Korea’s Real Semiconductor Battleground Isn’t Memory, but the Gaps Beyond EUV
These days, whenever people talk about semiconductors, all eyes go to HBM. Fair enough. The names carrying Korea’s banner right now are SK hynix and Samsung Electronics. But I’d argue we should be looking at what sits behind that. Korea’s semiconductor profits still come from memory and advanced process nodes, but the industry’s choke point is firmly controlled by EUV lithography and inspection/metrology.
Why does that matter? Because the areas Korea excels in, it really excels in. DRAM/NAND memory design, HBM, and DRAM-class 10nm process technology are all clear domestic strengths. On the other hand, the gap in EUV lithography is 8. Inspection and metrology equipment stands at 5, while EUV photoresists and etching equipment are at 4, and AI accelerators/NPUs are also at 4. In other words, this is not an industry that is equally strong at both the front end and the back end. Put positively, it’s focus and specialization. Put less kindly, the bottlenecks are obvious.
Money Is Made in Memory, but Control Shifts to Equipment and Systems
The semiconductor value chain is long, but the flow of profits is surprisingly simple. It starts with materials and chemicals, moves through equipment, picks up major value in design and manufacturing, passes through packaging/testing and modules/components, and ultimately the final winners are decided at the application/system level. Korea is strong in parts of this chain: some materials, cleaning equipment, memory design, HBM, and advanced DRAM process technology. Representative examples include high-purity hydrogen fluoride from Soulbrain, SK Materials, and Foosung, cleaning equipment from SEMES and KC Tech, and of course the memory giants Samsung Electronics and SK hynix.
The problem is that the greatest bargaining power does not always sit in the segments where Korea is strongest. If there is a gap in AI accelerators/NPUs on the design side, and if Korea remains weak in front-end EUV lithography as well as inspection and metrology equipment, then no matter how well it performs in memory, the industry's overall profit ceiling gets capped. Put simply, Korea may be running the process well, but someone else is still holding the pen that draws the board.
| Stage | Description | Subsegments |
|---|---|---|
| Materials/Chemicals | Raw materials, chemicals, and materials required for semiconductor manufacturing (including materials for HBM, EUV, and advanced packaging processes) | Wafers, photoresists, specialty gases, CMP slurry, targets/precursors, cleaning/etching chemicals, etc. |
| Equipment | Semiconductor manufacturing equipment (including front-end, back-end, and HBM/advanced packaging equipment) | Lithography equipment, deposition equipment, etching equipment, cleaning equipment, thermal processing equipment, ion implantation equipment, etc. |
| Design (Fabless) | Semiconductor design and IP (including AI accelerators, HBM/CXL controllers, and silicon photonics) | AP/mobile chips, AI/NPU, display ICs, power semiconductors (PMIC), analog ICs, CIS/sensors, etc. |
| Manufacturing (Foundry/IDM) | Wafer manufacturing and foundry operations (including HBM, GAA 2nm, and SiC/GaN) | Memory (DRAM), memory (NAND), foundry, IDM, power semiconductor manufacturing, compound semiconductors |
| Packaging/Testing | Back-end packaging and testing (HBM, CoWoS, hybrid bonding, glass substrates) | Substrates, bumps/solder balls, OSAT, advanced packaging, test equipment, bonding equipment |
| Modules/Components | Semiconductor modules and components (including CXL memory and PCIe 5/6 SSDs) | Memory modules, CXL memory/memory expanders, SSDs, MLCC/passive components, PCBs, camera modules, etc. |
| Applications/Systems | End applications and systems (including AI data centers, on-device AI, and physical AI) | Smartphones, PCs/laptops, servers/data centers, automotive semiconductors, on-device AI/edge AI, AR/VR/XR (spatial computing), etc. |
What Korea does well is world-class. What it doesn’t have, it really doesn’t have.
The term “localizing” should be used very carefully in semiconductors. This is not an industry where everything can simply be localized. What matters is taking a hard look at where Korea has genuinely caught up—and where it has barely even gotten a foothold.
Korea’s technology map shows its strengths heavily concentrated in memory, while its weaknesses are clustered in EUV, inspection/metrology, and non-memory compute. On the strong side, you have high-purity hydrogen fluoride, cleaning equipment, DRAM/NAND memory design, HBM, and 10nm-class DRAM. These are not paper strengths—they are solid assets tied directly to real productivity, yield, and supply stability. The fact that Samsung Electronics and SK hynix appear together across so many categories also matters. It suggests this is not a one-company fluke, but an actual industrial base.
But the weak spots are even more painful. An 8-point gap in EUV lithography tools is not just “large”—it is massive. A 5-point gap in inspection and metrology is no joke either, while EUV photoresists and etch equipment sit at 4, and AI accelerators/NPUs are also at 4. The message from that combination is clear: Korea is missing the core tools for process scaling, the “eyes” that control yield, and the compute brains beyond memory. No matter how well Korea makes HBM, ultimate system power will still tilt toward players like NVIDIA and TSMC. Put bluntly, Korea may be dominant in the areas it excels in, but it is not the rule-maker for the industry as a whole.
Table below: Domestic vs. global TRL (Technology Readiness Level 1–9) by technology, and key companies
| Technology | Domestic | Global | Assessment | Key Companies | ||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Materials | ||||||||||||||||||||||||||||||||||
| EUV photoresist | 5 | 9 | Weak · Gap 4 | Dongjin Semichem, SK Materials Performance | ||||||||||||||||||||||||||||||
| This is the “photosensitive liquid” — essentially light-reactive ink — that chemically dissolves or hardens only in the areas exposed to light when circuit patterns are etched onto a wafer. It works on the same principle as photographic film forming an image when exposed to light. EUV (extreme ultraviolet) has a wavelength of 13.5nm, about 14 times shorter than conventional 193nm light, which means much higher photon energy and fewer photons delivered at a time. That makes it extremely difficult to design a high-sensitivity, low-defect resist that can produce sharp patterns with very little light. More recently, the center of gravity has been shifting from conventional organic polymer resists toward metal oxide resist (MOR), which uses metal oxide particles to overcome those limitations. Three Japanese companies control roughly 80–90% of the global market, so while Korea has largely caught up in ArF resists, EUV resists are still only at the early mass-production validation stage domestically. | ||||||||||||||||||||||||||||||||||
| High-purity hydrogen fluoride (etch gas) | 8 | 9 | Strong · Gap 1 | Soulbrain, SK Materials, Foosung | ||||||||||||||||||||||||||||||
| This is an ultra-high-purity gas or liquid used to chemically remove unwanted portions of a wafer (etching) or to clean it. Purity is everything here: it must reach 11 nines (99.999999999%), meaning impurity levels in the parts-per-billion range, and leading-edge processes increasingly require 12 nines. Even a single particle of metal contamination can create leakage current in a fine circuit and ruin the chip. The function may look simple, but the process of refining, transporting, and storing material at this purity level is itself highly advanced chemical engineering. When Japan imposed export restrictions in 2019, Korean companies succeeded in localizing mass production in a short period, making this one of the flagship “overcoming dependency” cases. The current push is to raise self-sufficiency further in gaseous anhydrous hydrogen fluoride as well. | ||||||||||||||||||||||||||||||||||
| Silicon wafer | 6 | 9 | Caution · Gap 3 | SK Siltron | ||||||||||||||||||||||||||||||
| This is the round silicon substrate on which semiconductors are built—the “canvas” for every chip. Ultra-high-purity silicon extracted from sand is melted, then a seed crystal is dipped into it and slowly pulled upward to grow a single massive monocrystalline ingot (the Czochralski process), which is then sliced thin and polished to a mirror finish. Managing crystal defects is critical, because the atomic lattice has to be perfectly aligned in one direction for electrons to flow smoothly. The larger the diameter—300mm is today’s standard—the more chips can be produced from a single wafer, lowering cost. More recently, SiC (silicon carbide) wafers for EV power semiconductors have emerged as a new growth axis, and SK Siltron entered the field by acquiring DuPont’s SiC business in the U.S. | ||||||||||||||||||||||||||||||||||
| CMP slurry | 6 | 8 | Catching up · Gap 2 | KC Tech, Dongjin Semichem | ||||||||||||||||||||||||||||||
| This is the polishing slurry used to grind the wafer surface flat like a mirror every time circuit layers are stacked. CMP (chemical mechanical polishing) uses a dual action: a liquid containing fine abrasive particles is flowed across the surface while a pad rubs it, mechanically removing material that has first been softened by chemical reaction. If the surface is even slightly uneven, the next circuit layer cannot be patterned sharply with EUV, so CMP is essential at every stage in modern multi-layer chip processes. Different target materials—oxide, copper, tungsten—require different particle types and chemical formulations, which is why there are hundreds of slurry variants. U.S.-based Cabot and Japan’s Fujimi still lead, but Korean firms are gradually replacing imports in selected products and increasing their share. | ||||||||||||||||||||||||||||||||||
| Precursor | 6 | 8 | Catching up · Gap 2 | DNF, Hansol Chemical, Soulbrain | ||||||||||||||||||||||||||||||
| These are specialty compound gases or liquids that serve as the “feedstock” for depositing atomically thin films on wafers. Inside deposition equipment, the precursor decomposes and leaves behind the desired element—silicon, metal, oxygen, and so on—one layer at a time on the wafer surface. In ALD (atomic layer deposition), the precursor is introduced, stopped, and followed by another gas in repeated cycles, allowing films to be built one atomic layer at a time. In that process, precursor purity and reactivity determine film quality. As circuits continue to shrink, thinner and more uniform films are required, driving rapid growth in demand for high-value precursors. Korean chemical companies are expanding into more difficult products, including those for High-k dielectric films. | ||||||||||||||||||||||||||||||||||
| Equipment | ||||||||||||||||||||||||||||||||||
| EUV lithography tool | 1 | 9 | Weak · Gap 8 | (ASML monopoly) Korean participation in parts and metrology | ||||||||||||||||||||||||||||||
| This is the pinnacle of semiconductor manufacturing equipment: the machine that “prints” sub-7nm ultra-fine circuits onto wafers using extreme ultraviolet light. Since EUV has a very short wavelength of 13.5nm, it cannot pass through ordinary glass lenses, so the system uses reflective optics that gather and direct the light entirely with mirrors. The light source itself is generated by firing a powerful laser at molten tin droplets 50,000 times per second to create plasma, and the entire process takes place in a vacuum. The technical difficulty is so extreme that ASML in the Netherlands is the only company in the world that makes these machines. A single unit costs more than KRW 200 billion, while next-generation High-NA models are in the KRW 500 billion range. Intel, Samsung, and TSMC have now begun introducing High-NA systems, and the number of machines a company secures directly determines its leading-edge process capacity. | ||||||||||||||||||||||||||||||||||
| Etch equipment | 5 | 9 | Weak · Gap 4 | SEMES, APTC, HPSP | ||||||||||||||||||||||||||||||
| This is the equipment that physically etches the wafer following the pattern drawn in photoresist. Gas is energized with electricity into plasma—an active state containing ions and electrons—and those reactive particles strike the wafer to etch it with atomic-level precision. The key is the ability to dig straight down vertically without lateral spreading. As NAND has moved beyond 300 layers, the battleground has become high-aspect-ratio etching: drilling deep, narrow holes with aspect ratios above 50:1. Alongside lithography and deposition, etch is one of the largest spending categories in front-end equipment investment. The market is effectively split between Lam Research in the U.S. and Tokyo Electron in Japan. Korean firms are increasing localization by supplying selected dielectric and metal etch chambers. | ||||||||||||||||||||||||||||||||||
| Deposition equipment (CVD/ALD) | 5 | 8 | Caution · Gap 3 | Jusung Engineering, Wonik IPS, TES | ||||||||||||||||||||||||||||||
| This is the equipment used to deposit atomically thin films—insulating layers or metal layers—onto wafers. In CVD (chemical vapor deposition), reactive gases are introduced and a film forms through chemical reaction on the surface. In ALD (atomic layer deposition), gases are introduced alternately so the film is built up one atomic layer at a time. Because ALD relies on a “self-limiting reaction” that deposits only one layer per cycle, thickness can be controlled at the angstrom (0.1nm) level. That makes it increasingly critical as circuits shrink and structures become more complex and three-dimensional. ALD’s biggest strength is step coverage: the ability to coat even the inner walls of deep holes uniformly. Applied Materials in the U.S. leads the field, but this is a promising area where Korean firms are internalizing technology in ALD and hard-mask deposition and gradually gaining share. | ||||||||||||||||||||||||||||||||||
| Cleaning equipment | 7 | 8 | Strong · Gap 1 | SEMES, KC Tech | ||||||||||||||||||||||||||||||
| This equipment removes microscopic dust, metal residues, and oxide films generated during processing. A single wafer goes through hundreds of steps before completion and is cleaned more than 50 times, because a single nanometer-scale particle landing on a circuit can immediately create a defect. The core technologies are wet cleaning with chemicals and precision cleaning that removes particles without damaging the circuit itself. As circuits get finer, the challenge of “removing contamination without shaving away the circuit” becomes harder. Korea has relatively high self-sufficiency here, with SEMES and KC Tech competitive against Japan’s SCREEN and Tokyo Electron. | ||||||||||||||||||||||||||||||||||
| Inspection and metrology equipment | 4 | 9 | Weak · Gap 5 | Nextin, Park Systems, AUROS Technology | ||||||||||||||||||||||||||||||
| These are the “inspection and measurement” tools used to verify whether circuits were fabricated as designed and whether invisible defects are present. As chips are built up through dozens of layers, these tools measure pattern width and alignment error at each step (metrology) and detect tiny particles or pattern defects (inspection). Nanometer-scale defects are identified through light scattering, electron beams, or atomic force microscopy (AFM), and as circuits shrink, the defects shrink too—driving inspection difficulty up exponentially. U.S.-based KLA effectively dominates this market and has the highest operating margins among semiconductor equipment companies. In Korea, Nextin (pattern inspection) and Park Systems (AFM metrology) are trying to carve out niches. | ||||||||||||||||||||||||||||||||||
| Design | ||||||||||||||||||||||||||||||||||
| Memory design (DRAM/NAND) | 9 | 9 | Strong · Gap 0 | Samsung Electronics, SK hynix | ||||||||||||||||||||||||||||||
| This is the technology for designing memory semiconductors that store data. DRAM distinguishes 0s and 1s by storing charge in tiny capacitors, but because the charge leaks away, it must be constantly refreshed and disappears when power is turned off—though it is very fast. NAND stores electrons trapped in a gate, so it retains data even without power, but is relatively slower. The core competitive edge lies in circuit and device design that packs more cells into the same area while preventing interference and leakage between them. Korea is world-class in both design and mass production here. More recently, AI demand has shifted the game beyond simple capacity competition toward high-value designs such as HBM and customized memory. | ||||||||||||||||||||||||||||||||||
| System semiconductors (SoC/AP) | 7 | 9 | Catching up · Gap 2 | Samsung Electronics (Exynos), LX Semicon | ||||||||||||||||||||||||||||||
| These are “brain” semiconductors that integrate multiple functional blocks—CPU, graphics (GPU), communications, AI compute (NPU), and more—onto a single chip, i.e. a System on Chip. The smartphone application processor (AP) is the classic example. While memory consists of regularly repeated identical cells, an SoC involves stitching together different circuits across billions of transistors, making design complexity incomparable. In this field, proven IP blocks and the software ecosystem are what decide the winner. Because the barriers to entering that ecosystem are so high, Korea is relatively weak here compared with memory. Samsung’s Exynos is its in-house AP challenge, while LX Semicon has built competitiveness in display driver ICs (DDIs). | ||||||||||||||||||||||||||||||||||
| AI accelerator / NPU | 5 | 9 | Weak · Gap 4 | Rebellions, FuriosaAI, Sapeon | ||||||||||||||||||||||||||||||
| These chips are designed to rapidly process the massive matrix multiplications that account for most AI computation—multiplying and adding huge numbers of values simultaneously. A general-purpose CPU handles tasks one by one, while NPUs and GPUs boost speed by deploying thousands to tens of thousands of multipliers in parallel. NVIDIA GPUs and their software stack (CUDA) are effectively the standard, which makes market entry extremely difficult. Still, Korean NPU startups are taking aim at inference—the stage where already-trained AI models are run—by optimizing performance per watt. Rebellions and Sapeon have merged under the Rebellions name to gain scale, while FuriosaAI is pushing ahead with performance validation for its own chips. | ||||||||||||||||||||||||||||||||||
| Automotive and power semiconductors | 5 | 8 | Caution · Gap 3 | Telechips, DB HiTek, ABOV Semiconductor | ||||||||||||||||||||||||||||||
| These are chips used either to control vehicles (automotive MCUs and SoCs) or to convert electricity efficiently (power semiconductors). In automotive chips, reliability and zero-defect performance across temperatures from -40°C to 150°C matter far more than the latest leading-edge node, and validation alone can take years. Power semiconductors act as switches that turn voltage on and off. Replacing conventional silicon with SiC (silicon carbide) or GaN (gallium nitride) allows operation at higher voltages and temperatures with lower losses, significantly improving EV range and charging speed. A single electric vehicle uses hundreds of power semiconductors, so demand is surging. European and Japanese players such as Germany’s Infineon are ahead, while Korea is chasing with DB HiTek in foundry and Telechips in automotive SoCs. | ||||||||||||||||||||||||||||||||||
| Manufacturing | ||||||||||||||||||||||||||||||||||
| HBM (High Bandwidth Memory) | 9 | 9 | Strong · Gap 0 |
Why the Lead Shifted from Intel to Samsung Electronics, and Then to NVIDIA and TSMCThe changing roster of top companies in each era isn’t a fad. It’s a record of where the money was made and who controlled the bottlenecks.
From the 1990s to 2016, Intel was the global No. 1, while Samsung Electronics led in Korea. In the centralized computing structure of the PC and server era, Intel’s dominance was overwhelming, while in Korea, Samsung Electronics anchored the memory segment. In 2017~2019, the global top spot shifted to Samsung Electronics, and domestically both Samsung Electronics and SK hynix expanded their presence. It was the period when the full force of the memory supercycle was on display. But the key shift is this: 2020~2024 belonged to NVIDIA and TSMC, and from 2025 to the present, SK hynix has become Korea’s No. 1. This isn’t just a market-cap parlor game. It means semiconductor leadership has moved from general-purpose computing to AI computing, and from standalone chips to system architecture and manufacturing ecosystems. Even within Korea, the fact that SK hynix now comes ahead of Samsung Electronics is symbolic. It tells you that HBM sits on the memory throne in the AI era. But don’t get this twisted. Being strong in HBM is not the same as controlling the system. Korea has moved closer to the center of the board, but it still isn’t the one designing the board itself. If You Keep Staring Only at HBM Numbers, You’ll Miss the Real Bottleneck AgainThe key points to watch from here are clear. First, whether Korea’s strengths—HBM and 10nm-class DRAM—continue to anchor profitability. Second, whether areas where competitiveness has already been proven, such as cleaning equipment and high-purity hydrogen fluoride, can spread into a broader process ecosystem. There’s a reason names like SEMES, KC Tech, Soulbrain, SK Materials, and Foosung keep coming up. In the end, industry runs on connections. But the real risk is relying on a memory upcycle while the gap in EUV lithography tools remains at 8 and the gap in inspection and metrology equipment stays at 5. It becomes even more frustrating when you add an AI accelerator/NPU gap of 4 on top of that. Korea can remain strong in the memory segments it already excels at. But the next wave of leadership will not be won on memory alone. Why do NVIDIA and TSMC still stand at the front? The answer is pretty obvious, isn’t it? Even if you’re bullish on Korean semiconductors, it’s dangerous to look only at memory earnings and fool yourself into thinking you’ve also bought into broader industrial dominance. ※ This article is an analysis compiled from public data and industry materials, and some figures and assessments are estimates. It should not be used as a direct basis for investment decisions. (Written: 2026. Ealexandro) | ||||||||||||||||||||||||||||||
댓글
댓글 쓰기