if it is not the fastest or most efficient. A lower-performance HBM3E can still support inference, smaller models, or less demanding training workloads. It can also reduce dependence on imported memory and grey-market channels. But it does mean CXMT will likely compete on availability and price rather than raw performance. Customers may use its products in systems where the latest HBM generation is not required, or where export controls make foreign alternatives unavailable.

Manufacturing, packaging, and customer validation

HBM production involves more than fabricating DRAM dies. The dies must be stacked, bonded, and packaged into a complete memory device. These steps require advanced equipment and process control. Yield losses can occur at every stage: during wafer fabrication, during thinning and stacking, during bonding, and during final testing. Improving yield therefore requires coordinated progress across design, process, and packaging. A company cannot fix a low yield simply by running more wafers; it must identify where defects arise and eliminate them.

Customer validation is another hurdle. AI chip designers need to integrate HBM into their products and test it under real workloads. They need to know how the memory behaves under sustained load, how much heat it generates, and how it interacts with their packaging and cooling systems. This validation takes time. CXMT's samples are an important milestone, but they are only the beginning of a longer process. The company must work closely with customers to tune performance and reliability while simultaneously improving yields.

The two versions of CXMT's HBM3E may help in this process. A standard design can be used for customer evaluation and performance testing, while an updated design can target yield improvement. If the updated design works, CXMT can shift more production to it and improve economics. But customers may need to revalidate the updated design, which adds time. The company must manage these trade-offs carefully, especially if it wants to begin limited production within two months.

What the effort means for the AI supply chain

The broader significance of CXMT's HBM3E plan is that it shows how the AI supply chain is fragmenting. In an ideal market, AI chip designers would buy the best memory from the best supplier at the best price. Export controls, national security concerns, and domestic industrial policy are changing that calculus. Chinese AI chipmakers are being pushed toward domestic suppliers, even if those suppliers are less advanced. The result could be a more resilient supply chain in China, but also a less efficient one, at least in the near term.

For CXMT, the challenge is to turn limited production into a reliable business. That means improving yields, meeting customer specifications, and controlling costs. It also means scaling production without sacrificing quality. The company has the advantage of a large domestic market and strong demand. It has the disadvantage of starting from a lower yield and a performance gap. How quickly it closes those gaps will determine whether it becomes a serious HBM supplier or remains a niche player.

For global memory suppliers, the rise of CXMT is a reminder that the HBM market may not remain as concentrated as it is today. Even a lower-yield, higher-cost product can find customers when supply is tight and alternatives are restricted. The leading suppliers are responding by moving to HBM4 and HBM4E, extending their technology lead. That strategy may protect their margins at the high end, but it could also leave space for domestic suppliers to serve less demanding segments. The competition will play out across performance, yield, cost, and access.

The immediate focus is on CXMT's limited production timeline and customer validation. If the company can begin shipping HBM3E within two months, it will mark a notable step for China's memory industry. If yields remain at 20% to 30%, the product will likely remain expensive and constrained. Either way, the effort highlights a central reality of the AI era: advanced memory is no longer a background component. It is a strategic resource, and the race to produce it is now part of the broader competition in AI chips.

Source:TNW | Semiconductors News