[fn Square] No reason to be rattled by China's CXMT catching up

Recently, when China's ChangXin Memory Technologies (CXMT) announced that it would unveil server DRAM with a 4F2 structure this year, the share prices of Samsung Electronics and SK hynix were badly shaken. But announcing a new technology is an entirely different matter from achieving high-yield mass production and securing price competitiveness. CXMT's progress should be watched carefully, but there is no reason to take its announcement alone as a sign that South Korean DRAM technology is about to be overtaken.
Samsung Electronics and SK hynix currently lead the market, mass-producing advanced 11–12nm-class 6F2 DRAM and high-bandwidth memory (HBM). In the second quarter of this year, Samsung Electronics held 39.4% of global DRAM sales, SK hynix 24.9%, and CXMT 9.5%, giving the two South Korean companies a combined share of more than 64%. Samsung Electronics presented 4F2 DRAM at an international conference back in 2009, and this year it also demonstrated the operation of a 10nm-class 16Gb 4F2 test chip. SK hynix is also developing vertical-gate and peripheral-circuit stacking technologies. In other words, 4F2 is not a novel proprietary technology developed by CXMT; Samsung Electronics and SK hynix have also been researching and preparing it for a long time.
Moving from 6F2 to 4F2 reduces the memory-cell area by about 33% under the same design rules. Combining this with a PUC structure, which vertically stacks peripheral circuits beneath the cells, could reduce the total chip area by as much as about 50%. However, this also brings new challenges, including leakage current in vertical channels, wafer bonding, process uniformity, and achieving adequate yields. This is also the context for CXMT's search for a breakthrough to increase density with a 4F2 structure: U.S. export controls restrict its access to extreme ultraviolet (EUV) equipment, making it harder to apply finer design rules. CXMT is applying the 4F2 structure to 14nm-class DRAM, while Samsung Electronics and SK hynix plan to use it for DRAM below 10nm. It is clear, therefore, that the 4F2 technology being developed by South Korean companies is more technically challenging and more advanced in terms of process miniaturization. Still, even if South Korea is considered to be at least three years ahead in DRAM process technology, should CXMT mass-produce a 4F2 product first, it could gain earlier experience in resolving defects arising from the new cell structure and improving yields. This is certainly something to watch carefully.
The recent sharp fall in South Korean semiconductor stocks should also be viewed separately from the industry's actual boom. Uncertainty over U.S. interest rates, financial risks facing AI data-center companies that rely on borrowing, foreign investors taking profits, and CXMT's announcement all combined to shake South Korea's stock market, which has a heavy semiconductor weighting. But anxiety over highly leveraged operators such as CoreWeave does not mean that all AI investment by cash-generative big tech companies such as Microsoft, Google, and Amazon will come to a halt. Share prices reflect expectations for future earnings in advance, but falling stock prices do not mean that orders for HBM and server DRAM will immediately disappear.
Samsung Electronics' semiconductor (DS) division's operating profit actually rose from 53.7 trillion won in the first quarter of this year to 89.2 trillion won in the second quarter. A slowdown in the rate of DRAM price increases or operating-profit growth is clearly different from a decline in the absolute amount of prices or operating profit. That is why concerns about a “peak-out” reflected in share prices alone cannot establish that the semiconductor industry has reached its peak.
AI is expanding from training into inference and agents, as well as sovereign AI and physical AI. As a result, demand is increasing not only for HBM but also for server DRAM, low-power DRAM, and enterprise SSDs. A continuing shortage of DRAM wafers for HBM is expected: in 2027, monthly demand of 680,000 wafers against supply of 510,000; in 2028, demand of 850,000 against supply of 650,000–750,000; and in 2029, demand of 1.1 million against supply of 800,000–1 million. It takes time for expanded fab capacity to come online, and HBM stacking and testing processes also constrain supply. Even as capacity expansions in Pyeongtaek, Yongin, Cheongju, and elsewhere proceed in stages, South Korea's memory industry is expected to see continued growth in revenue and operating profit through 2031, given expanding AI demand, supply bottlenecks, and a rising share of high-value-added products.
Ultimately, the answer is neither to fear China nor to grow complacent amid the current boom. Samsung Electronics and SK hynix must carry out their road maps for DRAM below 10nm, 4F2 and 3D DRAM, HBM4E, and advanced packaging without disruption, and begin full-scale production validation of 4F2 and yield improvement in the second half of 2029. The government, too, should prepare power and water supplies, along with infrastructure for validating materials, components, and equipment, before fabs begin operating. Semiconductors are a timing-driven industry. The future of South Korea's semiconductor industry depends not on China's announcements, but on how accurately we keep to the technology and mass-production timetable we have set.
Park Jae-geun, president of the Korea Semiconductor and Display Technology Society and distinguished professor at Hanyang University