Inside the Data Layer at a Glance: From Web Crawl to Training Batch
The data supply chain of an AI factory on one page: the eight stages that turn the web, books, code, licensed archives, synthetic text and human judgement into training batches; the numbers, players and risks at each; a trace of one document through the whole chain; and a strategy builder. Everything links to the eight-part series.
Explore the supply chain ↓ Start at Part 1 →1. The supply chain
Eight stages from source to GPU. Tap a stage.
2. Follow one document
A web page explaining tides travels from the open web into a GPU. Press Play. A simplified illustration.
3. Headline numbers
One or two figures per part. Tap to open the article.
4. Three kinds of data, by training stage
An illustrative split of natural, synthetic and human-made data. Most labs do not publish theirs; NVIDIA reported 98%+ synthetic alignment data for Nemotron-4 340B.
5. Who discloses what
What each organisation has published about each stage, as read across the series. ● detailed · ◐ partial · — little or nothing. Tap a cell.
6. The data economy
Reported deal sizes and valuations. Hatched = press report, estimate or talks, not a confirmed figure.
7. Timeline, 1994 → 2026
8. What can go wrong at each stage
9. Build a data strategy
Four questions, one suggestion per stage. An editorial starting point combining the choosers in each part.
10. Quiz
Eight questions, one per part.
11. Reading paths
Related
The sibling hub Inside the ML Layer at a Glance covers the software that consumes this data. See also Producing Intelligence Part 2 (Llama 3's data, end to end) and Who Runs the AI Factory on the data platform and storage teams.