SK Hynix (000660)

[Business Overview] SK Hynix is a leading memory-semiconductor specialist with a portfolio spanning DRAM, NAND Flash, and HBM (High Bandwidth Memory), supplying global hyperscalers, smartphone OEMs, and PC vendors. Following its 2020 acquisition of Intel's NAND business (now Solidigm), SK Hynix is the world's #2 NAND supplier alongside its long-standing #2 position in DRAM. The company's HBM3 and HBM3E lines have made it the dominant supplier of memory for NVIDIA's AI GPUs (H100, H200, B100/B200), positioning SK Hynix as the single most-leveraged Korean equity to the AI infrastructure build-out. [Core Competitiveness] SK Hynix's competitive moat is its early lead in HBM stacking technology (TSV, MR-MUF), which translates into a manufacturing yield advantage and a multi-quarter time-to-market edge over both Samsung and Micron in the highest-margin slice of the DRAM market. Roughly half of group operating profit in 2024–2025 came from HBM, despite HBM accounting for a much smaller share of bit shipments — illustrating the unusually wide margin spread on AI memory. [Earnings Analysis] After the 2023 memory trough (operating loss of roughly KRW 7.7 trillion), SK Hynix swung to record profits as HBM3 ramped through 2024 and HBM3E supplied NVIDIA's H200/B100. The company's 2025 earnings benefited from sustained HBM ASP strength and a recovery in conventional DRAM/NAND pricing. [Financial Health] HBM CapEx and the Solidigm consolidation pushed leverage higher through 2023, but improving cash generation and HBM gross-margin tailwinds materially deleveraged the balance sheet through 2024–2025. Continued HBM pre-orders from hyperscalers further reduce inventory risk relative to the pre-2024 cyclical pattern. [Outlook] HBM3E 12-Hi shipments and HBM4 development for 2026/2027 GPU sockets are the dominant earnings drivers. Risks include NVIDIA roadmap shifts, Samsung's HBM3E qualification cadence narrowing the gap, and broader spending discipline among hyperscalers. The Solidigm enterprise QLC SSD business should benefit as data centers ingest growing AI training/inference data volumes.