SK Hynix

SK Hynix is a South Korean semiconductor company and one of the world's largest manufacturers of dynamic random-access memory (DRAM). It has become the single most important supplier to the AI industry through its dominance of high-bandwidth memory (HBM), the vertically stacked DRAM that sits next to AI accelerators and determines how fast a GPU can feed data to its cores.

History and Background

The company was founded in 1983 as Hyundai Electronic Industrial and became Hynix Semiconductor in 2001 before being acquired by SK Telecom's parent group in 2012 and renamed SK Hynix. Through the 2000s it built scale in commodity DRAM and NAND flash, later acquiring Intel's NAND business (rebranded Solidigm) in 2021.^[1]

Its pivotal AI move came in 2015, when it co-developed the first high-bandwidth memory with AMD for the Fiji GPU generation. While competitors treated HBM as a niche product, SK Hynix invested through the downturn and became the lead development partner for NVIDIA's accelerator roadmaps. It mass-produced HBM3E from 2024 and completed the world's first HBM4 development, readying mass production in September 2025 as demand shifted toward NVIDIA's Rubin platform.^[2]

Key Concepts

• Why HBM matters for AI. Modern accelerators are limited less by arithmetic than by memory bandwidth. HBM stacks several DRAM dies on a silicon interposer directly beside the processor, giving hundreds of terabytes per second of bandwidth at the cost of a larger, more expensive package. NVIDIA Blackwell and Rubin-class GPUs, Google's TPUs and other accelerators all depend on it.^[1]
• Market position. Counterpoint Research ranked SK Hynix No. 1 in HBM shipments with roughly 62% share in mid-2025, easing to about 50% by mid-2026 as Samsung and Micron qualified their own parts. Analysts expect it to hold the majority of the HBM4 segment on NVIDIA's Rubin platform.^[3]
• Capacity as strategy. HBM output is constrained by advanced packaging — stacking, through-silicon vias and interposers — rather than by cleanroom wafer starts alone. SK Hynix's early lead in packaging yield is why its capacity has been sold out years in advance, and why memory, not logic chips, has repeatedly been the binding constraint on AI buildouts.^[2]
• Broader portfolio. The company remains a major supplier of conventional DDR5 server and mobile DRAM, NAND flash and enterprise SSDs, giving it a cyclical base beneath the AI-driven HBM business.

Applications and Impact

Every leading AI training and inference cluster in the world runs on HBM-equipped accelerators, so SK Hynix's output directly sets the pace of global AI capacity. The HBM market grew from roughly $38 billion in 2025 to an estimated $58 billion in 2026, with projections approaching $100 billion by 2028 as content per accelerator rises.^[1] Tight HBM supply has been cited in accelerator lead times, price increases and the emergence of a broader "memory supercycle", and it has reshaped supplier power: memory makers that historically competed on price now negotiate multi-year take-or-pay contracts with hyperscalers and GPU vendors.^[3]

>See Also

• High Bandwidth Memory
• NVIDIA
• AI Accelerators
• TSMC
• Penang Semiconductor and AI Ecosystem
• AI Data Centres Malaysia