Korean Semiconductor Supply Chain Explained (2026)
HBM, AI accelerators, foundry — how a single NVIDIA GPU sale flows back into the entire Korean chip ecosystem
TL;DR
- HBM (High Bandwidth Memory): SK Hynix leads globally (50%+ share), Samsung catching up, Micron a distant third
- NVIDIA H100/H200/B100 dependency: Each GPU stacks 6-8 HBM units → revenue flows directly to SK Hynix
- Foundry race: Samsung vs TSMC — the 3nm and 2nm yield gap is the critical battleground
- Equipment & materials beneficiaries: Hanmi Semiconductor (HBM bonders), Wonik IPS (deposition), Dongjin Semichem (photoresists)
Why Korea Sits at the Center of AI Silicon
AI accelerators (NVIDIA H100, AMD MI300, Google TPU, etc.) are packages combining a GPU die and HBM stacks. The GPU dies come from TSMC, but HBM is effectively monopolized by SK Hynix. Every AI GPU shipment, in other words, becomes Korean semiconductor revenue.
Anatomy of a Single AI GPU
- GPU die: Designed by NVIDIA/AMD, fabricated by TSMC at 4nm/3nm
- HBM stack: H100 uses 6, H200 uses 6, B100 uses 8 → SK Hynix (#1), Samsung (#2)
- CoWoS packaging: TSMC monopoly — the bottleneck constraint
- Substrate & interposer: Some Korean substrate makers (Samsung Electro-Mechanics, LG Innotek) supply
HBM Market: SK Hynix's Runaway Lead
HBM is regular DRAM stacked 8-12 layers high and connected via TSV (Through-Silicon Via). It determines AI GPU memory bandwidth and sells at 5x+ the price of regular DRAM.
Current share (2024 estimate): SK Hynix 50-55%, Samsung 30-35%, Micron 10-15%
Why SK Hynix Dominates
- HBM3E 12-layer mass production lead: First to qualify for NVIDIA H200/B100
- MR-MUF packaging process: Better thermal performance and yield than competitors
- NVIDIA's primary supplier: NVIDIA effectively pre-bought SK's entire HBM capacity in 2023-2024
- HBM4 roadmap: 2025 mass production, 2026 NVIDIA Rubin generation likely uses SK
Samsung's Catch-Up Scenario
Samsung qualified for NVIDIA HBM3E 8-layer in 2024 and is undergoing 12-layer evaluation. For HBM4, it has heavily expanded packaging lines to close the gap. Short-term, however, thermal and yield issues have prevented Samsung from reaching Tier-1 status with NVIDIA.
Foundry: Samsung vs TSMC Gap
TSMC dominates foundry (60%+ share), with Samsung at #2 (~15%) and Intel chasing as #3. AI-era GPU and AP fabrication is effectively a TSMC monopoly.
- 3nm node: TSMC mass production from 2022; Samsung started 2022 with GAA (Gate-All-Around) but yield issues lost NVIDIA/AMD
- 2nm production race: TSMC H2 2025 vs Samsung 2025 — simultaneous entry but TSMC has the customer base
- Samsung's differentiation: Earlier GAA adoption, diverse customer IP pool (Qualcomm, Tesla D1)
- Texas Taylor fab: Samsung's $17B investment, 2026 startup — backed by US CHIPS Act subsidies
Equipment & Materials Beneficiaries
The most interesting part of the Korean semi supply chain is the small-cap equipment and materials suppliers feeding Samsung and SK. They often grow faster than the megacaps during AI cycles.
- Hanmi Semiconductor (042700): Global #1 in TC bonders for HBM stacking — direct beneficiary of SK HBM capacity expansion
- Wonik IPS (240810): CVD/ALD deposition equipment — supplies Samsung and SK memory lines
- HPSP (403870): Monopoly on high-pressure hydrogen annealing — essential for gate dielectric, also used by TSMC
- Soulbrain (357780): HF (hydrofluoric acid) and etchants — symbol of Korea's de-Japan supply diversification post-2019
- Dongjin Semichem (005290): Photoresists — domestic EUV PR development underway
- Leeno Industrial (058470): Global #1 in semiconductor probe card pins
Risk Factors
- US-China chip war: US export controls on advanced equipment to China constrain Samsung/SK fabs in Xi'an and Wuxi
- HBM competition: Micron accelerating HBM4 entry in 2025; as a US company it may receive political preference
- Memory cycle volatility: AI HBM demand is steady, but conventional DRAM/NAND prices remain volatile
- Single-source supply chain: Dependency on Japan EUV photoresists and Netherlands ASML EUV machines
- Taiwan geopolitics: A TSMC disruption would halt NVIDIA/AMD GPU production — Korean memory revenue takes the hit too
Investment Scenarios
- Megacap core: SK Hynix (000660) for direct HBM exposure; Samsung (005930) for catch-up multiple expansion potential
- Equipment beta plays: Hanmi Semiconductor, HPSP, Wonik IPS — react faster to HBM/EUV cycles
- Foundry recovery bet: Samsung wins NVIDIA/AMD orders at 2nm GAA → multiple re-rating
- Consumer memory rebound: 2026 PC/mobile refresh cycle could lift both SK and Samsung
- ETF route: KODEX Semiconductor, TIGER Fn Semiconductor TOP10 for diversified exposure
Frequently Asked Questions
Why did SK Hynix become the HBM #1?
The decisive factor was packaging technology. SK was first to mass-produce HBM1 in 2013 and pioneered the MR-MUF (Mass Reflow Molded Underfill) process, which delivers better thermal dissipation and yield than Samsung's NCF (Non-Conductive Film) approach. NVIDIA's H100 and H200 qualifications in 2022-2023 went first to SK's 12-layer HBM3E, locking in dominant supplier status for essentially all NVIDIA AI GPUs.
Why did Hanmi Semiconductor surge so quickly?
Hanmi is the global #1 in TC Bonders (Thermo-Compression Bonders), the equipment used for stacking HBM 12 layers. As SK Hynix expanded HBM capacity, Hanmi received proportional bonder orders. When Samsung also began HBM3E production and adopted Hanmi bonders, the multiple re-rated again. Hanmi is the cleanest pure-play on HBM capacity expansion.
What if Samsung never gets NVIDIA HBM qualification?
Samsung is still 80%+ general DRAM and NAND, so single qualification failure doesn't crater the whole company. But HBM is rapidly growing to 30-40% of memory operating profit — without NVIDIA Tier-1 status, the operating profit gap with SK widens. Samsung can partially compensate with AMD, Google TPU, and Amazon Trainium customers.
If TSMC goes down, do Korean semis go down too?
Yes, directly. NVIDIA and AMD can't make GPUs without TSMC, so SK and Samsung HBM demand evaporates with them. Samsung also outsources some Exynos AP production to TSMC and depends on TSMC for modem and connectivity chips. Taiwan geopolitical risk is the single biggest external variable for Korean semiconductors.
What happens after the AI semiconductor cycle ends?
Short-term correction is unavoidable, but structural support remains: 2026 PC/mobile refresh cycle, automotive memory growth, and data-center conventional DRAM demand. Plus, on-device AI puts NPUs in phones and laptops, driving LPDDR5X/LPDDR6 demand. The current cycle is less a one-off boom and more a step-change in the memory cycle's upper bound.
How can I get this exposure via ETFs?
Korean-listed ETFs include KODEX Semiconductor (091160) and TIGER Fn Semiconductor TOP10 (396500), heavily weighted toward SK Hynix and Samsung but with some equipment-supplier exposure. US-listed SOXX or SMH have heavy NVIDIA/TSMC/AMD weights with limited Korea exposure. For direct Korea exposure, EWY (MSCI Korea) or individual ADRs are better.
This analysis provides general market information, not personalized investment advice. Consider semiconductor cycle, FX, and geopolitical risks before making investment decisions.