The Open-Source AI Stack
The open-source AI stack has ten layers in production-pipeline order (what gets built on top of what). It also has five cross-cutting meta-layers that observe or constrain the whole stack rather than sitting between two other layers. Hover any layer for what is in it; click to dive in.
Core stack · surface
Foundation · substrate
How to read this map
The diagram above is a stack: the foundation (silicon) sits at the bottom, and each layer above it builds on what is below. Above silicon, the chips run somewhere (compute) on data (data) shaped by training tools (training) to produce weights. A runtime (runtime) serves those weights through an agent (agents) that talks to other agents and tools via wire protocols (protocols). The agent also reaches into a retrieval and memory store (retrieval and memory) at call time. Protocols sit at the top because that is the layer closest to the user.
The cross-cutting meta-layers run alongside the stack rather than stacking on it. They observe (evaluation), define (governance), trust-anchor (identity and trust), constrain (safety and guardrails), and shape the political framing (sovereignty and decentralization primitives).
Each layer page is a five-in-one view: a short editorial explainer, a living catalog of projects in it, the latest news at the layer, the grants flowing into it, and a curated reading list for going deeper.