Follow one Blackwell GPU from beach to chatbot and you traverse five continents and a dozen near-monopolies. This is the chain the whole site maps; here it is in one pass.
Upstream: materials and machines
Ultra-pure silicon (Japan dominates wafers), specialty gases and photoresists (Japan again), and the tools: ASML's EUV monopoly, Applied/Lam/KLA/TEL for everything else. Chokepoint count so far: four countries, six companies.
Midstream: fab, memory, package
TSMC fabricates virtually every leading AI die (Taiwan risk in one word); SK hynix, Samsung and Micron supply HBM (Korea concentration); CoWoS packaging joins them — the 2023–24 bottleneck. Then ODMs (Foxconn, Quanta, Taiwan again) build the servers, with connectors, optics and power conversion from their own specialist chains.
Downstream: facilities and energy
Land, permits, grid positions; thermal and electrical chains with their own multi-year backlogs (transformers, turbines); then the operators — hyperscalers, neoclouds, colos — and the generation deals feeding them.
The intelligence layer
On top: software (CUDA's moat, open frameworks), data (web corpora, labeling pipelines, synthetic generation), models, and products — where the only step consumers ever see happens.
Reading the chain like an analyst
Three habits. Find the single points of failure: EUV, TSMC leading edge, HBM trio, CoWoS, transformer suppliers — shocks there propagate everywhere. Watch lead times, not prices: this chain rations by queue. Track vertical integration: hyperscalers designing chips, labs signing power deals, NVIDIA investing downstream — every layer is trying to swallow its neighbors, and each such move redraws this map. The Stack Explorer keeps the live version.
