The yield rate is less than 30%, and Changxin Memory Technologies is vigorously promoting HBM3E.

In the era of explosive growth of AI computing power, HBM (High Bandwidth Memory) is exactly the core "high-quality feed" that powers large language models. In this extremely cutting-edge field that is absolutely dominated by South Korean giants, the breakthrough made by local Chinese manufacturers is presenting an extremely harsh yet solemn and stirring picture.

According to the latest industry news, CXMT (Changxin Memory Technologies), the largest DRAM manufacturer in China, has been fully prepared for mass production of the fifth-generation high-bandwidth memory HBM3E. However, accompanying this ambitious goal is a glaring data point: the yield rate is only 20% to 30%.

For comparison, the yield rate of SK Hynix, the global HBM leader, is around 70%. CXMT's yield rate is less than one-third of its rival's. Moreover, in terms of performance, CXMT's HBM3E is closer to SK Hynix's previous-generation HBM3, rather than the current top-tier HBM3E.

Under normal business logic, a company with such extremely low yield, lagging performance and high production costs would not even be qualified to enter the market. But CXMT has not only entered the market, but also delivered samples to leading domestic AI chip companies including Huawei, Cambricon, MetaX and Iluvatar CoreX, and plans to launch small-scale production within two months.

This is by no means a simple technological breakthrough, but a "bloody breakthrough" forced by US export controls.

Cut Lifeline and the Cost of "Informal Channels"

To understand why CXMT's yield rate is so low, we must look at the macro backdrop of the China-US tech game.

HBM is not ordinary computer memory. It is a system that vertically stacks multiple DRAM chips and connects them together through extremely complex Through Silicon Via (TSV) and advanced packaging technologies. It requires extremely high manufacturing precision, heat dissipation control and yield management. In the era of large AI models, computing chips (such as GPU/NPU) run at extremely high speeds. If the memory bandwidth cannot keep up, the computing power will be severely idled. HBM is the only solution to break the "memory wall" bottleneck.

In December 2024, the United States further tightened its export restrictions on advanced HBM to China. This ban directly cut off the "lifeline" of Chinese AI manufacturers that once relied on imports. More fatally, the United States not only restricted the export of HBM chips themselves, but also restricted the export of related manufacturing equipment. Without the most advanced Extreme Ultraviolet (EUV) lithography machines, nor top-tier hybrid bonding equipment, CXMT has to complete extremely complex stacking processes on a production line with incomplete equipment.

According to industry sources, after the ban took effect, some Chinese manufacturers had to turn to "informal channels" to purchase HBM from overseas black markets, whose price is said to be several times the normal level, and the supply is extremely unstable. The shadow of possible supply cutoff at any time has made the entire Chinese AI industry extremely anxious.

This has led to an extremely brutal situation: memory shortage has directly pushed up the manufacturing cost of Chinese AI chips. Enterprises including Huawei, Cambricon, MetaX and Iluvatar CoreX have been forced to raise the prices of their existing and next-generation AI chips by 20% to 50%, and their capacity expansion plans have also been severely disrupted. On the cliff edge of "supply cutoff", CXMT has become the only "last hope of the whole village" in China capable of filling this huge gap. Even if the yield rate is only 20%, even if the performance lags behind by one generation, as long as the product can be manufactured, Chinese AI chips will still have a glimmer of life.

Breakthrough First, Optimization Later: Harsh Engineering Compromise

Facing the extremely low yield rate and equipment blockade, CXMT has adopted an extremely pragmatic engineering strategy.

In the normal evolution of semiconductor technology, a chip company usually waits until the yield rate climbs to over 80% to ensure profitability before delivering large batches of samples to customers. But CXMT cannot afford to wait. It has adopted a batch promotion idea of "making breakthroughs first, then fixing problems".

According to people familiar with the matter, CXMT has provided two sets of solutions to its customers: one is the standard design product, and the other is the improved version. The improved version allows customers to make certain design changes (such as adjusting PCB layout, modifying underlying firmware) to match CXMT's chips, so as to improve the overall system yield.

This practice is not uncommon in the semiconductor industry. It is essentially a "collaborative compromise" between upstream and downstream of the supply chain when facing technical bottlenecks. With relatively immature local HBM chips, and cooperating with the underlying design adjustments of domestic AI chip manufacturers, CXMT first builds a workable system, and exposes and solves problems in actual operations.

But the cost of compromise is high. The extremely low yield rate means that the manufacturing and depreciation cost allocated to each qualified HBM is extremely high. Sources say that the pricing of CXMT's HBM3E will be higher than similar overseas products. This is a phenomenon that extremely violates common business sense: your product has worse performance and lower yield than others, but its price is higher.

This is the bitter fruit of "loss-making domestic substitution" that China's technology industry has to swallow under US sanctions. Without scale effect, without the most advanced equipment, it has to rely on capital and manpower to force a bloody path. For large domestic manufacturers such as Huawei, what they buy is not just memory, but also the right to survive of China's AI industrial chain. As long as CXMT can supply goods, even if the price is a little higher and the performance is a little worse, it is better than having their supply cut off by others at any time.

Unchanged Landscape: The Reality of One Generation Gap

Although CXMT has done its utmost, in the global HBM competitive arena, the competition pattern is still difficult to subvert in the short term.

According to data from Counterpoint Research, the global HBM revenue share in the second quarter of this year is: SK Hynix 50%, Samsung Electronics 33%, Micron Technology 18%. These three giants almost monopolize the entire market. Although CXMT ranks fourth in the overall DRAM market with a 10% share, it is still a very marginal new player in the high-end HBM field.

What is more brutal is the technological generation gap. While CXMT is still struggling to improve the yield of HBM3E (which is benchmarked against the previous-generation HBM3), Samsung has expanded the production capacity of HBM4, and even delivered samples of HBM4E. HBM4 has a crushing advantage over HBM3E in bandwidth, stacking layers (such as 12 layers or even 16 layers) and energy efficiency ratio.

For South Korea's SK Hynix and Samsung, CXMT's breakthrough is not a direct threat in the short and medium term. They are relying on the extremely high premium of HBM4 to frantically harvest orders from giants such as NVIDIA, and consolidate their technological hegemony. CXMT's production capacity is too small, and its yield rate is too low, which is not enough to trigger a price war in the international market. There is a clear consensus in the industry: CXMT is at least one full generation behind the global leaders in commercial HBM technology.

The 2027 Supply Chain and Long-term Variables

According to industry estimates, local AI products equipped with CXMT's HBM3E will not be available on the market until 2027 at the earliest. This means that in the next two to three years, China's AI industry will still have to endure the pain of "computing power bottleneck".

But this is by no means meaningless.

The small-scale mass production of CXMT's HBM3E is an important signal for Chinese AI chip companies to test the domestic supply chain and achieve diversification. It proves that Chinese enterprises have the ability to "manufacture" HBM relying on DUV multi-patterning and self-developed advanced packaging technologies, even in the dilemma of being completely blocked from high-end equipment.

In the long run, this forced "internal circulation" will have a far-reaching impact. Once Chinese manufacturers get used to using local HBM, even if the yield rate gradually climbs, it will put pressure on the global market. Industry experts point out that if CXMT can finally achieve scale expansion with the support of government subsidies, and even launch a low-price offensive, it is not impossible for the global HBM price competition pattern to change.

Conclusion: No Choice but to Rise to the Challenge

CXMT's aggressive promotion of HBM3E with only 20% yield is an extremely rare scene in the history of semiconductors.

This is not a complete technical victory, but a solemn and stirring "strategic defense". With extremely high cost and relatively lagging performance, it preserves a faint lifeline for the survival of China's AI industry.

The longer the US export controls last, the more China will be forced to build a completely independent domestic supply chain. This road is extremely muddy, full of sweat, toil and high costs brought by yield climbing. But for CXMT and China's AI industry behind it, there is no choice but to rise to the challenge. Because in this era where computing power means sovereignty, without self-developed memory, no matter how powerful the computing power engine is, it can only become a pile of unusable scrap iron. In this long tug-of-war, surviving first is more important than anything else.

This article is from the WeChat Official Account "Tech Business", written by Tech Business, and authorized for release by 36Kr.