JIL's Take on AI: Korea's AI Leap
South Korea holds a near-monopoly in HBM memory — the component that determines AI training speed. SK Hynix's structural position is a chokepoint with pricing power that persists for at least three years, and Korea's $7B AI investment programme reinforces the advantage.
The global AI infrastructure debate focuses obsessively on GPUs, data centres, and power. It systematically underweights memory. High-Bandwidth Memory — HBM — is the component that determines how fast a GPU can actually process data. Without sufficient HBM bandwidth, the most powerful GPU sits idle. In the HBM market, South Korea holds a position that is close to monopolistic — and this structural advantage has profound implications for investors seeking AI infrastructure exposure.
The HBM Technology Barrier
SK Hynix and Samsung collectively supply approximately 90% of global HBM. SK Hynix alone holds roughly 50% share and has been NVIDIA's primary HBM supplier for its H100, H200, and Blackwell series. HBM manufacturing requires mastery of through-silicon via (TSV) stacking technology — a process that bonds multiple DRAM dies vertically through microscopic copper connections, creating a memory stack that delivers 5-10x the bandwidth of conventional DRAM at comparable power consumption. The yield management for this process has taken SK Hynix a decade to optimise. Competitors including Micron have invested heavily but have struggled to achieve comparable yields at volume.
The demand picture is equally compelling. Every AI training cluster requires HBM, and the hyperscaler AI buildout trajectory projects forward at a compound rate that exceeds SK Hynix's planned capacity additions through 2027. This creates sustained pricing power: SK Hynix is a price-maker in HBM, not a price-taker. The company raised HBM prices by approximately 15% in Q3 2025 with no meaningful customer pushback — the demand inelasticity of constrained AI infrastructure.
Korea's $7 Billion AI Investment Programme
Beyond semiconductor supply, Korea is deploying $7 billion in national AI R&D through its AI One Team initiative, coordinated by ETRI (Electronics and Telecommunications Research Institute) with participation from Samsung, SK, LG, KT, and NAVER. The programme has three objectives: develop Korean-language LLMs competitive with global frontier models; build AI-native infrastructure across Korea's telecoms and internet backbone; and create a pipeline of AI talent sufficient to sustain Korea's technology competitiveness against China and the US.
KT Corporation has deployed commercial AI infrastructure across Korea's nationwide telecoms network, creating an edge computing platform that enables low-latency AI inference for enterprise and government clients. LG AI Research's EXAONE model is deployed across LG's manufacturing and service operations. NAVER's HyperCLOVA X has demonstrated performance competitive with GPT-4 on Korean-language tasks and is being commercialised across NAVER's dominant search, commerce, and financial platforms — creating a revenue-generating AI product that most non-US AI labs have yet to achieve.
The Geopolitical Premium
Unlike Chinese AI investment — which carries US export control risk, sanctions risk, and capital account restriction risk — Korean AI investment benefits from a strongly positive geopolitical backdrop. Korea is a US treaty ally under the Mutual Defense Treaty. The US-Korea bilateral semiconductor partnership, formalised in the CHIPS Act framework, provides both Samsung and SK Hynix with access to US manufacturing subsidies and de-risks their supply chains against any future technology decoupling scenario. Samsung is building a $20 billion wafer fabrication facility in Taylor, Texas; SK Hynix is constructing a $3.87 billion advanced packaging facility in Indiana. These investments reduce the geopolitical risk premium that sophisticated investors must attach to Korean semiconductor exposure.
Investment Positioning
SK Hynix (KRX: 000660) is the most direct and highest-conviction expression of HBM pricing power. The company's revenue and earnings are highly levered to HBM ASP and volume — it is effectively a pure-play on AI infrastructure memory demand. Samsung Electronics (KRX: 005930) provides broader exposure across memory, logic, foundry, and display but at the cost of diversification that dilutes the pure AI semiconductor thesis. For investors who cannot access Korean equities directly, the US-listed ETF KORU provides leveraged Korea equity exposure; EWY provides broader, unlevered exposure.
The key risk to monitor is the pace of Micron Technology's HBM ramp. Micron's HBM3E qualification with NVIDIA's Blackwell platform — expected in H2 2026 — would introduce a third credible supplier and could compress SK Hynix's price premium. However, qualification is not supply displacement: even a successful Micron qualification would likely result in a three-way supply split rather than a meaningful share transfer in the near term. The structural supply deficit in HBM relative to AI demand makes market share competition less relevant than total supply expansion.
