The 10% from Changxin is what remains when giants turn.
Hefei Changxin is the only company in mainland China capable of mass-producing memory chips. It listed on the STAR Market on July 27 at an issue price of 8.66 yuan, surged more than fourfold on its first day, and within just over a month became one of the largest companies on the A-share market.
Counterpoint's report, released on September 2, showed that CXMT's global DRAM market share, based on revenue, rose from 4% a year ago to 8% in the first quarter and reached 10% in the second quarter—marking its first time entering double digits.
The report was released, and the stock market celebrated immediately—on September 7, ChangXin's stock price rose another 5.8%.
First, observe the direction in which the giants are turning.
Memory comes in two types. Ordinary memory is installed in computers, phones, and servers for data storage and retrieval; it uses mature technology and is produced in a standard, routine manner. The other type is called HBM, a high-speed memory specifically designed for AI chips.
Regular memory is laid out flat, while HBM stacks memory units layer upon layer like building a skyscraper, allowing AI chips to instantly access massive amounts of data. But these layers aren’t built for free—the taller the stack, the more likely it is to tilt, resulting in more defects; from the same wafer, significantly fewer usable chips can be produced.
AI competitions are fierce, requiring large models to load all their parameters into memory and stay ready; insufficient memory means the models can't run. An AI server needs as much memory as 8 to 10 ordinary servers combined.
HBM has become the most sought-after product, with Samsung, SK Hynix, and Micron allocating nearly all of their new capacity to it.
Just look at Micron’s own calculations to see how thoroughly it’s committed: producing one HBM3E chip consumes the capacity of three standard memory chips.
New production lines are being built in batches, with Samsung’s Pyeongtaek and SK Hynix’s new facilities focusing almost entirely on HBM and high-end memory. The expansion plans for conventional memory show almost no activity.
Left unattended, it gets stuck in the cycle.
Memory manufacturers' production timelines, from groundbreaking to shipment, are measured in years. Micron’s new factory in the U.S. won’t begin ramping up production until the second half of 2027; even if major players now shift back to conventional lines, the products won’t hit the market for another two to three years.
As a result, this gap remained unfilled for two years.
Who suffered the most? Hynix—its DRAM market share dropped from 39% to 25% in just one year, with the most comprehensive production shift.
Samsung didn't allow it; its share rose from 33% to 38%, as it focuses on the high-end AI server market, which is not in the same league as CXMT; Micron slightly increased to 24%.
Prices have gone wild. A year ago, a 16GB memory module cost $4.80; now it’s $37.50—nearly seven times higher. Memory modules for servers with 256GB have surged past 50,000 RMB.
Distributors say memory prices change almost daily; some vendors in Huaqiangbei have stopped posting price tags altogether and now quote prices verbally, afraid that listing too low will cause losses, while listing too high will deter buyers.
Eventually, the profit margin of conventional memory surpassed that of HBM; some estimates suggest that the land left behind by giants became gold mines through neglect.
At this point, Changxin held firm to its position. It never changed course—it has always been focused on this.
It has competitive products on hand—both DDR5 and LPDDR5 are next-generation offerings. In third-party testing conducted in February, its consumer-grade DDR5 performed nearly on par with Samsung and SK Hynix, while priced only 5% to 10% lower—not through deep discounting.
The factory is operating at full capacity for all DDR and LPDDR lines, with a capacity utilization rate of 95.7% last year and full capacity this year.
Q2 revenue reached RMB 99.5 billion, a 977% year-over-year increase. First-half net profit amounted to RMB 77.6 billion, with a gross margin of 84.84%. For every RMB 100 of goods sold, over RMB 84 is profit—a figure that is astonishing in any industry.
To earn this amount despite having a cost per bit that is 30% higher than its Korean and American competitors, under conditions where it is at a cost disadvantage.
But for now, the giants are simply not ready to pick up this opportunity—the new HBM factory is still under construction, the window has been opened, and it will eventually close, differing only in timing.
......
What is Changxin aiming to gain by rushing these past two years? After analyzing the data, Zhiyuan identified three points:
First deposit, placed on the order.
In July this year, ByteDance signed a five-year contract worth over $7 billion to purchase server memory. Alibaba not only buys products but also became a shareholder, making it both a shareholder and a customer—tied together most tightly.
Tencent is also on the long-term order list, alongside other major domestic AI infrastructure companies—all orders are booked through 2027, with future years' inventory already pre-sold.
Cloud providers are willing to sign agreements to secure supply; last year, buying memory was a race, but this year, they’re directly locking in production capacity. One secures supply, the other locks volume—each gets what they need. Long-term agreements lock in volume, with prices tied to market conditions; by securing volume during the window period, they effectively weld their share into the production schedule.
The second bet is on production capacity.
Shares are determined by production capacity—look at how much wafers the fabs produce to see how much each player holds in the global market. In the memory industry, there is no foundry model; to ship products, you must build your own fabs—factories are everything.
All three factories are now operating at full capacity, with a monthly output of approximately 300,000 wafers, which, according to broker estimates, amounts to about 80% of Micron’s capacity.
By the end of 2026, capacity will expand to 350,000 wafers; by 2027, it will reach 420,000 wafers; and the target for 2028 is 500,000 wafers. A super factory will also be built in Shanghai, with a planned monthly capacity of 400,000 to 600,000 wafers, set to come online in 2027 and reach full capacity in 2028—equivalent to the combined output of three current factories.
The funds for expansion were already secured during the IPO; $57.9 billion was raised, while the planned investment was only $29.5 billion, resulting in over twice the required amount—this surplus was specifically raised to finance the expansion.
The third bet is placed on the outlet.
Its reach has extended far beyond expectations: overseas revenue reached RMB 95.8 billion in the first half of the year, accounting for 63.8% of total revenue—up from 36.9% a year ago, with the absolute amount increasing 17-fold over the past year.
The rapid increase is partly due to price appreciation, and the growth in volume is also genuine.
Note that this figure includes Hong Kong re-exports and may not represent solely overseas end-user orders—it does not export directly but relies on Chinese brands to distribute its products. Xiaomi, OPPO, vivo, Transsion, and Lenovo incorporate its memory modules and sell them globally, allowing it to leverage the distribution channels of these Chinese brands.
Xiaomi's new foldable phone is the first in the world to use mass-produced LPDDR6. The U.S. market is legally restricted, but elsewhere, it's already on the move.
After placing three bets, I looked back—this betting strategy wasn’t something I’d tried just once.
In 2023, industry prices dropped by 40%, and its quarterly gross margin reached -112%, meaning it lost more than one unit for every unit sold—yet it continued expanding production, increasing monthly capacity from 90,000 to 150,000 units.
Back then, the industry was in ruins, with giants retreating—but it bet that the cycle would return.
Over the three years from 2022 to 2024, it accumulated losses of 31.8 billion. While others gradually caught up generation by generation, it skipped generations in R&D and directly moved to the next product line. It wasn’t until 2025 that it posted its first full-year profit, netting 1.875 billion—now it can earn that amount in just one week.
Zhu Yiming once said that semiconductors are a marathon—while others compete to see who runs fastest, he bets on who lasts the longest. He bets during downturns and reaps rewards during upturns; this has always been how CXMT operates.
But this time, the bet was placed with borrowed money.
By the end of 2025, Changxin's interest-bearing debt reached RMB 152.784 billion, with long-term loans alone accounting for RMB 118.8 billion. This debt did not accumulate overnight; over the past decade, factory after factory was built, with most funds tied up in facilities and equipment.
At that time, the debt-to-asset ratio was 54.24%, but by the end of this June, it dropped to 42.16%. The funds from the IPO and profit gains from price increases have strengthened the balance sheet, leaving 143.4 billion yuan in cash on hand—things are going very well right now.
Experienced players use profits to build factories; this one uses promissory notes. The most critical issue is the interest hurdle—known in industry terms as the interest coverage ratio—at -6.18 in 2023 and -1.79 in 2024.
The money earned over those two years didn’t even cover the interest; it wasn’t until 2025 that I recovered and climbed to 3.43.
Money is easy to make during the window period, but once the window closes, prices collapse while interest payments continue, and funding for the next generation of factories must be sought elsewhere. It’s betting everything on getting through the window—converting all the window’s windfall gains into production capacity and market share to grow itself bigger.
The profits on the books are real, and the off-the-books debts are real too—both are solid, and that’s the full stance of its bet.
......
The stake is fully locked in—can it feed itself above the safety line before the window closes in 2028?
The 15% safety line in the industry was paid for in blood.
The annotation for this line was provided by the Research Director at Counterpoint, a veteran with three decades of storage research experience at Samsung and Goldman Sachs.
His exact words:
DRAM manufacturers' market share has fallen below 15%; without funding to build next-generation wafer fabs, they will continue to shrink to just 3%, becoming mere spectators.
This line works—Taiwanese manufacturers have verified it, at the cost of several wafer fabs.
During the 2008 financial crisis, DRAM prices collapsed, causing Nanya Technology, Hynix, and ProMOS to see their global market share fall below 15%, prompting banks to avoid them.
In the years that followed, Hynix was acquired by Micron, ProMOS went bankrupt, and Nanya Technology survived to this day thanks to financial support from the Formosa Plastics Group, its market share reduced to just 2%; of Taiwan’s once-dominant three DRAM giants, only half remains as a mere spectator today.
The line is drawn at 15%; what’s being blocked is money. DRAM is a capital game—the investment for the next-generation wafer fab is only going to rise; ChangXin’s own Hefei Phase I project already invested RMB 56.5 billion, and it will only become more expensive going forward.
As long as the factory is operational, depreciation continues daily. ChangXin's annual depreciation exceeds 20 billion yuan and is still rising. During the downturn, with no sales volume to support it, cash flow has been completely depleted.
When the economy is booming, banks chase after lending, and anyone can secure funding. When the economy slows, banks pull back first, leaving only the largest players able to borrow.
Taiwan's third factory died in 2008 during the price crash, same cause—funding ran out.
The trend line is clearly drawn; now, how far is鑫 from it?
10% is calculated based on revenue. Memory prices surged dramatically this year, driving up incoming revenues. ChangXin also acknowledged in its prospectus that a key reason for the significant performance improvement was the increase in product unit prices and gross profit margins—in other words, most of this round of profits came from price hikes.
In the industry, competitiveness is measured by outflow volume; that metric remains around 9%, still far from the 15% target; the 10% the market is celebrating, by industry standards, has only completed about two-thirds of the journey.
The countdown for the remaining journey has begun.
Goldman Sachs expects memory prices to peak around mid-2027. Counterpoint anticipates that the major players' current HBM capacity expansion will conclude around 2028, followed by a subsequent capacity rebound.
According to TrendForce’s production capacity data, global新增 DRAM capacity in 2028 alone will reach 550,000 wafers per month—exceeding ChangXin’s entire 2028 capacity target. The giants’ response will not be slow.
Changxin certainly won't sit idle waiting for a rebound—with orders in hand, production capacity doubled, and export channels opened, every one of the three investments has paid off.
The product continues to rise, with server memory revenue increasing from 8.4% to 26.5% over the past year.
Institutional views on its prospects are wildly divergent. Nomura predicts its market share will reach 18% by the end of 2028, comfortably surpassing the threshold; Counterpoint’s baseline forecast, however, estimates only 11% by 2028 and just 15% by 2035.
The same company—optimistically, it will pass the threshold in two years; conservatively, it will take another nine. The disagreement is placed on the same bet: one bets it will move fast, the other bets the window will close quickly.
I heard someone say that it has backing from the state, so this resistance level can't hold it.
State capital has indeed been involved from start to finish—the initiator is Hefei’s state-owned capital, and Fund II of the National Integrated Circuit Industry Investment Fund also holds a significant stake; this is true. However, state funding has its limits; each factory built comes with a bill in the billions.
The company is well aware of this; during the IPO roadshow, Secretary Yuan Yuan stated that if new production capacity is released en masse, the industry could return to a downturn.
It is also securing a backup position—HBM is the juiciest target of the AI era and the most tightly guarded fortress of the three giants.
ChangXin's HBM has just been sampled; small-scale pilot production won't be discussed until year-end, and yield is still improving. Institutional models estimate that currently three-quarters of the output must be scrapped. Alibaba's Pingtouge and Cambricon are currently testing it.
Regarding 3D DRAM, domestic equipment has just passed the etching stage. These are the stories that come after the window of opportunity—whether they succeed or not, we’ll know in two years.
Go back to the number at the beginning.
Since the report came out, the market has been celebrating a 10% gain, and the stock price continues to rise; the countdown within the industry has been ticking since the giants turned their backs.
Whether it can get inside the line before the window closes, 2028 will provide the answer.
Data source:
[1] Counterpoint Q2 2026 Global Memory Tracking Report, ChangXin Technologies H1 2026 Annual Report and Prospectus, TrendForce Counterpoint capacity data, and institutions such as Goldman Sachs, Nomura, and ZhongTai Securities; this article does not constitute investment advice.
This article is from the WeChat public account "Wang Zhiyuan" (ID: Z201440), authored by Wang Zhiyuan.