How Korea’s chip supply chain fits together: wafers to AI servers

This guide is for expats, investors, and Korea-watchers who keep seeing “Samsung” and “SK Hynix” in AI headlines but want to understand where those companies sit inside a longer production chain. The one-sentence answer: Korea dominates the memory and high-bandwidth memory (HBM) links of the global chip supply chain, holds a weak position in contract chip-making (foundry), and relies on Japan and the West for the upstream inputs—materials and equipment—that make everything else possible.

How it works

A finished AI server contains processors, memory, a circuit board, and a metal chassis. The memory alone passes through at least six distinct production steps before it leaves Korea.

Step 1: Raw materials and chemicals

Every chip starts as a silicon wafer—a thin, polished disc of ultra-pure silicon. SK Siltron is Korea’s sole wafer maker and ranks third globally in 12-inch wafer production. Alongside silicon, chipmakers consume enormous quantities of specialty chemicals: photoresist (a light-sensitive coating used to draw circuit patterns), hydrogen fluoride (HF, used to etch away unwanted material), and fluorinated polyimide (used in display and packaging layers). Japan dominates global supply of all three. As of 2024, Korea had reduced its dependence on Japanese photoresist from 93.2 percent in 2018 to 65.4 percent, a significant improvement but still a meaningful reliance. Dry hydrofluoric acid—critical for advanced lithography—remains largely supplied by Japan. Korean materials companies such as Dongjin Semichem and SK Materials are expanding domestic output, but full substitution has not been achieved.

Step 2: Equipment

Chipmakers use lithography machines to project circuit patterns onto wafers. The most advanced machines—EUV (extreme ultraviolet) lithography tools—are made exclusively by ASML, a Dutch company. Without EUV, no chip manufacturer can reach the leading nodes (3 nm and below) that power today’s AI accelerators. Korea’s own equipment makers—SEMES (a Samsung subsidiary), Wonik IPS, Eugene Technology, and Jusung Engineering—supply deposition and etch tools, a real competitive asset. But they cannot yet build lithography machines, the single most capital-intensive link in the chain. That gap leaves Korea dependent on ASML for its most critical tool.

Step 3: Memory fabrication

This is Korea’s clear strength. Samsung Electronics and SK Hynix together control roughly 60 to 70 percent of global DRAM (dynamic random-access memory, the standard working memory in any computer) and 45 to 50 percent of NAND flash (the storage memory in phones and SSDs). In Q1 2026, Samsung led DRAM revenue with a 38 percent share; SK Hynix followed with 29 percent (Counterpoint Research, Q1 2026).

Step 4: High-bandwidth memory (HBM)

HBM is a specialized form of DRAM stacked in multiple layers and connected by tiny vertical channels called TSVs (through-silicon vias). It moves data far faster than conventional DRAM, which is why every Nvidia AI chip uses it. SK Hynix held 57 to 62 percent of global HBM revenue in early 2026; Samsung was ramping its HBM4 product toward mass production. The two Korean companies together account for roughly 65 percent of global HBM revenue (Korea Herald, 2026).

Step 5: Foundry (contract chip-making)

A foundry manufactures chips designed by other companies—Nvidia, Apple, AMD—without designing chips itself. TSMC of Taiwan captured 69.9 percent of the global foundry market in 2025; Samsung Foundry’s share stood at approximately 7 percent, a significant gap (Counterpoint Research, 2025). Samsung’s foundry division is working on its SF2 (2-nanometer) process but has faced yield challenges that have kept major customers with TSMC. DB HiTek handles mature-node foundry work for analog and display driver chips. This remains Korea’s most significant structural weakness in the supply chain.

Step 6: Packaging and assembly

After wafers are cut into individual chips (dies), they must be packaged—connected to substrates and sealed for use in a final product. Advanced packaging is increasingly important for AI chips because it allows multiple dies to be combined in a single package. Samsung Foundry offers FOWLP and FOPLP (fan-out wafer/panel level packaging) technologies. SK Hynix uses TSV and hybrid bonding for HBM stacking. Amkor Technology Korea (the Korean operations of a U.S.-headquartered firm) and Nepes provide OSAT (outsourced semiconductor assembly and test) services. Korea is competitive here, though Taiwan and the U.S. hold advantages in the most advanced CoWoS (chip-on-wafer-on-substrate) packaging used by TSMC for Nvidia’s highest-end GPUs.

Step 7: AI servers

The final link is the AI server itself. Korean memory sits inside Nvidia’s AI accelerators, which in turn go into servers built by companies such as Dell, Supermicro, and Taiwanese ODMs (original design manufacturers). Korea does not have a major domestic AI server assembler, but its memory is the critical content of every major AI training cluster. In 2025, Nvidia and SK Hynix locked in long-term HBM supply agreements reported at over $500 billion combined with Samsung, through 2030 (Silicon Analysts, July 2026).

Why Korea does it this way

Korea’s semiconductor industry was deliberately built by the government and chaebol (large family-controlled conglomerates) from the 1980s onward. Samsung entered DRAM production in 1983 and pursued a “counter-cyclical” investment strategy: it spent heavily during downturns, when competitors cut back, gaining capacity share each time prices recovered. SK Hynix followed a similar path after multiple ownership changes and financial crises. By concentrating capital in a small number of very large fabs (fabrication plants), Korea achieved scale economies that smaller rivals could not match. The country chose memory over logic chips (CPUs, GPUs) because memory is more commoditized and lends itself to mass production—a better fit for the chaebol model of capital-intensive, process-driven manufacturing. The foundry gap is a product of this history: building a competitive foundry requires a different customer-service culture and a broader portfolio of process nodes, areas where Samsung has struggled against the more specialized TSMC.

The numbers that matter

What it means for you

Expats and workers in the semiconductor industry: The cluster of fabs and R&D centers is concentrated in Gyeonggi Province—Hwaseong, Icheon, and Pyeongtaek are Samsung and SK Hynix home bases. If you are considering a role in this sector, proximity to these cities is practically relevant. The industry employs tens of thousands of engineers, and the K-CHIPS Act incentives have accelerated capex (capital expenditure) expansion plans.

Investors and Korea-watchers: Korea’s chip supply chain is now structurally linked to the global AI build-out. HBM is the current chokepoint—without it, Nvidia cannot ship its highest-end accelerators. Any supply disruption at SK Hynix or Samsung’s HBM lines has direct global consequences. Monitor quarterly HBM market-share reports from Counterpoint Research and TrendForce for early signals. The export control environment—especially U.S. rules restricting advanced chips and equipment from reaching China—is also worth tracking, as it reshapes where Korean companies can sell and from whom they can buy equipment.

Foreign residents generally: Korea’s semiconductor exports accounted for roughly 24 percent of total goods exports in 2025. When chip prices fall sharply, the won (Korea’s currency, symbol: KRW) tends to weaken relative to the dollar, which affects purchasing power for people holding foreign-currency income.

Regarding the K-CHIPS incentives: These apply to Korean legal entities investing in approved facility categories inside Korea. Foreign companies establishing Korean subsidiaries can access the same credits under the same rules. The 2026 production-volume credit is new and differs from the earlier investment-only credit—consult the National Tax Service (nts.go.kr) or a licensed Korean tax accountant for current eligibility criteria before relying on specific figures.

FAQ

Does Korea make the AI chips that power ChatGPT and similar tools?

Korea makes the memory those chips depend on, not the chips themselves. The processors (GPUs) in AI data centers are designed by Nvidia (U.S.) and manufactured by TSMC (Taiwan). But every one of those Nvidia GPUs uses HBM made almost entirely by SK Hynix or Samsung. Without Korean memory, the AI hardware supply chain would stop.

What would happen if Japan restricted chemical exports to Korea again?

Japan imposed licensing controls on three key chemicals in 2019, causing short-term supply uncertainty. Since then, Korea has diversified some sources and localized some production, but dry hydrofluoric acid and advanced photoresists remain areas of reliance. A severe restriction would slow advanced fab output within weeks, given low inventory buffers for specialty chemicals. Korea continues investing in domestic materials capability to reduce this risk, but full independence has not been achieved as of 2026.

Why can’t Samsung simply catch TSMC in foundry?

Catching up requires more than capital. TSMC has decades of experience serving hundreds of chip designers, a deeply optimized process library, and a reputation for high yields (the share of working chips per wafer) that attracts the most demanding customers. Samsung Foundry serves some external clients but also competes with them through its own chip design division, a conflict of interest that makes some fabless companies reluctant to use it. The 2025 foundry share figures—69.9 percent for TSMC versus roughly 7 percent for Samsung—reflect this structural gap, not just a temporary technology lag.

Browse all Le Won guides