🖥️ AI Compute Stack · Overview

The AI Compute Stack at a Glance: From Electrons to Tokens

Every AI answer travels through 18 layers, from power plants and transformers through chip fabs, memory and networks to the price on your invoice and the debt that paid for it all. This page puts the whole series on one map: click any layer for its key number, its bottleneck and who controls it, then jump into the full article.

Explore the stack ↓ Start the series at Part 1 →
⚠️ About these numbers. Every figure here is taken from one of the 18 articles in the series, written between June and October 2026, and inherits that article's sources and caveats (many are estimates, forecasts or vendor claims). The bottleneck ratings are our editorial judgement. Hyperscaler capex figures predate their late-October 2026 results. Click through to each article for sources.

🧱 The stack

Read from the bottom up: energy feeds facilities, facilities house silicon, silicon becomes systems, and systems produce the tokens whose economics fund the next round. The red bars show how tight each layer is today.

🔥 Bottleneck heatmap

Each layer rated 1–5 on three questions: how tight supply is today, how concentrated the suppliers are, and how long it takes to add capacity. Click a column heading to sort.

🪙 Follow one token

What has to happen, layer by layer, for a model to write a single word of an answer.

🔢 Key numbers

One headline figure per layer. Each tile links to the article it comes from.

🧭 Reading tracks

Every part stands alone. If you do not want to read all 18 in order, pick a track.

🔭 What to watch in 2027

📖 Glossary