The AI Infrastructure Stack

The AI Infrastructure Stack

A layered map of the AI supply chain — from the apps people pay for, down through the chips, the fabs, the power plants, and the rocks pulled out of the ground. Each layer is a separate market with its own players, moats, and capital cycle.

The stack is read top-down: every layer above creates demand for everything below it. Money flows the opposite direction — an OpenAI revenue dollar fans out into hyperscaler capex, GPUs, HBM, fab capacity, lithography tools, gases, copper, and electricity.

The 18 Layers, in 7 Tiers

Tier I · Demand — Software
Where end-user dollars enter the stack. These are the products and APIs that prove AI is paid for.
LAYER 01
Applications

Consumer assistants, dev tools, vertical AI, embedded copilots inside incumbent SaaS.

LAYER 02
Foundation Models

The frontier labs and open-weight challengers that train the models everything else runs on.

LAYER 03
AI Developer Tooling & Middleware

Vector DBs, agent frameworks, evals, observability, fine-tuning, labeling — the picks and shovels for AI builders.

Tier II · Compute as a Service
Where GPU time is rented. The financial pivot point of the stack — hyperscaler capex lives here.
LAYER 04
AI Cloud & Inference

Hyperscalers, AI-native neoclouds, inference specialists, sovereign clouds, AI-aware data platforms.

Tier III · Physical Data Center
The buildings, the metal, the wires. Real estate plus the systems that keep racks alive.
LAYER 05
Data Center Real Estate

Colo REITs, hyperscale developers, miner-to-AI pivots, land & power asset holders.

LAYER 06
Facility Systems

Cooling (air / liquid / immersion), PDUs, UPS, switchgear, backup gensets, racks, DCIM.

LAYER 07
Networking

Switches, optical transceivers, NICs/DPUs, DSPs, retimers, fiber, co-packaged optics.

Tier IV · Compute Hardware
The actual silicon and the boxes it ships in. The most concentrated profit pool in the stack.
LAYER 08
Accelerators (GPUs & ASICs)

Merchant GPUs, hyperscaler custom ASICs, ASIC design partners, training/inference startups, photonic compute.

LAYER 09
CPUs, Memory & Storage

Server CPUs, HBM, DDR, NAND/SSDs, HDDs, CXL pooling, on-board power management, substrates.

LAYER 10
Server & Rack Integration

AI server OEMs, Taiwanese ODMs, rack-scale integrators, EMS for AI boards.

Tier V · Silicon
Designing and manufacturing the chips. Where the deepest moats and the highest capital intensity live.
LAYER 11
Chip Design Software & IP

EDA tools, IP cores, photomasks — the duopolies you can’t design a modern chip without.

LAYER 12
Fabrication (Foundries)

Leading-edge logic, specialty/analog, memory IDMs. The TSMC layer.

LAYER 13
Advanced Packaging

CoWoS, HBM stacking, OSAT, IC substrates, bonders — the real GPU bottleneck.

LAYER 14
Semiconductor Equipment

Lithography (ASML), deposition, etch, CMP, ion implant, metrology, test, probe.

LAYER 15
Semiconductor Materials

Wafers, photoresist, CMP slurries, specialty gases, filtration, sputtering targets.

Tier VI · Energy
The bottleneck of this AI cycle. Whoever can deliver electrons to a rack wins.
LAYER 16
Power Generation

IPPs, regulated utilities, renewables, gas turbines, nuclear (incl. SMRs), uranium, gas E&P, BESS.

LAYER 17
Transmission, Distribution & Grid Build-out

Transformers, switchgear, cable, EPC contractors, electrical components — the years-long grid backlog.

Tier VII · Raw Inputs
Rocks, gases, water. Cyclical commodity exposure for the patient.
LAYER 18
Raw Materials

Copper, aluminum, steel, rare earths, silica, helium, gold, gallium/germanium, water.

How to read this site

Every layer page has the same shape: a brief description of what the layer does, a list of sub-categories, and the public + private players in each. Analysis — moats, unit economics, capital cycle, and stock-specific notes — gets filled in over time as research is done. Pages marked Analysis coming soon are scaffolding waiting for content.

Tickers are color-coded: NVDA public US, TSM public ADR, Private private company / no direct ticker.