Why China Wants to Make HBM Even at 20% Yields
CXMT plans limited HBM3E production within two months, Reuters reported. Reported yields trail SK hynix by a wide margin, but under US export controls, supply itself has economic value in China.
Why make a product when only two or three of every ten units pass inspection? In an ordinary chip business, that number would raise immediate doubts about cost competitiveness.
Making more defects than good parts
China's math is different. On Oct. 28, Reuters, citing two anonymous sources, reported that CXMT, China's largest DRAM maker, plans to start limited production of HBM3E within two months. CXMT has reportedly already sent samples to several Chinese AI chip designers. It is also said to have offered a modified version aimed at improving yield.
The key point: making HBM available inside China comes before making it cheaply. Under US export controls on advanced memory, the mere possibility of supply carries real economic value for Chinese AI chip companies.
How different is the cost at 20% versus 70% yield?
Yield is the share of units produced that pass inspection. Put 100 units through the same process at the same cost. A line that yields 20 good units and one that yields 70 spread that cost very differently across each good unit.
| Yield | Relative cost per good unit |
|---|---|
| CXMT, low end: 20% | 5.00 |
| CXMT, high end: 30% | 3.33 |
| SK hynix: 70% | 1.43 |
Calculation: relative cost per good unit = 1 ÷ yield. It assumes identical input cost, product specification, process steps and cost allocation. At 20% versus 70%, the gap is about 3.5 times. At 30% versus 70%, it is about 2.3 times. These figures do not represent actual selling cost or profit margin.
The yield figures are estimates from anonymous sources. Neither CXMT nor SK hynix has confirmed them, and it is not clear that product mix and the definition of yield are the same for both. The comparison above is a hypothetical to illustrate manufacturing logic, not a real cost table for the two companies.
China also pays a cost for not producing
HBM stacks DRAM vertically to feed large volumes of data quickly to an AI processor. Without enough HBM, it is hard to build AI accelerators at all. Workarounds add cost.
According to Reuters, Chinese AI chip makers such as Huawei and Cambricon raised prices of finished AI chips by 20% to 50% amid memory cost pressure. That does not mean HBM prices rose 20% to 50%. It signals that the difficulty of securing memory is being passed through to the whole AI chip business.
So CXMT's early product could appeal to some Chinese customers even if it costs more to make than rival products abroad. What buyers are paying for is not only memory performance but availability and control over the supply chain. That does not guarantee CXMT a profit, however.
An HBM3E label does not mean equal performance
According to Reuters' sources, some evaluations found CXMT's HBM3E performing closer to HBM3, the generation before SK hynix's HBM3E. Even under the same product name, bandwidth, power efficiency, stability and customer qualification can differ.
The competitiveness of SK hynix and Micron does not rest on chip design alone. It combines difficult stacking processes, high yields, customer-specific qualification and experience supplying at volume. A Chinese maker starting production does not make that moat vanish at once.
Nor is there any guarantee the current gap is permanent. Repeated production runs and customer feedback can speed up yield improvement. US equipment export controls make catching up harder, but it cannot be said that they stop the learning itself.
This article is for information and investor education and is not a recommendation to buy or sell any security.
Insight Times Editorial Desk
