Overview

NVIDIA is the representative product owner for this model — a fabless designer whose GPU-class accelerators are the finished output of the chain, and where the idiot index is largest (die value ≫ raw inputs).

What's in a finished accelerator

A high-end accelerator (H100/H200-class) is far more than the logic die:

• Logic die — a ~$18k 3nm wafer (TSMC) yields the compute die.
• HBM — 6–8 stacks at ~$300–500 each ⇒ ~$2,000–4,000 of memory (SK Hynix), often rivaling the die cost.
• CoWoS packaging + interposer + ABF substrate — ~$10k CoWoS wafer ASP plus substrate.

The market for these is the growth engine — see the Overview market chart (~25–31% CAGR for AI accelerators).

To verify next

Per-die cost split between wafer, HBM, packaging, and margin (proprietary).