← The AI Systems Map
L4 · ToolsMemory / HBM⬦ ChokepointSouth Korea
HBM4
Next-gen memory gating 2026 accelerators
HBM4 widens the interface to 2048-bit and (for the first time) puts a logic base die often fabbed at a foundry under the DRAM stack, tightening the SK Hynix–TSMC–Nvidia coupling. HBM capacity is the single tightest input to AI accelerators; HBM4 qualification timing directly gates Rubin/MI400 volumes.
How it fits the stack
HBM4 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.
HBM4Depends on ↑Feeds ↓
HBM4 in the AI stack. HBM4 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) — 7 entities, 6 relationships
- HBM4 —depends on→ BE Semiconductor (Besi)
- HBM4 —depends on→ HBM lithography bonding (hybrid bonding)
- HBM4 —manufactures→ SK Hynix
- HBM4 —uses→ TSMC (Taiwan Semiconductor Manufacturing Company)
- Nvidia Data-Center GPU (Blackwell/Rubin) —supplies→ HBM4
- Nvidia Vera Rubin platform —uses→ HBM4
Depends on ↑ · 4
depends onBE Semiconductor (Besi)⬦Higher HBM stacks move toward hybrid bondingdepends onHBM lithography bonding (hybrid bonding)⬦HBM4 stacking needs hybrid bondingmanufacturesSK Hynix⬦SK Hynix leads HBM4 qualification for 2026 acceleratorsusesTSMC (Taiwan Semiconductor Manufacturing Company)⬦HBM4 logic base die is foundry-fabbed, tightening SK Hynix–TSMC–Nvidia coupling