54.4 billion yuan in profits and a 20% yield rate: CXMT's memory super cycle stands at a narrow doorway.
Changxin Technology has released a quarterly report that catches the attention of every semiconductor company: its attributable net profit in the third quarter reached 54.4 billion yuan, with revenue hitting 118.71 billion yuan, representing a year-on-year increase of 613.2% and a quarter-on-quarter increase of 19.3%. A DRAM manufacturer that was still struggling on the verge of losses a few years ago has earned a figure in a single quarter that many people could not have imagined. Turning losses into profits is newsworthy enough in itself, but what is worth further analysis is exactly where this 54.4 billion yuan came from and how long it can last.
Almost at the same time, another set of figures is also noteworthy: the HBM3E chips that this company is about to mass produce have a yield rate of only 20% to 30%, while its South Korean competitor SK Hynix has a yield rate of around 70% for the same product. The 54.4 billion yuan profit and the 20% yield rate, placed side by side, exactly form the entry point to understand Changxin's current situation — its profits come from the industry cycle, and its future depends on whether it can raise the yield rate from "functional" to "competitive" before the cycle narrows.
Cyclical Profits
The primary explanation for the 54.4 billion yuan net profit does not lie in Changxin's own technological leap, but in a more macro structural fact: HBM is systematically "siphoning off" the global wafer production capacity of DRAM, and Changxin is just standing on the vacant land left after this siphoning.
Producing each bit of HBM consumes about three times the wafer capacity of standard DRAM, due to the larger chip area, complex stacking process, and yield loss throughout the whole process. When Samsung, SK Hynix, and Micron divert more and more DRAM wafers to HBM, the production capacity of standard DRAM for PCs, mobile phones and general-purpose servers continues to be compressed. The proportion of HBM wafer starts of the three major original manufacturers in the total DRAM wafer starts rose from about 18% at the end of 2025 to 22% at the end of 2026, but the proportion of HBM bit supply in the total DRAM bit supply only rose from 8% to 9% — nearly a quarter of the wafers are allocated to products that only produce about one-eleventh of the bit supply. The supply gap of general-purpose DRAM cannot be naturally bridged by slowing demand, but is determined by the product structure itself.
Changxin is the domestic DRAM supplier with the largest revenue scale in the world. Its global DRAM revenue share rose to 8% in the first quarter of 2026, and its customers cover core domestic computing power and terminal buyers such as Alibaba Cloud, ByteDance, Tencent, and Lenovo. Under the pattern where overseas giants shift production capacity to HBM and the supply of standard DRAM is systematically tightened, Changxin's products have no trouble selling, and the price is determined by the seller. The DRAM contract price in the third quarter still maintained a double-digit quarter-on-quarter increase, and Changxin's own statement in the Q3 report also confirms this point: revenue growth is jointly driven by "a sharp rise in sales volume and sales price".
In other words, the essence of the 54.4 billion yuan profit is a supply vacuum created by geopolitics and product structure together, not that Changxin has caught up with Samsung and SK Hynix in DRAM core technology or cost control. This distinction is crucial, because it determines the sustainability boundary of this profit. Profits given by the cycle will also be partially taken away by the cycle; the companies that can really retain profits on the balance sheet are those that complete technology climbing and cost curve reduction at the prosperity peak. What Changxin has obtained now is a ticket to enter the next round of competition, not a report card that has fully realized all its value.
From the perspective of competition pattern, this supply vacuum is not accidental. Samsung, SK Hynix, and Micron are all tilting their product portfolios to high-margin, high-barrier AI storage products such as HBM, so the supply elasticity of standard DRAM has decreased: even if prices continue to rise, original manufacturers lack the motivation to transfer wafers back from HBM to general-purpose DRAM, because the output value per wafer and strategic priority of HBM are higher. This leaves a relatively loose window for Changxin — in the general-purpose DRAM market that overseas giants are unwilling to fully cover, Changxin can use its price advantage and local supply capacity to exchange for market share and profits. The reason why a window is called a window is that it will not stay open forever.
HBM Admission Ticket
What truly reveals Changxin's technological position is exactly the story it most wants to tell: HBM.
Changxin plans to start limited production of HBM3E within two months, and has delivered samples to many domestic AI chip design companies. However, details disclosed by insiders outline a not-relaxing picture: the yield rate is only 20% to 30%, which means a large number of chips have to be scrapped or reworked, directly pushing up the unit cost; in tests, the performance of Changxin's HBM3E is closer to SK Hynix's previous-generation HBM3, rather than the current-generation HBM3E; its expected pricing will also be higher than that of overseas competitors' same products.
This set of facts points to a clear judgment: Changxin's HBM business at the current stage is a strategic "loss-making for admission", rather than a profit contributor. The reason why it still pushes forward is not the calculation of commercial returns, but that Chinese AI chip manufacturers "cannot obtain enough advanced storage chips to expand production". Under the background that US export controls list Changxin's chip facilities as "controlled facilities", the demand of the domestic AI computing power chain for local HBM supply has gone beyond the simple consideration of cost performance. Changxin's HBM3E is first a supply chain security issue, and then a business issue.
The economics of yield rate are extremely harsh here. A yield rate of 20% to 30% means that for the same wafer, the number of qualified HBM3E chips that can be delivered is only about one-third of that of SK Hynix; the wafer cost, packaging cost and test cost allocated to each qualified chip will be significantly amplified. HBM itself is an advanced packaging-intensive product. The more stacking layers there are, the more complex the yield loss will be, and the cost curve brought by low yield is not linear, but will accelerate to deteriorate with the increase of stacking layers and packaging difficulty. Therefore, even if Changxin's HBM3E is priced higher than overseas similar products, it does not necessarily mean high gross profit, and it is more likely to be a passive coverage of low yield cost. Its current production capacity scale is limited to "limited production", which essentially uses a small batch to control the loss exposure while accumulating engineering data.
And the difficulty of this problem is still increasing. Samsung took the lead in starting mass production of HBM4 globally in February 2026, and its yield rate has increased from less than 60% in the initial stage to about 80%; SK Hynix still occupies about 54% of the share in the HBM4 market. Changxin's catch-up on HBM3E is facing a time window when competitors have already switched to next-generation products. There is not much space left for the 20% yield rate to climb slowly. More critically, HBM competition is not only the competition of wafer manufacturing yield, but also the competition of packaging capacity, customer certification cycle and system-level verification. Domestic AI chip manufacturers are willing to give Changxin a chance for samples, which is driven by the logic of supply chain security; but from sample to large-scale procurement, there are multiple tests of stability, consistency and long-term supply capacity in between.
What the Stock Price Is Afraid Of
Changxin's stock price has fallen by 32% from its August high (as of mid-to-late October), which happened around the announcement of the 54.4 billion yuan profit. It seems contradictory, but the logic is self-consistent. The market is not pricing the profit of the past quarter, but how long the demand assumption that this profit relies on can hold.
The sustainability of AI capital expenditure is the biggest point of divergence in the current storage sector. The optimistic side has sufficient data support: Nomura estimates that the global new data center capacity will increase by about 41% next year, and the storage shortage may continue until 2028; Citi expects HBM demand to increase by about 62% year-on-year in 2027, and there will still be a supply gap for DRAM from 2027 to 2028. But cautious signals are equally real: the quarterly increase of DRAM contract price has narrowed from about 80% in the second quarter to about 10% to 15% in the fourth quarter, the rising speed is braking, while the supply contraction has reached a maximum of 50% — price increase is decelerating, and shortage is accelerating, both happening at the same time. The willingness of PC and mobile phone manufacturers to accept continued price increases has decreased significantly. The proportion of storage chips in the total BOM of complete machines has risen from 10%-15% to 30%-40%, surpassing SoC for the first time and becoming the most expensive single component. When storage changes from "cheap enough to ignore" to "the most expensive part", the elasticity of end demand begins to be activated.
The fall of Changxin's stock price essentially prices a possibility in advance: if the growth rate of AI capital expenditure slows down, or new production capacity is released intensively from 2027 to 2028, the balance between DRAM supply and demand may swing back faster than the current consensus expectation. At that time, the part of Changxin's profit that comes from the "standard DRAM supply vacuum" will be the first to come under pressure. Its HBM business, with the yield rate still at the 20% level, is not enough to form a hedge on the profit side.
Another factor priced into the stock price is the switch of valuation anchor. Changxin's stock price soared more than five times on the first day of listing in Shanghai, and it became the listed company with the highest market value in China in July. Its 58 billion yuan IPO is the largest IPO in Asia this year. Such pricing includes strong domestic substitution premium and AI storage narrative premium. When the narrative enters the realization period, the market will start to examine with stricter standards: how much of the profit comes from sustainable product upgrading, and how much comes from cyclical price increases; when the HBM business can transform from strategic input to profit contribution; whether capital expenditure and R&D investment will form depreciation pressure in the next downward cycle. A profit of 54.4 billion yuan can support valuation, but cannot define valuation alone. The 32% fall of stock price from the high point is not a denial of the Q3 performance, but a revision of the assumption that "the super cycle will extend infinitely".
The Narrow Gate
Changxin is in an unusual time window: the cycle brings it profits, geopolitics brings it demand, but technology sets a narrow gate for it.
The quarterly profit of 54.4 billion yuan gives it ammunition for R&D investment and capacity expansion — the R&D investment in the third quarter is about 4.2 billion yuan, which is not aggressive compared with the net profit of 54.4 billion yuan. Although export controls list its facilities as controlled objects, Changxin has previously stocked up DUV lithography equipment for about three years of use, so short-term capacity expansion will not be directly strangled by equipment supply cuts. What is really scarce is not money or equipment, but the time needed to raise the yield rate from 20% to 70%.
In the semiconductor industry, there is no shortcut for yield climbing. It depends on the precision of process control, the decline rate of defect density, the maturity of the packaging link, and the experience curve accumulated by a large number of engineering batches. It took SK Hynix several years to reach 70%. Changxin needs to complete the same climbing in a shorter time in the rhythm that competitors have switched to HBM4. Moreover, the target it faces is not static: when Changxin pushes the HBM3E yield to a commercially viable level, overseas competitors' HBM4 may have entered the stage of large-scale volume release, and customers' pricing tolerance for HBM3E will also decline accordingly. The pursuer not only needs to run fast, but also needs to recover the investment of the previous generation of products before the next generation of products switches, which puts extremely high requirements on capital discipline and engineering efficiency.
The 54.4 billion yuan profit proves that Changxin has become a DRAM supplier that cannot be ignored. But the 20% yield rate reminds another thing: it is still separated from becoming a company with commercial competitiveness in high-end storage by a narrow gate guarded by engineering discipline and time together. The cycle will reward companies that stand in the right position, but will not cross that gate for them. For Changxin, the best news is that it now has enough profits to win time for technology climbing; the most dangerous possibility is to misinterpret the profits given by the cycle as the completion of technology catch-up. 54.4 billion yuan is ammunition, 20% is a road sign, and the narrow gate is still ahead.
This article is from the WeChat official account "Tech Business", author: Tech Business, published with authorization by 36Kr.