Haelangdal Insight: AI Infrastructure Value Chain
Investment opportunities across the full AI training and inference infrastructure value chain, from semiconductors to power, cooling, and networking. Analyzing 12 sub-themes spanning GPU/HBM to transformers and cooling from a unified perspective.
Key Drivers
- Agentic triple power chain (sub-agent token/memory I/O 10-100× + CPO 2026 commercial + BYOP ~14GW)
- Samsung Electronics Q1 OP KRW 57.2T (YoY +755%, OPM 43%) — Memory super cycle confirmed
- Big Tech AI Capex surpasses $650B/year (YoY +80%) — locking up production capacity for years
- HBM4 mass production begins (Samsung Feb) — H2 2026 OpenAI Titan chip supply scheduled
- DRAM prices quadrupled in H2 2025, supply shortage forecast through 2030 (SK Group)
Risk Factors
- Samsung Electronics May strike risk (93.1% approval) — 50% Pyeongtaek production impact
- Qatar helium supply 30-38% offline + Hormuz blockade disrupting fab processes
- Valuation in overextended territory
- China export controls expansion + gallium retaliatory restrictions
- Energy cost surge → potential hyperscaler Capex reduction
Catalysts
- NVIDIA Quantum-X H1 / Spectrum-X H2 shipments + GEV quarterly backlog disclosure
- ASML (4/15) and TSMC (4/17) earnings — AI demand and pricing guidance
- Samsung strike early resolution → supply relief rally
- Big Tech capex guidance upgrades
- Data center construction acceleration
Value Chain Analysis
12 sub-themes
AI Accelerators & Fabless
Fabless companies designing GPUs, TPUs, and custom ASICs that are the core of AI training and inference. Direct beneficiaries of big tech AI capex expansion.
AI Memory & HBM
The performance bottleneck for AI accelerators has shifted from compute to memory bandwidth. HBM (High Bandwidth Memory) and DDR5 demand is surging explosively, ushering in a memory super-cycle.
Semiconductor Packaging Value Chain
AI chips are evolving from single dies to integrated packages combining chiplets, HBM, and interposers. 2.5D/3D packaging, hybrid bonding, and test/burn-in are emerging as new bottlenecks.
Power Semiconductors (SiC/GaN)
Critical for data center power efficiency. SiC/GaN power semiconductors reduce losses by 30-50% compared to conventional silicon. Benefiting from the 48V to 800V transition trend.
Transformers & Power Distribution Infrastructure
Data center and renewable energy expansion have made transformers and switchgear critical bottlenecks. Lead times of 2-3 years and constraints from permits and supply chains limit expansion speed itself.
Data Center Power & Cooling
The era of 100kW+ rack power density. Power distribution (PDU/UPS/busway) and cooling (liquid cooling/DLC) infrastructure have emerged as critical bottlenecks in data center construction.
AI Networking & Optical Communications
Data movement costs are as critical as compute in AI clusters. Switch ASICs, optical transceivers, CPO (co-packaged optics), and silicon photonics are the next-generation bottlenecks.
Nuclear Power & SMR
24/7 carbon-free power demand from AI data centers is triggering a nuclear renaissance. Big tech companies are directly signing nuclear PPAs, driving a revaluation of the uranium and nuclear value chain.
Gas Turbines & Power Generation Infrastructure
Surging data center power demand is driving a gas turbine demand explosion. On-site power generation bypassing grid connection delays and power grid EPC/construction demand are also growing in tandem.
Fuel Cells & Clean Energy
A clean alternative for data center on-site power. Fuel cells attract growing big tech interest due to modular scalability, low carbon emissions, and grid independence. However, economic viability remains the key challenge.
Data Center Physical Infrastructure
The physical build-out infrastructure for data centers. Optical fiber (connectivity), enclosures/racks (thermal management), and cables/connectors are essential real-asset infrastructure in AI Factory construction.
Korean AI Infrastructure Beneficiaries
Korean beneficiaries of the global AI infrastructure investment expansion. Korean companies are gaining global competitiveness in the semiconductor ecosystem (memory/foundry) and power infrastructure (distribution/transformers).
Key Related Assets
10 stocks
Top Investment Ideas
View AllARM — A bet on the AI-era compute platform built on CPU IP licensing
ARM is expanding beyond its dominance in smartphone SoCs into data-center Arm AGI CPUs, automotive, and IoT. License revenue accelerating at +29%, the Total Access / Flexible Access structural revenue model, and the new Arm AGI CPU data-center chip are the key variables.
Jeju Semiconductor — The Only Listed Korean Fabless Beneficiary of Memory Big 3's Legacy Exit
As Samsung/SK hynix/Micron accelerate LPDDR4/4X EOL, Jeju Semiconductor sits as the only listed Korean fabless play on the resulting low-density LPDDR supply void. Capital-light fabless model plus a mix shift toward high-margin automotive/IoT revenue drives structural re-rating. 2025Q3 revenue +197% YoY, operating profit +762% YoY confirms the turnaround.
NVIDIA — Undisputed AI Accelerator Leader, B200/B300 Cycle in Full Swing
80%+ market share in AI training/inference GPU market. B200 Blackwell shipments ramping up, driving data center revenue to record highs. Agentic AI inference demand is the next growth catalyst.
Broadcom — Dual Beneficiary of Custom AI ASICs & Networking
Expanding custom AI ASIC design wins including Google TPU and Meta MTIA. Networking semiconductor demand (Jericho3-AI, Tomahawk 5) growing in parallel. AI revenue mix now exceeds 40%.
SK hynix — Global #1 in HBM, Top Beneficiary of the Memory Super Cycle
Absolute #1 with 50%+ global market share in HBM3E. HBM4 mass production ready. Operating profit forecast of KRW 100T+ for 2026. Dominant supply position for NVIDIA-bound HBM continues.
Micron — HBM3E Shipments Ramping, America's Memory Champion
Full-scale HBM3E production driving rapid increase in high-margin product mix. Data center revenue share exceeds 50%. CHIPS Act subsidies funding U.S. manufacturing expansion.