Not investment advice. Not a recommendation to buy/sell. Educational only. Do your own research. Full disclaimer
Node 02 / 06 · educational map
The stacking problem inside the package — FLOPs starve without bandwidth.
AI Bottleneck Map
02
High-bandwidth memory and advanced packaging decide how many complete accelerators can ship, not just how fast they theoretically calculate.
Tickers are educational examples of names often cited in public AI-infra discussions. Not investment advice. Not a recommendation to buy/sell. Educational only. Do your own research.
Modern accelerators are hungry for bytes, not just math. High-bandwidth memory (HBM) stacked on or beside the compute die is what keeps those chips fed. Without it, a “GPU shortage” is sometimes a memory-and-packaging shortage wearing a GPU costume.
HBM is a different industrial problem than commodity DRAM. It needs through-silicon vias, tight stacking yields, and a packaging line that can attach memory to an expensive logic die without scrapping both.
Public commentary from chipmakers and cloud builders keeps returning to memory bandwidth as a limiter even when headline FLOPs look abundant. The map treats that as a first-class node.
If HBM supply loosens, watch whether the next bind is optics (moving data between chips) or power (running the denser boxes you can finally fill).
Example public names often cited in AI infrastructure discussions — not a buy list, not a rating.
NASDAQ
MU
Micron Technology
Often cited among the handful of firms that can manufacture high-bandwidth memory at scale.
KRX
000660.KS
SK hynix
Often cited in public AI-infra discussions as a leading HBM supplier used in advanced accelerators.
KRX
005930.KS
Samsung Electronics
Often cited as another major memory manufacturer in the HBM conversation — not as a score or forecast.
Example public names often cited in AI infrastructure discussions — not a buy list, not a rating.
Free weekly map update
A short educational note when the map changes. No paid course. No WhatsApp blast. Email only.