Definition
The AI stack is the set of layers that make up an AI product: hardware at the bottom, then data, frameworks, models, APIs and applications on top. Most people and companies build only at the upper layers.
In simple words
It's like running a restaurant. Someone makes the stoves (chips), someone grows the food (data), someone writes recipes (frameworks and models), and you serve the meal (the app). You don't have to farm in order to open a restaurant.
Why it matters
It tells you where you can realistically work. Training a frontier model costs a fortune, while building an app on an API can start today at very little cost.
How it works
| Layer | What it is | Examples |
|---|---|---|
| Application | What users touch | ChatGPT, Copilot, your app |
| API | Access to a model over the web | OpenAI, Anthropic, Google APIs |
| Model | The trained network | GPT, Claude, Gemini, Llama |
| Framework | Tools to build and train | PyTorch, TensorFlow, scikit-learn |
| Data | What models learn from | Datasets, databases, documents |
| Compute | Chips and data centres | GPUs, TPUs, cloud |
Common mistakes
- Believing you must train your own model. Most products use an existing one.
- Building a product only as a thin layer on one vendor, with no plan if prices or terms change.
Check yourself
Summary
- Compute, data, frameworks, models, APIs, applications.
- You can build at the top without owning what's below.
Finished reading?