The AI supply chain.
Eight physical layers — from rare earths to data-center power — sit between hyperscaler capex and a finished Rubin GPU. Every layer has its own physics, its own bottleneck, and its own catalyst calendar. This is the picks-and-shovels theme.
The market still stares at NVIDIA. The actual physics chokepoints sit five layers further upstream — at hybrid bonders that don't have alternatives, at indium phosphide wafers grown by three companies in the world, at gas turbines with three-year lead times. Hyperscalers have guided $700B of 2026 capex; that number is abstract until you ask what it physically has to become. We map it.
★ Live signal: AI Semis readout Book RSI None, breadth None%, 0 names overbought and 0 oversold. Prose generation requires ANTHROPIC_API_KEY — see SKILL.md. Thesis intact until data says otherwise.
From silicon to data center.
Eight physical layers, one narrowing aperture. Each layer has its own physics, its own moat, its own catalyst calendar — and a different bottleneck dynamic.
Hyperscalers / Neoclouds
7 names mapped to this physical layer of the AI supply chain.
Compute / Silicon
7 names mapped to this physical layer of the AI supply chain.
Foundry / EDA / WFE
8 names mapped to this physical layer of the AI supply chain.
Memory · HBM, NAND, HBF
8 names mapped to this physical layer of the AI supply chain.
Bonding / Photonics / Substrate
24 names mapped to this physical layer of the AI supply chain.
Bottlenecks / Materials
15 names mapped to this physical layer of the AI supply chain.
Test
16 names mapped to this physical layer of the AI supply chain.
Power & Thermal
13 names mapped to this physical layer of the AI supply chain.
Build · Trim · Coiled.
Three rotating buckets from the latest scan + research overlay. Each card carries entry · stop · invalidation and a thesis grounded in the bottleneck physics.