The memory that makes an accelerator usable — stacked DRAM bonded beside the GPU. Bandwidth, not compute, is the binding constraint, and supply is sold out across all three makers through 2026.
Tightness
-3 this quarter — Eased slightly as Samsung qualifies HBM4 into Nvidia and all three suppliers ramp — but 2026 supply is sold out across the board and orders are booked into 2028.
HBM is the stacked DRAM bonded next to the GPU die — bandwidth, not raw compute, is what limits large-model training and inference. Every accelerator needs multiple stacks, and supply is set by just three manufacturers, two of which control ~90% of output. So memory allocation gates how many usable AI chips actually exist, in lockstep with CoWoS packaging downstream.
Unlike CoWoS's single source, three suppliers are all ramping hard — Samsung and Micron qualifying HBM4 into Nvidia broadens supply, and TrendForce flags a possible 2026 demand correction from AI-chip delays and inventory. Structurally, though, memory stays sold out through 2026 and most expect real relief only around 2028.
Get the squeeze before the headlines.
A weekly read on the AI buildout's tightest constraints. Free, sourced.