Silicon
Compute begins here. So does the capex.
Chips for AI training and inference. The foundational layer where every model run starts. Designers (NVDA, AMD, AVGO) sit on top of foundries (TSM) and lithography (ASML) — capital-intensive, cyclically demand-driven, and the layer with the most concentrated economics in the entire stack.
Memory & Networking
Compute without bandwidth is silicon idling.
Memory bandwidth and rack-scale interconnect — the bottleneck after compute. A modern GPU sits idle without HBM memory bandwidth feeding it and high-speed networking connecting it into a cluster. Smaller market than silicon but increasingly strategic as model sizes outgrow single accelerators.
Cloud Infrastructure
Where chips become recurring revenue.
Hyperscalers that aggregate compute and rent it to everyone else. They buy chips by the tens of thousands, build datacenters in regions, and offer the picks-and-shovels for the rest of the market. The most profitable layer measured by gross margin, with deep moats from scale and customer lock-in.
Foundation Models & Tooling
Most of the IP. Few of the public pure-plays.
The model layer plus the picks-and-shovels around running them in production. Most pure-play foundation model companies (OpenAI, Anthropic) are private — public exposure comes via their cloud parents (MSFT, GOOG, AMZN). Tooling for observability, vector search, and orchestration is where the public market plays directly.
AI-native & Vertical Adopters
Where chip dollars finally become customer dollars.
Companies whose product is AI, or who derive durable advantage from deploying AI in their existing business. The most diverse layer — from pure AI-software (PLTR's AIP) to AI-as-feature companies (CRWD applying ML to security, NOW embedding AI in IT workflows). Where margins from chips, cloud, and models eventually meet the end-customer.
Power & Cooling
AI's electricity bill. The grid is the next chokepoint.
The layer most analysts skip but where the next bottleneck of 2026-2028 actually sits. Datacenter power demand is doubling, the grid is constrained, and cooling has moved from a side-business to mission-critical. Independent power producers (CEG), utilities (NEE), and electrical/cooling infrastructure (ETN, VRT) capture the spend that flows through every other layer.