SK Hynix
Compare SK Hynix…SK Hynix Inc. manufactures and sells semiconductor products including DRAM, NAND flash, and High Bandwidth Memory worldwide.
SectorsAI Chips
Register for a premium account to gain access to Sterling AI.
Get StartedThings you can ask Sterling:
SK HynixSK Hynix Inc. manufactures and sells semiconductor products including DRAM, NAND flash, and High Bandwidth Memory worldwide.
| Generation / Date | Capacity / Bandwidth | Customer Wins |
|---|---|---|
| HBM2 (2018-2020) | 8 GB / ~300 GB/s | NVIDIA V100 + A100; AMD Instinct MI50/MI60; initial deployment in HPC + early AI training workloads |
| HBM2E (2020-2022) | 16 GB / ~460 GB/s | NVIDIA A100 80GB variant; SK Hynix + Samsung roughly equal share |
| HBM3 (2022-2024) | 24 GB / ~819 GB/s | NVIDIA H100 first volume customer (SK Hynix was sole supplier at launch March 2022; Samsung qualified Q3 2023, Micron qualified Q4 2023); AMD MI300X uses HBM3 8-stack with 192GB total |
| HBM3E (2024-2025) | 36 GB / ~1.2 TB/s | NVIDIA H200 (SK Hynix as primary supplier; Micron as second source Q4 2024; Samsung struggled to qualify until late 2024); NVIDIA B100 + B200 use 192GB HBM3E across 8 stacks |
| HBM3E 12-Hi (2024-2025) | 48 GB / ~1.2 TB/s | 12-die-stack variant; SK Hynix shipped first in March 2024; targeted at top-bin NVIDIA + AMD + Broadcom AI accelerators; capacity-constrained through 2025 |
| HBM4 (2026 target) | 48-64 GB / ~2.0 TB/s | SK Hynix targets 2026 production launch; NVIDIA Rubin + AMD MI400 + Broadcom + Google TPU v7 all expected as anchor customers; TSMC + Samsung Foundry both producing the base die |
| HBM4E (2027 target) | Higher capacity + bandwidth | Roadmap disclosed at SK Hynix CES 2025 + investor day; targets HBM4-tier customers + emerging custom-ASIC partnerships |
| Market share trajectory (TrendForce) | Reference | 2022: SK Hynix 50% + Samsung 40% + Micron 10%; 2024: SK Hynix 52% + Samsung 38% + Micron 10%; 2025E: SK Hynix 50% + Samsung 33% + Micron 17% (Micron gaining share) |
| HBM revenue trajectory (financial context) | Reference | SK Hynix 2024 HBM revenue ~$13B; 2025E ~$28B (per Q3 2025 earnings call commentary); HBM contributes ~50% of memory operating profit through 2025 despite being ~25-30% of memory revenue (high margin vs DRAM commodity) |
| Editorial. Why SK Hynix lead matters | Strategic context | Memory was commoditized for decades. HBM broke the commodity model: NVIDIA + AMD + Broadcom prefer the supplier that ships first because designing accelerator + memory together is engineering-intensive. SK Hynix's HBM3 first-mover advantage cascaded to HBM3E and is structurally hard for Samsung + Micron to close in HBM4. The customer-relationship moat is the SK Hynix equity thesis |
Q1 2023-Q1 2026
$000660.KS
World's second-largest memory chipmaker and leading supplier of High Bandwidth Memory (HBM) essential for AI training GPUs. ~$35B annual revenue company supplying the majority of NVIDIA's HBM3E chips, giving it a dominant position in the AI memory market. Investing heavily in next-gen HBM4 and advanced DRAM packaging to maintain its lead in AI-critical memory solutions.
The HBM kingpin — supplies the memory that makes AI GPUs actually work.
HBM3E
High Bandwidth MemoryPrimary supplier to NVIDIA for H100/B200
HBM4
Next-Gen HBMIn development for 2026 AI accelerators
DDR5
Server DRAMHigh-performance server memory
NAND Flash
StorageEnterprise SSD and data center storage
Market Share
~50% of HBM market (dominant), ~28% of total DRAM
Competitive Moat
First-mover in HBM3E, exclusive NVIDIA supply relationship, process technology lead
Key Risk
Revenue highly concentrated in a cyclical memory market; Samsung and Micron catching up
Without SK Hynix's HBM chips, NVIDIA's GPUs can't function. It's that simple. SK Hynix controls half the HBM market and is NVIDIA's primary memory supplier. They're the hidden bottleneck of the entire AI buildout.
Key Milestones
Founded as Hyundai Electronics in Icheon, South Korea by Hyundai Group founder Chung Ju-yung; initial focus on DRAM manufacturing as Korea's second memory player behind Samsung.
Renamed Hyundai Electronics Industries; consumer electronics arm and memory unit operated separately under the Hyundai chaebol umbrella through the 1997 Asian Financial Crisis.
Forced merger with rival LG Semicon under Korean government Big Deal industrial policy; combined entity briefly became world’s largest DRAM maker by capacity.
Hyundai Electronics renamed Hynix Semiconductor March 26 amid debt restructuring; Korean creditor banks took control after the 1997 financial crisis nearly bankrupted the company.
Acquired by SK Group for B February 14; renamed SK Hynix and entered a 14-year run of stable strategic capex under Chairman Chey Tae-won.
Released the world's first HBM (High Bandwidth Memory) device, co-developed with AMD on TSV (through-silicon-via) packaging; foundation of all modern AI accelerators a decade later.
Joined Bain-led consortium taking 49% stake in Toshiba Memory (now Kioxia) for ~B; secured non-controlling NAND access to balance Samsung-dominated supply chain.
Announced B M16 chip-cluster investment in Yongin, Korea; the cluster ultimately housed leading-edge DRAM and HBM production lines coming online through 2027.
Announced B acquisition of Intel’s NAND/SSD business October 19; closed in two phases through 2025 with Solidigm as standalone subsidiary running ex-Intel Dalian fab.
Began industry-first HBM3 mass production in June, sole supplier to NVIDIA H100 at launch; locked in 5+ years of dominant AI memory share and became NVIDIA's most strategic supplier.
Granted one-year US license waiver for advanced equipment imports into its Wuxi DRAM fab, easing immediate concerns about strandable Chinese memory capacity.
Received indefinite VEU authorization from US Commerce, securing Wuxi DRAM fab equipment access long-term; protected ~40% of SK Hynix DRAM capacity that resides in China.
Began industry-first HBM3e 8-layer mass production for NVIDIA H200 ramp; SK Hynix supplied roughly 80% of NVIDIA's HBM3e through 2024 alongside Micron's 24GB units.
Began HBM3e 12-layer mass production at 36GB/9.6Gbps in September, six months after the 8-layer ramp; locked in 12-layer supply for Blackwell B200 and B300.
December 2 BIS rules added country-wide HBM controls, restricting any HBM exports to China above 2 GB/s/mm2; effectively cut SK Hynix off from Chinese AI accelerator customers.
Delivered industry-first 12-layer HBM4 samples to NVIDIA, ahead of plan and gating Rubin platform launch; cemented projected ~70% share of Rubin HBM4 demand.
Delivered final paid HBM4 samples to NVIDIA targeting February 2026 mass production for Rubin; secured estimated 70% of NVIDIA HBM4 supply allocation.
Began HBM4 mass production for NVIDIA Rubin; shipped 12-layer stacks at 9.6Gbps with 2TB/s bandwidth per stack, six months ahead of original 2026 H2 target.
