What this topic means

South Korea's semiconductor industry is the most concentrated of the major chip economies, in two senses. It is concentrated in two companies — Samsung Electronics and SK Hynix — and it is concentrated in one product category, memory.

That concentration is not a weakness in the way it sounds. Memory turned out to be the AI era's second choke point, after leading-edge logic. High-bandwidth memory, or HBM, is what determines how quickly an AI accelerator can be fed with data, and Korean companies supply most of it — SK Hynix alone has held roughly 58% of the HBM market, while Samsung leads overall DRAM at around 38%. Every large AI training cluster built anywhere in the world contains Korean memory. There is no route around it.

Those two numbers also encode the most important competitive fact in Korean semiconductors: the company that leads memory overall is not the company that leads the memory that matters most for AI. Samsung is the larger memory maker; SK Hynix owns the AI-critical segment.

But the concentration is real, and it cuts both ways. Korea is world-leading in memory and comparatively weak in almost everything adjacent: semiconductor manufacturing equipment, advanced packaging, chip design services, and fabless design. Understanding Korea's position means understanding both halves.

Why it matters

For anyone following AI infrastructure, Korea matters for four reasons:

• HBM is a genuine bottleneck. Accelerator supply has repeatedly been limited not by logic wafer capacity but by memory and packaging. A shortage in Korea propagates directly into AI compute availability globally.
• The memory market is a duopoly plus one. Samsung, SK Hynix, and Micron account for essentially all DRAM production. Two of the three are Korean. There is no fourth supplier that could absorb a disruption.
• Memory is cyclical and AI broke the cycle. Memory has historically been a brutal commodity business with boom-and-bust pricing. AI demand for HBM changed the economics, because HBM is sold on long-term contracts at high margins and consumes disproportionate wafer capacity. This is the single most important shift in the memory industry in decades.
• Korea is the clearest example of upstream dependency. Korea's memory dominance rests on Japanese materials, Dutch and American equipment, and Taiwanese packaging capacity. It is a case study in how a country can lead a critical industry while remaining structurally dependent on others.

The two companies

Samsung Electronics

Samsung is the only company in the world with serious scale in memory, logic foundry, and its own end products. Its semiconductor division spans:

• DRAM and NAND flash — historically the world's largest memory producer by revenue, with fabs concentrated at Pyeongtaek and Hwaseong.
• Foundry — contract manufacturing of logic chips for external customers, including advanced nodes using gate-all-around transistor architecture. Samsung is the only credible alternative to TSMC at the leading edge, and it has struggled to convert that into market share.
• System LSI — Samsung's own chip designs, notably Exynos application processors and image sensors.
• Advanced packaging — a comparatively newer effort, built partly to serve HBM and partly to offer integrated logic-plus-memory packages.

Samsung's structural advantage is that it can offer a customer memory, logic, and packaging from one company. Its structural problem is that this same breadth makes it a competitor to many of the customers it wants to serve, and that its foundry business has consistently trailed TSMC on yield reputation at the most advanced nodes.

SK Hynix

SK Hynix is smaller, more focused, and — in the AI era — arguably better positioned. It is a memory company: DRAM, NAND, and HBM, with fabs at Icheon and Cheongju and a NAND business substantially expanded by the acquisition of Intel's flash operation, which now operates as the subsidiary Solidigm.

SK Hynix's decisive move was committing to HBM early, when it was a niche product for graphics and supercomputing with unclear commercial prospects. That bet made it the leading HBM supplier to the dominant AI accelerator vendor, which in turn made it the most important memory company in the AI supply chain.

SK Hynix also sits inside the SK Group alongside SK Telecom, which gives Korea one of the few examples anywhere of a memory maker and an AI service business under common ownership.

HBM: why it changed everything

High-bandwidth memory is DRAM stacked vertically, connected by through-silicon vias, and placed directly next to a processor on the same package. The point is bandwidth: an AI accelerator is useless if it cannot be supplied with data fast enough to keep its compute units busy.

Four things about HBM matter for understanding Korea's position:

1. It consumes far more wafer capacity per bit than conventional DRAM. Yields are lower and the die area is larger. Every wafer moved to HBM is a wafer not making commodity DRAM, which tightens the conventional memory market as a side effect.
2. It is sold differently. HBM is contracted well in advance, often with capacity committed before it is produced. This converted memory from a spot-priced commodity into something closer to a long-term supply relationship — a structural change in how memory companies earn money.
3. It requires advanced packaging, which Korea does not dominate. The finished accelerator package — logic die plus HBM stacks on an interposer — is assembled largely in Taiwan. Korea makes the memory; Taiwan integrates it. Neither can ship an AI accelerator without the other.
4. Stacking is a bonding problem. The equipment that bonds HBM dies together is a specialised niche, and it is the one area where a Korean equipment supplier holds a commanding global position rather than a subordinate one: Hanmi Semiconductor has held upwards of 70% of the market for thermocompression bonders. In an industry where Korea otherwise buys its tools from abroad, this is the exception that proves the rule — Korea built the capability precisely where the customer requirement was Korean.

Each HBM generation raises stack height, bandwidth, and thermal difficulty. The competitive question in Korean memory is simply which company qualifies each new generation with the major accelerator vendors first, and at what yield.

The clusters

Korea's semiconductor manufacturing is geographically dense, in an arc south of Seoul.

• Pyeongtaek and Hwaseong — Samsung's principal memory and foundry campuses.
• Icheon and Cheongju — SK Hynix's main sites.
• Yongin — the site of very large planned expansion by both companies, part of a national strategy to build the world's largest concentrated semiconductor cluster.

The Yongin projects are worth watching for a reason that has nothing to do with chips: power and water. Concentrating that much fab capacity in one area requires new generation capacity, new transmission lines, and large volumes of ultrapure water. Grid connection has been a recurring source of delay. This is the same constraint that shows up in AI data centre siting across the region, and Korea faces it in an acute form because its industrial electricity demand is concentrated and its power system is effectively an island — Korea cannot import electricity from a neighbour.

The gaps

This is the half of the Korean story that gets underreported, and it is what distinguishes this guide from a Samsung-and-SK-Hynix profile.

Equipment

Korea buys the machines that make its chips. Lithography comes from ASML in the Netherlands, with no alternative at the leading edge. Deposition, etch, and metrology come largely from American and Japanese suppliers. Korea has domestic equipment companies — Semes, Samsung's own toolmaking affiliate, plus several independent firms in deposition and etch — but they compete in specific niches rather than across the tool set.

Hanmi Semiconductor's position in HBM bonding, described above, is the single clear exception, and it is not enough to change the overall picture. On the tools that determine whether a fab can run at the leading edge, Korea has no domestic option.

Materials

Korea's materials dependency is the subject of the formative event in modern Korean semiconductor policy. In 2019, Japan imposed export restrictions on three chemicals critical to Korean chip and display manufacturing — a hydrogen fluoride etching gas, photoresist, and a fluorinated polyimide — in the context of a broader political dispute.

The dispute was formally resolved in mid-2023, when Japan reinstated Korea to its preferential trade "White List" and the restrictions were lifted. The measures are not in force today.

Their effect outlasted them entirely. The restrictions never shut down Korean production — what they did was demonstrate that they could, and Korean localisation of specialty chemicals accelerated sharply afterwards. Companies including Soulbrain, Dongjin Semichem, and SK's own materials arm expanded domestic supply. Localisation is real but partial: high-end photoresist in particular remains largely Japanese.

This is the most useful thing to understand about Korean industrial policy. Every semiconductor strategy document Korea has produced since is written in the shadow of 2019 — not because the restrictions still bite, but because they established that supply chain dependency is a political vulnerability, not merely a commercial one. A resolved dispute changed Korean behaviour permanently. If you want one fact that explains why Korea spends public money on materials localisation for inputs it can currently buy freely, it is this one.

Packaging and test

Korea has no equivalent of Taiwan's independent packaging and test sector. Advanced packaging is done in-house by Samsung or contracted out, largely to Taiwanese providers. Given that packaging is now a primary constraint on AI accelerator output, this is Korea's most consequential structural gap.

Fabless and design

This is the mirror image of Taiwan. Taiwan has a world-class fabless sector and no memory industry to speak of; Korea has world-class memory and almost no fabless sector. Korea's chip design companies are small and concentrated in display drivers and analogue parts. Korea has a mature-node foundry industry, but it is modest.

Korea's AI accelerator startups are the interesting counter-trend here, and the reason to care about them is precisely that they would begin to fill this gap. (They are covered in the Korea AI Industry Map.)

Talent

Korean semiconductor executives have described engineering talent shortage as a binding constraint on expansion, driven by demographic decline, competition from medical and other prestige careers, and recruitment of experienced engineers by foreign competitors. Fab capacity can be built in three years; a senior process engineer takes fifteen.

China exposure

Both Korean memory makers operate significant fabs in mainland China — Samsung in NAND, SK Hynix in DRAM and NAND. This was straightforward when it was simply low-cost capacity serving Chinese customers. Under US export controls on advanced semiconductor equipment to China, it became a live strategic problem: those fabs need equipment upgrades to stay competitive, and their ability to receive that equipment depends on US licensing decisions.

For several years both companies held validated end-user (VEU) status, a blanket authorisation that let them ship US-controlled equipment into their Chinese fabs without applying case by case. That status was revoked in 2025. Samsung and SK Hynix now operate those fabs under annual licences instead, and have been reported to be testing domestically produced Chinese tools as a partial substitute.

The shift from blanket authorisation to annual renewal is the whole story in one procedural change. It converts a stable operating environment into one where a multi-billion-dollar asset's viability is reviewed every year by a foreign government. No company can plan capital expenditure on that basis, which is why the practical consequence is a slow structural shift of advanced capacity back toward Korea and, in some cases, toward the United States — expensive, and unavoidable.

It also produces a genuinely awkward second-order effect: Korean fabs evaluating Chinese-made equipment to keep Chinese sites running is precisely the kind of validation that Chinese toolmakers need in order to mature. Export controls intended to slow China's domestic equipment industry are, at the margin, handing it a demanding customer. This is the clearest instance of Korean firms caught between their largest customer market and their security guarantor — a position Korean policy has to manage rather than resolve.

Government policy

Korean semiconductor policy has three consistent features:

• Cluster building. Public support has focused on concentrating capacity, infrastructure, and supplier ecosystems geographically rather than spreading it.
• Tax credits over direct subsidy. Korea has leaned on investment tax credits rather than the direct grants used in Japan and the United States. Under the K-Chips framework, facility investment attracts credits in the range of 15–30% and R&D spending 30–50%, with the higher bands aimed at smaller companies. A production tax credit was added for 2026, which is a meaningful change of instrument: crediting output rather than investment rewards fabs that actually run at volume, not merely those that get built.
• Supply chain localisation. Post-2019 policy consistently targets domestic materials, equipment, and parts capability, with explicit localisation ratios as goals.

Korea's approach differs from Japan's in an instructive way. Japan is using state money to re-enter leading-edge logic manufacturing it had lost. Korea is using state money to defend a leading position it already holds and to reduce the upstream dependencies that make that position fragile. (Compare Japan Semiconductor Ecosystem.)

What to watch

• HBM generation transitions. Which company qualifies each new generation first, at what yield, and with which customers. This is the single most consequential recurring event in Korean semiconductors.
• Samsung foundry's leading-edge customers. Samsung's foundry problem is credibility with large external customers. A major win would change the competitive picture at the leading edge; continued absence of one confirms a TSMC monopoly.
• Whether commodity DRAM stays tight. Because HBM eats wafer capacity, conventional memory pricing is now partly a function of AI demand. This links Korean earnings to AI capital spending in a way it never was before.
• Yongin's power and water timeline. Grid and water infrastructure, not fab construction, is the plausible constraint.
• Materials localisation progress. Specifically high-end photoresist, the least localised of the 2019 items.
• China fab licence renewals. With VEU status gone, each annual renewal is now a discrete event with real consequences. Watch the renewals themselves, and watch how far Chinese-made equipment gets qualified in those fabs as a hedge.
• Any move into packaging. If Korea decides its packaging gap is strategic rather than acceptable, that would be a significant shift, and it would put Korea into direct competition with Taiwanese providers it currently depends on.

Related guides

Korea AI Industry Map covers the model builders, the AI chip startups, and Korean AI policy. AI Chips and Hardware in Asia places Korean memory inside the full regional stack alongside Taiwanese logic and Japanese materials — read that one if you want the interlocking picture rather than the country picture. Japan Semiconductor Ecosystem and China Semiconductor Ecosystem cover the two neighbours Korea is most dependent on and most exposed to respectively.

Last updated: August 2026