000660.KS
SK Hynix (Korea)
Summary
What they do:
Manufacture HBM (High Bandwidth Memory), DRAM, and NAND flash — the memory chips physically bonded inside every AI GPU package — as the world's dominant HBM producer with 50-62% market share, sitting at Layer 08 as the primary memory supplier for NVIDIA's AI GPU production.
Why they matter:
Every NVIDIA Blackwell GPU requires 8 HBM3E stacks; every Vera Rubin GPU will need 12 HBM4 stacks. SK Hynix produces the majority of them. HBM allocation has been the binding constraint on GPU shipments — ahead of even GPU die fabrication. The company has sold out its entire 2026 and 2027 HBM production.
Recent performance:
FY2025 revenue KRW 97.1T (+47% YoY), operating profit KRW 47.2T (49% margin, +101% YoY). Q1 2026 projected revenue ~KRW 55.8T (+211% YoY). Stock at ~₩1.13M, all-time high. Market cap ~$541B. First to mass-produce HBM4 (September 2025). UBS projects 70% HBM4 share for NVIDIA Rubin.
Our Verdict
The HBM monopolist of the AI era — 50-62% market share in the single most supply-constrained component in AI infrastructure, with NVIDIA as anchor customer, 49% operating margins, and sold-out production through 2027 — but memory cyclicality and Samsung recovery are the ceiling.
Structural trends
Structural
91
/ 100
Moat
9/10
Dominant HBM producer with 50-62% share, 12-18 month yield lead, NVIDIA primary partner
AI Exp.AI Exposure
Pure Play~70% AI
Play Type
EstablishedAI Growth
~211%
Rel. Value
85
COMPELLINGPriceLIVE
$783.45
+2.99%
Live via Yahoo Finance · refreshes every 5 min
Market Cap
$554.9B
P/E Ratio
N/A
P/S Ratio
8.3x
52W High
$808.97
52W Low
$118.48
52W Chg
561.2%
Beta
1.75
SK Hynix is the world's dominant producer of High Bandwidth Memory — the stacked DRAM chips that sit inside every AI GPU package. The company operates massive fabrication plants in Icheon and Cheongju, South Korea, where billions of transistors are etched onto silicon wafers that become the memory chips AI depends on.
Open an NVIDIA Blackwell GPU package. The GPU die sits in the center. Surrounding it, bonded onto the same silicon interposer, are eight tall, thin rectangles — those are HBM stacks, and the majority are made by SK Hynix. Each stack is a tower of memory dies, 8-12 layers high, connected by thousands of through-silicon vias (TSVs). The entire stack is about 1mm tall and delivers over 1 terabyte per second of bandwidth to the GPU — 10x faster than conventional DDR5. A single HBM3E stack costs $100-$150; eight stacks per GPU means $800-$1,200 in memory cost per GPU.
The financial transformation is staggering. FY2025 revenue hit KRW 97.1 trillion (+47% YoY) with 49% operating margins — record results. Q1 2026 is projected at KRW 55.8T, implying ~211% YoY growth. The HBM business is the growth engine: SK Hynix began mass production of HBM4 in September 2025 — the first in the industry — and has been tapped as a primary supplier for NVIDIA's Vera Rubin platform. Both Samsung and SK Hynix reportedly planned ~20% HBM3E price increases for 2026 as demand from NVIDIA and ASIC designers overwhelms supply.
SK Hynix trades on the Korea Exchange (000660.KS) — not directly accessible on US exchanges. Investors need Korean market brokerage access or ETF exposure. The stock hit an all-time high of ₩1.13M in April 2026, up ~500% from its 52-week low. Market cap is approximately $541B.
Supply Chain Dependencies
Upstream Suppliers
The Catch
Memory is the most cyclical business in semiconductors. SK Hynix has risen ~500% from its 52-week low on the most powerful HBM supercycle in history — 49% operating margins, sold-out production, NVIDIA as anchor customer. But every memory supercycle ends. Samsung is investing tens of billions to close the HBM yield gap. Micron's share is growing from 13% to 24%. When the cycle turns — whether from AI capex slowdown, inventory builds, or competitive yield convergence — SK Hynix's margins compress from 49% to 25-30% within 3-4 quarters, earnings collapse, and the stock gives back a significant portion of its gains. The 500% move means the absolute dollar downside is enormous. The HBM content multiplier (8→12 stacks per GPU) provides a structural floor that previous memory cycles lacked, but it delays the peak rather than eliminating cyclicality.
If They Win
If SK Hynix maintains its HBM yield lead through HBM4, HBM5, and beyond — if each new generation widens rather than narrows the competitive gap — SK Hynix becomes the TSMC of memory: the company that controls the most critical component in AI infrastructure with near-monopoly positioning. HBM TAM reaches $100B+ by 2028. SK Hynix captures 55%+ with 45-50% operating margins sustained across the cycle. Revenue exceeds KRW 200T. Market cap approaches $1T. The memory cyclicality that has defined the industry for decades is partially broken because HBM supply remains structurally short as AI compute demand compounds faster than fab capacity can be built. SK Hynix becomes the second most important company in the AI hardware supply chain after TSMC — the memory monopolist whose production schedule determines when the world's AI infrastructure gets built.
Others in Memory for the Server
Not financial advice. All scores generated via AI algorithms using public data.