Post by Monty Weber

The AI chip supply chain has 9 layers. Each one is a potential point of failure. A new map of the global AI chip supply chain — from raw earth to data centre — reveals just how concentrated the infrastructure behind artificial intelligence actually is. Layer 1 — Raw materials China controls roughly 60% of the rare earth metals used in chip production. The tin needed for EUV plasma generation and the specialty gases required for fabrication come from a small number of concentrated sources. Layer 2 — Silicon wafers Japan and Germany together supply around 80% of the advanced silicon wafers used in chip manufacturing. Three companies dominate: Shin-Etsu (Japan), SUMCO (Japan), and Siltronic (Germany). Layer 3 — EUV components The mirrors and precision optics inside EUV machines come from a single source: Carl Zeiss SMT in Germany. There is no substitute supplier. Layer 4 — The critical chokepoint ASML in the Netherlands is the only company on Earth that manufactures EUV lithography machines. Each machine costs $150M–$350M, contains over 100,000 components, and took 30 years and $9B in R&D to develop. ASML produces roughly 90 units per year. Layer 5 — Chip design Nvidia designs the chips. It doesn't manufacture a single one. Design software comes from Synopsys and Cadence, both US companies. Layer 6 — Fabrication TSMC in Taiwan manufactures approximately 90% of the world's most advanced chips. The $165B US expansion is underway — but Taiwan remains the primary production centre. Layer 7 — Advanced packaging CoWoS packaging — required for Nvidia's H100 and B200 chips — is itself a bottleneck. TSMC cannot build CoWoS capacity fast enough to meet demand. Layer 8 — The AI chip The Nvidia H100/B200 Blackwell runs on a 3nm process, contains 208 billion transistors, and costs $30,000–$40,000 per unit. 92% of all AI accelerators globally run on Nvidia architecture. Layer 9 — Data centres Microsoft Azure ($80B+ capex 2025), Google/AWS ($75B+ each), Meta and others are the end buyers — absorbing the output of this entire chain to run AI workloads. The final note on that map deserves repeating: every layer above must function without interruption for AI to run. Remove any single node and the chain breaks. That's not a metaphor. It's an engineering constraint — and increasingly a geopolitical one. #TheAIEspresso ______________ The AI Espresso · by Monty Weber Your weekly shot of AI news. Fast. Fun. Zero fluff. ☕ Already 8.700+ readers — join them.