The Anatomy of an AI Chip
A Supply Chain Journey from Silicon Valley to Taiwan
Scroll to explore
Logic Design
Nvidia designs the H100 'Hopper' architecture in Santa Clara. This massive chip packs 80 billion transistors and pioneered the Transformer Engine for AI. It is designed specifically for TSMC's custom 4N process node.
Strategic Role
The Brain - Controls 90%+ of global AI training
EDA Software & IP
Synopsys and Cadence provide the Electronic Design Automation (EDA) tools essential for laying out 80 billion transistors. These US-controlled software tools are the first line of defense in semiconductor export controls.
Strategic Role
The Pencil - Design is impossible without this software
Instruction Architecture
While the GPU cores are proprietary, the H100's management subsystems rely on Arm CPU cores. Arm's instruction set architecture (ISA) remains the ubiquitous standard for low-power control logic within high-performance chips.
Strategic Role
The Blueprint - Foundation of chip control logic
HBM3 Memory
SK Hynix is the primary supplier of the H100's HBM3 memory, providing 80GB of capacity at 3.35TB/s bandwidth. SK Hynix holds ~62% of the HBM market as of 2025. Samsung is now qualified for HBM3E, and Micron is qualified for HBM3E while preparing HBM4 for next-generation platforms.
Strategic Role
The Fuel - Throughput bottleneck for LLMs
ABF Substrate
Ibiden manufactures the Ajinomoto Build-up Film (ABF) substrates. These complex multi-layered boards route thousands of signals from the nanometer-scale chip die to the millimeter-scale motherboard pins.
Strategic Role
The Foundation - Japan dominates this supply bottleneck
EUV Lithography Machine
ASML's Twinscan NXE:3800E systems are used to print the 4nm features. Each machine costs ~$180–200M and contains over 100,000 parts. ASML's next-generation High-NA EXE:5000, priced at ~$380M, is now shipping for future nodes. ASML remains the only company in the world capable of building these machines.
Strategic Role
The Monopoly - Single point of failure for global tech
Silicon Wafers
Shin-Etsu Chemical produces the ultra-pure 300mm silicon wafers. The silicon must be 99.999999999% pure ('11 nines') to avoid defects that could ruin the massive H100 die.
Strategic Role
The Canvas - Japan controls ~60% of wafer market
Wafer Fabrication
TSMC's Fab 18 fabricates the H100 using the '4N' process, a 4nm-class node derived from N5 and optimized for Nvidia. It takes ~3 months to process a wafer, with each yielding only roughly 65 perfect H100 dies due to their massive size.
Strategic Role
The Forge - The most advanced factory on Earth
CoWoS Packaging
TSMC's Chip-on-Wafer-on-Substrate (CoWoS) technology fuses the GPU die and 6 HBM3 memory stacks onto a silicon interposer. This packaging step was the primary global supply bottleneck through 2023–2024. TSMC has since tripled CoWoS capacity, easing the shortage, though advanced packaging remains in high demand.
Strategic Role
Critical Bottleneck - The true limit on AI scaling
Server Integration
Foxconn and Quanta assemble H100s into HGX baseboards and DGX servers. A single rack of these servers consumes as much power as a small neighborhood, requiring advanced liquid cooling integration.
Strategic Role
The Final Product - Shipped to AWS, Azure, & Meta
Journey Complete
The H100 is now ready to power AI workloads in data centers worldwide.
Back to Stories