High-Bandwidth Memory (HBM)
The stacked-DRAM bottleneck that gates every AI accelerator
HBM is a 3D-stacked DRAM standard (JEDEC) that stacks 8-16 DRAM dies connected by through-silicon vias (TSVs) and sits beside the GPU/ASIC on a silicon interposer, delivering the terabytes-per-second of bandwidth that transformer training and inference demand. It is the single tightest chokepoint in the AI hardware stack alongside CoWoS packaging: HBM3E is the 2024-2025 volume product and HBM4 ramps in 2025-2026 with a wider 2048-bit interface and, for the first time, a logic base die often fabbed by a foundry. Capacity has been sold out through 2025 and largely booked into 2026, with SK Hynix, Samsung, and Micron the only three suppliers on Earth.
Suppliers worldwide
3 (SK Hynix, Samsung, Micron)
Current gen
HBM3E (8-hi/12-hi); HBM4 ramping 2025-2026
HBM4 interface
2048-bit (2x HBM3E), logic base die
Capacity status
Sold out through 2025, booked into 2026
How it fits the stack
High-Bandwidth Memory (HBM) with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.
High-Bandwidth Memory (HBM) in the AI stack. High-Bandwidth Memory (HBM) with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.
Graph data (text) — 14 entities, 15 relationships
- High-Bandwidth Memory (HBM) —manufactures→ CXMT (ChangXin Memory)
- High-Bandwidth Memory (HBM) —depends on→ DISCO Corporation
- High-Bandwidth Memory (HBM) —supplies→ DISCO Corporation
- High-Bandwidth Memory (HBM) —uses→ DRAM
- High-Bandwidth Memory (HBM) —depends on→ EUV Lithography (13.5nm)
- High-Bandwidth Memory (HBM) —depends on→ Liquid Cooling
- High-Bandwidth Memory (HBM) —manufactures→ Micron Technology
- AMD —used by→ High-Bandwidth Memory (HBM)
- AMD —uses→ High-Bandwidth Memory (HBM)
- ASE Technology (ASE Group / SPIL) —uses→ High-Bandwidth Memory (HBM)
- AWS Trainium / Inferentia —uses→ High-Bandwidth Memory (HBM)
- Broadcom —used by→ High-Bandwidth Memory (HBM)
- Etched —uses→ High-Bandwidth Memory (HBM)
- Google TPU —uses→ High-Bandwidth Memory (HBM)
- High-Bandwidth Memory (HBM) —invests in→ South Korea (K-Chips / HBM policy)
Depends on ↑ · 13
Feeds ↓ · 16