By Decode
In February 2012, a bankruptcy protection order from the Tokyo District Court delivered a death certificate to Japan's DRAM industry.
Elpida Memory, the memory giant inheriting the semiconductor legacies of Hitachi and NEC, collapsed after exhausting its last reserves of cash flow. Six months later, it was acquired by Micron for a mere $2 billion—less than Samsung Electronics' quarterly net profit.
Back then, no one would have imagined that twelve years later, a company from Hefei, China, which started from the ashes of Qimonda, would challenge those monopolists who crushed Elpida in almost exactly the same way.
July 27, 2026, a day destined to be written into the history of China's A-share market.
The domestic DRAM leader, CXMT (ChangXin Memory Technologies Inc.), listed on the STAR Market. It opened at 49.5 yuan, soaring 471.59% from its issue price of 8.66 yuan. By the close, its total market capitalization exceeded 3.31 trillion yuan, dethroning Industrial and Commercial Bank of China to become the new market value king of the A-share market. The day's trading volume reached 141.19 billion yuan, setting a historical record for single-day turnover of an individual stock.
The market frenzy was not without logic.
In Q1 2026, CXMT's global DRAM market share reached 8%. While this figure still seems small compared to Samsung (38%), SK Hynix (29%), and Micron (22%), it represents the first time in twenty years that a non-Korean, non-American company has torn a crack in the DRAM iron curtain.
And this crack was precisely pried open using the "counter-cyclical" weapon most familiar to its rivals.
The Epitaph of Elpida
To understand CXMT's victory, one must return to a war twenty years ago.
In 1999, to cope with fierce competition in the global DRAM market, the Japanese government orchestrated the merger of Hitachi and NEC's DRAM businesses, forming Elpida Memory. In 2003, Mitsubishi Electric's DRAM business was also merged into it.
Elpida's birth itself was an embodiment of "national will" by the Japanese government to save its domestic DRAM industry, seen as Japan's last semiconductor "fortress" and "quasi-national champion."
But Elpida was eventually acquired by Micron Technology in 2012. Japan's former semiconductor "hope" was completely reduced to a part of Micron's global footprint.
Takashi Yunoue, former director of the Japan Precision Processing Research Institute and a 16-year semiconductor R&D veteran at Hitachi and Elpida, wrote a book specifically about this titled "The Lost Manufacturing."
From Yunoue's perspective on Elpida's defeat, Samsung as an external factor wasn't the main culprit. Elpida's own arrogance, sluggish perception of market trends, and organizational fragmentation were the real culprits.
In his book, Yunoue used yield as a case study to detail how Elpida was crushed by Samsung in cost control, and the internal cause of the obsession with yield was the concentrated manifestation of the three aforementioned fatal flaws.
In 2005, while visiting Elpida, Yunoue discovered a shocking fact: the yield rate for the most advanced 512Mb DRAM chips at Elpida was as high as 98%, while Samsung's was only 83%.
Based on this data, various analysts concluded that Elpida's technical level far surpassed Samsung's.
But Yunoue calculated a different set of numbers.
Samsung's 512Mb DRAM chip area was 70 sq. mm, while Elpida's was 91 sq. mm. This meant that from a 300mm wafer, Samsung could produce about 830 chips, while Elpida, with its 98% yield, could only produce around 700.
Even more fatal, in pursuit of the 98% yield, Elpida's manufacturing equipment throughput was only half of Samsung's. In other words, producing the same number of chips cost Elpida double the equipment cost compared to Samsung.
Improving yield from 60% to 80% is relatively easy, but pushing from 80% to 95% comes at an extremely high price. Elpida became obsessed with pushing yield to the extreme, neglecting that unit cost is the life-or-death metric.
In the end, the more technically advanced Elpida had a profit margin of only 3%, while Samsung Electronics boasted 30%.
Yunoue thus summarized a line worthy of being engraved in DRAM industry textbooks: If the unit cost per DRAM increases, even 100% yield is meaningless.
Samsung's philosophy was the complete opposite of Elpida's.
For mass-produced DRAM, Samsung did not deliberately pursue extreme yield but focused on developing next-generation products with smaller area and lower cost.
Simultaneously, Samsung adopted a multi-generational parallel R&D strategy. While Team A developed the 100nm process, Teams B through E tackled 95nm, 90nm, 85nm, and 80nm, respectively.
Whichever team first achieved a breakthrough in integration technology would move to mass production, while other teams shifted to even more advanced processes. This pace of parallel development and rapid iteration allowed Samsung to maintain a leading edge in technology generations.
More crucially, Samsung believed in counter-cyclical investment. Expanding production when others cut back during industry downturns, expanding when others contracted. Exchanging capital endurance for market share, crushing competitors with economies of scale.
It was during one industry winter after another that Samsung used counter-cyclical expansion to exhaust Elpida and ascend to the DRAM throne.
Elpida's collapse revealed the brutal laws of the DRAM industry. This is not a contest of technical precision but a death race of cost and scale. And Samsung was the player who understood this deeply.
Twenty years later, CXMT has almost replicated this playbook. Only this time, the target being crushed by counter-cyclical tactics is Samsung itself.
From Qimonda to Global No.4: What Did CXMT Do Right?
CXMT's story began with a bet on "picking up chips."
In 2016, CXMT's predecessor was registered and established in Hefei. At that time, the global DRAM market had been monopolized by Samsung, SK Hynix, and Micron for over twenty years, collectively holding over 90% of the global market share. China imported nearly $60 billion worth of memory chips annually, with a domestic production rate close to zero.
The first hurdle facing CXMT was the patent barrier. The patent walls built by Samsung, Hynix, and others were enough to halt any new entrant in its tracks.
CXMT's solution was quite ingenious: it picked up chips from the ruins of the bankrupt German DRAM giant, Qimonda. Qimonda was once the world's second-largest DRAM supplier, collapsing in 2009 due to the financial tsunami and a plunge in DRAM prices.
In 2019, CXMT signed an agreement with Canadian company Polaris Innovations, obtaining licenses for approximately 7,000 DRAM patents left by Qimonda, along with over 10 million documents (about 2.8 TB) of comprehensive DRAM technical documentation. The core purpose of this deal was not to directly acquire technology but to secure legal space for subsequent R&D.
Simultaneously, CXMT absorbed former technical talent from Qimonda, including Karl-Heinz Kuesters, who spent 24 years at Siemens, Infineon, and Qimonda as VP of Technology and Early Development. A group of overseas Chinese engineers who had worked at Hynix, Micron, and TSMC were also gradually recruited back, forming CXMT's earliest technical backbone.
In September 2019, CXMT launched its independently designed and produced 8Gb DDR4 product, achieving the breakthrough from zero to one for domestic DRAM.
Mass production was the starting point, but the generational gap with international giants was the real test. CXMT chose a risky path: skip-generation R&D.
After mass-producing 19nm in 2019, the process roadmap skipped 18nm and directly tackled 17nm. On the product front, it skipped the long, gradual iteration rhythm of DDR4 and swiftly shifted focus to DDR5 and LPDDR5. This strategy traded massive capital and market investment for a time window to catch up. CXMT's founding team revealed in 2019 that it had spent $2.5 billion on R&D and capital expenditure since its establishment.
The spiritual core of this approach is identical to Samsung's parallel R&D and rapid iteration.
Although the specific operational forms differ (Samsung uses multi-generational parallel R&D, CXMT uses leapfrogging to conquer more advanced nodes), the core logic is consistent: using multi-threaded advancement instead of linear catching-up, using capital investment to buy time windows.
Thereafter, CXMT gradually progressed from its first-generation process platform to its fourth. Products evolved from DDR4 and LPDDR4X to DDR5, LPDDR5, and LPDDR5X. Current DDR5 chip speeds reach 8000 Mbps, with LPDDR5/5X up to 10667 Mbps.
If skip-generation R&D was CXMT's strategy for catch-up speed, then yield ramp-up was its core capability to translate technology into competitiveness—precisely where Elpida fell.
According to industry reports, CXMT's initial DDR5 chip yield at mass production start was only about 50%, gradually improving through continuous process optimization. By the second half of 2025, its 17nm process DDR5 yield had surpassed 90%. Its DDR4 chip yield had long been stable around 90%.
Unlike Elpida, CXMT did not sacrifice equipment throughput and unit cost to pursue extreme yield. Its IDM (Integrated Device Manufacturer) model enabled deep synergy between process development and chip design, achieving systemic efficiency advantages in iteration speed, yield ramp-up, and performance optimization.
The rapid yield improvement directly translated into lower unit cost, the most core competitiveness in the DRAM industry. CXMT's gross margin rose from -1.93% in 2023 to 40.99% in 2025, approaching the levels of overseas players like Samsung and Micron.
In the DRAM industry, where cost is the line between life and death, a yield of 90% is sufficient. CXMT did not repeat Elpida's mistake. The cost of pushing yield from 90% to 98% far outweighed the benefits gained from those 8 percentage points.
This is precisely the core tenet of Samsung's cost discipline: DRAM competition is essentially about unit cost, not a numbers game of yield.
If "skip-generation R&D" and yield ramp-up were CXMT's technological catch-up, then the counter-cyclical capacity expansion in 2023 was its strategic-level direct replication of Samsung's counter-cyclical playbook.
In 2023, the global memory industry experienced its coldest winter in fifteen years. DRAM prices plummeted over 40%, with lows down about 50% from the highs in the first half of 2022. Samsung, SK Hynix, and Micron were forced to announce production cuts and reduced capital expenditure. Industry consensus held that new entrants were "doomed" in such an environment.
CXMT, however, made a startling decision: counter-cyclical expansion.
Despite a net loss attributable to shareholders of 16.34 billion yuan, annual R&D expenditure reached 4.67 billion yuan. Monthly 12-inch wafer capacity was raised from 90k to 150k. Simultaneously, it relied on aggressive pricing to enter customer supply chains, with product quotes once as low as half of overseas competitors' prices.
Counter-cyclical expansion coupled with price-cutting strategy, while causing high inventory and severe book losses in the short term, forcefully tore away market share during the giants' production cut period.
This gamble followed the exact same logic Samsung used to defeat Elpida. Yunoue writes in his book that Samsung, through round after round of counter-cyclical investment—expanding when others cut back, expanding when others contracted—exchanged capital endurance for market share and crushed competitors with economies of scale.
Supporting this gamble was a group of patient capital behind CXMT. Hefei State-owned Capital, the National Integrated Circuit Industry Investment Fund Phase II, followed by industrial capital from Alibaba Cloud, Tencent, Lenovo, Xiaomi, and others, successively entered the fray. In the second half of 2023, when DRAM prices halved and market-oriented investment institutions hesitated, it was this long-term capital that helped CXMT weather the most difficult moment.
CXMT's counter-cyclical gamble paid off in 2025.
In 2025, the AI wave erupted, completely altering the demand structure for memory chips.
Samsung, SK Hynix, and Micron shifted major capacity and R&D resources heavily toward higher-margin HBM (High Bandwidth Memory) and server DRAM, actively "ceding" part of the consumer electronics and traditional DRAM market. The capacity CXMT had built counter-cyclically perfectly plugged this structural gap.
According to Counterpoint Research data, CXMT's global DRAM market share climbed from 3% in Q1 2025 to 8% in Q1 2026. Omdia data showed its share jumping from 4.7% in Q4 2025 to 7.6% in Q1 2026.
During the same period, the market shares of Samsung, SK Hynix, and Micron were approximately 38%, 29%, and 22%, respectively.
The leap in market share directly translated into an explosion in financial data.
From 2023 to 2025, CXMT's revenue soared from 9.087 billion yuan to 61.799 billion yuan. Net profit attributable to shareholders turned from a loss of 16.34 billion yuan to a profit of 1.875 billion yuan. In Q1 2026, quarterly revenue reached 50.8 billion yuan, with net profit of 24.762 billion yuan.
Profit from a single quarter almost wiped out accumulated losses from previous years. The company forecasts revenue of 110-120 billion yuan and net profit of 50-57 billion yuan for the first half of 2026.
Product structure also improved simultaneously. In 2025, the LPDDR series contributed about 66% of revenue, with the DDR series accounting for about 32%.
Currently, CXMT's LPDDR products have achieved adoption rates exceeding 30% in Chinese Android brand smartphones (excluding Huawei). End customers cover leading manufacturers like Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, Honor, OPPO, and vivo.
Conclusion
CXMT's story is a classic case study in counter-cyclical strategy. But standing at the 8% market share node, the real challenges may have just begun.
A trickier hurdle comes from HBM.
The AI wave has spawned structural differentiation within the memory chip sector, with HBM becoming the juiciest profit source. This market is almost entirely carved up by SK Hynix and Samsung.
Over 98% of CXMT's revenue still comes from traditional DRAM, i.e., commodity-grade chips for servers and smartphones.
According to industry sources, CXMT technically possesses HBM3 mass-production capability, but yield is only 25%-35%, still some distance from the commercial threshold. The chase is already underway; CXMT plans mass production of HBM3E by 2027. But by then, competitors may have moved to the next generation.
The process node gap is another unavoidable hard constraint.
Hynix's sixth-generation 10nm-class DDR5 is in full mass production, while CXMT's mainstay is still 17nm (G4). TechInsights teardowns show CXMT's bit density is about 0.239 Gb/mm², still lagging behind top international levels. Closing the technology gap takes time, and iteration in the DRAM industry waits for no one.
A more hidden threat comes from the giants' counterattack.
Samsung, SK Hynix, etc., are already expanding capacity, with industry expectations that new capacity will gradually come online starting in 2027. Over the past two years, the three giants shifted main capacity to HBM, inadvertently ceding space in the traditional DRAM market to CXMT.
But this window won't remain open forever. Once they complete HBM capacity deployment and return to the traditional DRAM market, a price war is inevitable. Then, CXMT will face a frontal onslaught it has never experienced before.
However, CXMT also holds chips. Of the net proceeds from this IPO, 7.5 billion yuan is invested in upgrading the memory wafer manufacturing mass production line, 13 billion yuan in DRAM memory technology upgrades, and 9 billion yuan in forward-looking technology R&D. SemiAnalysis estimates CXMT's monthly capacity will reach about 350k wafers by the end of 2026.
From picking up the first chip on Qimonda's ruins to a STAR Market capitalization exceeding 3 trillion yuan; from the first DDR4 mass production in 2019 to an 8% global share in Q1 2026. CXMT completed a classic "counter-cyclical" campaign in ten years.
But the reason this battle could be won can find all its answers on Elpida's body.
The "autopsy report" Yunoue wrote for Elpida in his book: it did not die from a single mistake but from the collapse of an entire system.
Technologically, Elpida's engineers achieved a 98% yield on 512Mb DRAM, while Samsung's was only 83%. But Samsung's chip area was 21 sq. mm smaller, with twice the equipment throughput. The result: Samsung, which lost on yield, had half the unit cost of Elpida.
Organizationally, Elpida was formed from the merged DRAM businesses of Hitachi and NEC, but the technical cultures of the two parent companies never truly fused.
Strategically, Samsung had a 230-person market research team deeply embedded on the customer frontlines to understand needs. Elpida's engineers were obsessed with technical metrics in the lab, blind to market temperature shifts. After the 2008 financial crisis, as the PC market shrank and smartphones rose, causing a dramatic shift in demand structure, Elpida reacted slowly, while Samsung had already planned ahead.
Three weaknesses pointed to the same root cause: Elpida treated the "technical precision race" as the goal itself, forgetting that the DRAM industry is essentially a comprehensive war of cost, organization, and market.
CXMT did not take this path. It picked up patents and technology from Qimonda's ruins, recalled overseas Chinese engineers from Hynix, Micron, and TSMC, used a unified technical roadmap and cost discipline to bind all forces into a single rope. It targeted visible, tangible customers like Chinese Android phones, servers, and cloud computing vendors.
CXMT understood the sentence engraved on Elpida's tombstone: In the DRAM industry, leading in a single dimension is meaningless.
