#The five layers of the AI industry
We hear about “the AI sector” or “AI companies” every day. That’s about as precise as “data” or “digital”.
Nvidia, which makes GPUs, doesn’t sit at the same level as OpenAI, which builds foundation models, or as Zoom with its new automatic call summaries. They’re all part of the AI value chain, just at different stages.
To make sense of it, I rather like Andrew Ng’s breakdown of the sector into five layers:
Semiconductors: the components, GPUs (Graphics Processing Units) in particular, are the base everything else rests on. They do the number crunching AI models require. Examples: Nvidia, AMD, Intel.
Cloud infrastructure: supplies computing power and storage to companies building AI products, so they don’t have to run their own data centers. Examples: Amazon AWS, Google Cloud, Microsoft Azure.
Foundation models: the models themselves (GPT, LLaMA, Claude…) are what AI applications are built on. They cost a fortune to train, but their quality sets the quality of every product downstream. Examples: OpenAI, Anthropic, Mistral.
Orchestration (of “agents”): frameworks for building intelligent workflows that combine several models and APIs. Examples: LangChain, MemGPT (now Letta), CrewAI.
Applications: the layer we see every day, with tools that bring AI into consumer or B2B products, and new entrants launching by the thousands. Examples: Perplexity, Jasper, Canva.
A quick recap table:
| Layer | Definition | Competition | Investment |
|---|---|---|---|
| Semiconductors | Compute hardware for AI | Low | Astronomical ($10B+) |
| Cloud infrastructure | Services for AI compute and storage | High | Very high (~$1B–$10B) |
| Foundation models | Algorithms that underpin AI applications | High | Very high (~$100M–$1B+) |
| Orchestration | Tools to coordinate models and APIs | High | Moderate (~$1M–$10M) |
| Applications | End products and services built on AI | Variable | Low to moderate (~$100K–$10M) |