[Edaily Reporter SOYEON KIM ] The long-standing pattern in the memory industry—where large-scale capital investment led to oversupply—is beginning to shift. This is because, alongside the explosive demand for artificial intelligence (AI), factors limiting production capacity on the supply side are growing. Not only has the amount of memory that can be produced from a single wafer decreased due to the advancement of next-generation high-bandwidth memory (HBM), but the time required to secure cutting-edge equipment is also lengthening. Some in the industry are concluding that the boom-and-bust cycle in the memory market has effectively disappeared.

According to industry sources on the 2nd, the combined capital expenditure of SamsungElectronics and SK hynix this year amounts to approximately 120 trillion won. SamsungElectronics’ Device Solutions (DS) division will invest 76.8 trillion won this year—a 62% increase from last year’s 47.4764 trillion won. SK hynix is reportedly expanding its investment from 30.173 trillion won last year to the high 40 trillion won range this year.

Internally, both SamsungElectronics and SK hynix believe that large-scale capacity expansions will not immediately translate into increased supply. SamsungElectronics expects supply and demand to remain tight through 2028, while SK hynix anticipates the same through 2030.

Memory supply constraints are directly linked to changes in flagship products. Technical limitations have grown as highly advanced HBM has taken center stage, replacing the so-called “mass-produced” generic products of the past. As we move toward next-generation HBM4E and HBM5, the area of the semiconductor die increases and the number of metal interconnect layers grows, resulting in less memory volume that can be produced with the same manufacturing capacity. According to U.S.-based Micron, HBM3E consumes about three times as many wafers as producing the same capacity of DDR5. Micron forecasts that HBM4E will exceed four times that amount. As production of high-value-added HBM increases, the number of wafers available for standard DRAM production decreases relatively.

Furthermore, even for the current flagship 12-layer HBM3E product, the yield rate is reportedly only around 50 percent. This stands in stark contrast to general-purpose DRAM, which consistently exceeds 80 percent. Consequently, supply volumes are bound to decrease. A senior industry official stated, “As memory products become more sophisticated, physical constraints on supply growth will be inevitable.”

It also takes more time to secure the equipment needed to handle the most advanced processes. For example, ASML, the world’s largest semiconductor equipment manufacturer, plans to increase its extreme ultraviolet (EUV) lithography systems from 65 this year to 85 next year and 110 by 2028. Even so, this falls far short of customer demand. Semiconductor manufacturers plan to proactively secure production facilities to maintain high profit margins, while adjusting equipment installation and mass production expansion in line with demand.

Kim Hyung-joon, head of the Next-Generation Intelligent Semiconductor Business Division (Professor Emeritus at Seoul National University), said, “Even if large-scale factories are built first, a strategy is needed to flexibly adjust equipment installation and line operations according to market conditions,” adding, “Memory cycles like those of the past are a thing of the past.”