AlphaOS

What is HBM memory and which companies make it?

AlphaOS investment intelligence · Research and education only — not investment advice · Updated Sep 27, 2026

About artificial-intelligence

Direct answer

High Bandwidth Memory (HBM) is a specialized DRAM architecture that stacks multiple memory dies vertically using through-silicon vias (TSVs), connected directly to a processor via a silicon interposer, delivering dramatically higher bandwidth and lower power consumption than conventional GDDR memory. HBM is the critical memory technology enabling modern AI accelerators, with SK Hynix holding approximately 50% of HBM market share, Samsung holding roughly 40%, and Micron Technology capturing the remaining ~10%. SK Hynix is the dominant supplier to NVIDIA's H100 and H200 GPUs. The HBM market is projected to grow from approximately $4 billion in 2023 to over $20 billion by 2026, driven almost entirely by AI training and inference demand.

This week

Companies with new signals this week

METAgrowth ranks with an opportunity score of 90, driven primarily by growth signals.90MXLgrowth ranks with an opportunity score of 89, driven primarily by growth signals.89

2 of 24 shown

See the full board →

Key Takeaways

  • SK Hynix is the current HBM market leader with ~50% share, serving as primary supplier for NVIDIA's flagship H100/H200/B200 AI GPUs
  • HBM3e (the latest generation) delivers up to 1.2 TB/s memory bandwidth per stack, versus ~1 TB/s for HBM3 — bandwidth is the critical bottleneck for large language model training
  • Each NVIDIA H100 GPU contains 80GB of HBM3 across 5 stacks, making HBM a significant cost component of the ~$30,000–$40,000 per-unit GPU price
  • Samsung has struggled with HBM3e yield issues, temporarily ceding ground to SK Hynix; Samsung is working to qualify HBM3e with NVIDIA
  • Micron entered HBM with HBM3e in 2024 and secured qualification with NVIDIA, positioning itself to gain meaningful share from 2025 onward
  • HBM supply is structurally constrained because it consumes 3–5x the wafer capacity of standard DRAM per unit of output, limiting how quickly production can scale
  • AMD's MI300X accelerator uses HBM3 with 192GB capacity per chip, the highest in the industry as of 2024, sourced from SK Hynix and Samsung
  • The entire AI infrastructure buildout — NVIDIA, AMD, Google TPUs, custom ASICs — is dependent on HBM supply from only three manufacturers globally

Evidence & Analysis

  • HBM global market revenue estimated at ~$4B in 2023, forecast to exceed $20B by 2026 per industry analyst consensus (TrendForce, IDC)
  • NVIDIA H100 SXM5 contains 80GB HBM3 delivering 3.35 TB/s aggregate memory bandwidth, compared to 900 GB/s for GDDR6X in gaming GPUs
  • SK Hynix reported HBM accounted for over 40% of its DRAM revenue in Q3 2024, up from under 10% in 2022
  • HBM3e (8-high stack) operates at 1.2 TB/s per stack vs 819 GB/s for HBM3, a ~46% bandwidth improvement relevant to next-gen AI model scaling
  • A single HBM3e stack requires approximately 10–12 DRAM wafers to produce due to stacking, creating inherent supply constraints independent of demand growth
  • TSMC and other advanced packaging providers (CoWoS) are a secondary bottleneck — HBM must be integrated with GPU dies via silicon interposer, and CoWoS capacity was rationed through 2024

Key Companies

000660.KS

SK Hynix

HBM market leader with ~50% share; primary supplier of HBM3/HBM3e to NVIDIA for H100, H200, and Blackwell GPUs

005930.KS

Samsung Electronics

Second-largest HBM producer with ~40% share; facing yield challenges on HBM3e but working toward NVIDIA qualification

MU

Micron Technology

Emerging HBM competitor with ~10% share; qualified HBM3e with NVIDIA in 2024, targeting accelerated share gains through 2025-2026

NVDA

NVIDIA Corporation

Primary HBM demand driver; each Hopper and Blackwell GPU requires 5–8 HBM stacks, making NVIDIA the largest single buyer of HBM globally

AMD

Advanced Micro Devices

Second-largest AI GPU customer for HBM; MI300X uses 192GB HBM3, the highest capacity AI chip available as of 2024

Related stock research

Connected companies and research pages from the AlphaOS knowledge graph.

Priority stock research

High-intent stock intelligence pages — connected from AlphaOS research hubs.

Related stock research

Structured stock intelligence for companies connected to this research.

Related Questions

Generated by AlphaOS from the Knowledge Graph, earnings intelligence, and industry analysis. Content is for research and education only — not investment advice.