🧠 Inside the ML Layer · Overview

Inside the ML Layer at a Glance: Eight Layers from Agent to Kernel

The machine-learning software of an AI factory on one page: the eight layers between a user's request and the GPU, the open tools and in-house systems at each, the numbers that matter, where each layer gets stuck, and a guided trace of one request through the whole stack. Everything links to the eight-part series.

Explore the stack ↓ Start at Part 1 →
⚠️ About the numbers. Everything here comes from the eight articles (as of 27 October 2026), which cite their sources. Figures marked ⚠️ come from secondary sources; vendor results ("up to") are labelled; the stack chooser and quiz are editorial. Disclosure: Claude, made by Anthropic, helps write this site; Anthropic projects (MCP, the Claude Agent SDK) appear alongside their competitors.

1. The stack

Eight layers, from the agent a user talks to down to the kernels running on the GPU. Tap a layer.

2. Follow one request through the stack

A user asks a coding agent to fix a failing test. Press Play to watch the request go down the stack and the answer come back up. A simplified illustration.

Press Play.

3. Headline numbers

One figure per part. Tap to open the article.

4. Open tools or in-house?

What each lab is publicly known to use at each layer. Most frontier labs disclose little; "—" means not disclosed.

Tap a cell.

5. Who governs the open stack

Much of the open ML layer has moved under neutral foundations since 2022; the rest is run by a vendor.

6. Timeline, 2007 → 2026

Tap a milestone.

7. Where each layer gets stuck

8. Build a stack

Four questions, one suggested tool per layer. An editorial starting point combining the choosers in each part.

9. Quiz

Eight questions, one per part.

Answer all eight.

10. Reading paths

Sources

Each article lists its own sources and a "Reading the numbers" section. Related hubs: Inside the Data Layer (the data this software consumes), Who Runs the AI Factory (the roles), Producing Intelligence (one training run end to end) and the AI Compute Stack (the hardware underneath).