The AI Web

This is the layer end-users actually touch — the assistants, agents, and AI-native apps that turn model capability into business value.

Layer 1

The foundation and frontier models that give every app its intelligence; the frontier leaders here are private (OpenAI, Anthropic).

Layer 2

The frameworks, MLOps, and data platforms that let models be trained, served, and monitored at scale.

Layer 3

The hyperscalers and neoclouds that rent out the physical compute everything above runs on; power availability is their binding constraint.

Demand engine

The GPUs, custom ASICs, and networking silicon that physically perform AI math — the most visible and valuable hardware layer.

Layer 5

HBM, DRAM and NAND — the binding constraint of the cycle, since GPUs starve without enough high-bandwidth memory.

Layer 6

The optics, retimers and cables that let thousands of chips act as one giant computer; increasingly the scaling bottleneck.

Layer 7

Advanced packaging (CoWoS) and substrates that physically assemble chips + HBM into one package — a hard chokepoint.

Layer 8

Where chips are actually manufactured; ~70% runs through TSMC, making Taiwan a single point of failure.

Layer 9

The machines that make chips — lithography, etch, deposition, test. ASML's EUV monopoly is the hardest chokepoint in the chain.

Layer 10

Wafers, photoresists, gases and CMP consumables — quiet oligopolies with deep qualification moats.

Layer 11

The bedrock — copper, rare earths, gallium, lithium. Often China-controlled, the ultimate upstream dependency.

Layer 12