Technology
Large Language Model
From HBM memory to grid power — the hidden stack behind every chatbot.
⚠ HBM memory & CoWoS packaging
Key chokepoints
- GPU / AI accelerator siliconSemiconductors
Leading-edge nodes (3–5 nm) and HBM packaging are concentrated among a handful of fabs and OSAT partners.
Related tickers: NVDA, TSM
- High-bandwidth memory (HBM)Memory
Only SK hynix, Samsung, and Micron supply HBM3e at volume; lead times stretch during AI build-outs.
Related tickers: 000660.KS
- CoWoS advanced packagingPackaging
TSMC CoWoS capacity has been a binding constraint on AI accelerator shipments.