# Project Plan — The Definitive Guide to AI Data Centers

## 1. Vision
A single, navigable, *vendor-neutral* reference that takes a reader from "I want to stand up AI compute" all the way through siting, powering, building, filling with GPUs, networking, commissioning, and operating it — with the depth a practitioner needs to actually make decisions, and the date-stamped numbers to defend them. It teaches **decisions and their consequences**, not just definitions.

## 2. Why this doesn't already exist (our differentiation)
The research sweep confirmed there is **no comprehensive, neutral, end-to-end AI-DC guide**. The closest things and their gaps:

| What exists | Gap we fill |
|---|---|
| **SemiAnalysis** "Datacenter Anatomy", "100k H100 Clusters", "Neocloud Playbook" — the best technical teardowns | Paywalled; fragmented across many posts; no single navigable site; light on siting/permitting, L1–L5 commissioning, and steady-state ops; US/hyperscaler-centric |
| **Vendor/standards reference architectures** (NVIDIA DGX SuperPOD, OCP, Uptime, ASHRAE) | Each siloed and vendor/standard-specific; nobody stitches them into a neutral lifecycle |
| **Books** (Data Center Handbook, BICSI 002, *The Datacenter as a Computer*) | Authoritative but pre-2024 in spirit; thin on the AI-specific shift (800VDC, NVL72, liquid, HBM, scale-up fabrics, goodput economics) |
| **SEO overviews** (TechPlus, IndexBox) | Shallow, no primary sourcing, no BoM-level depth |

**Our edges:** (1) end-to-end lifecycle in one place; (2) vendor-neutral side-by-side matrices; (3) decision-and-consequence pedagogy; (4) a **live, date-stamped numbers register** that beats frozen print and paywalled spreadsheets; (5) free and openly navigable.

## 3. Structure
17 Parts (0–16) + Appendices A–G — see `TABLE_OF_CONTENTS.md`. The spine is the lifecycle:

Foundations → **Strategy/Archetypes/Economics** → **Project Delivery/Procurement/Contracts/Risk** → **Siting/Power/Permitting** → **Electrical** → **Cooling** → **The Building (civil/structural/fire/construction)** → **Compute/Silicon** → **Networking/Optics** → **Storage** → **Software/Orchestration/Delivery** → **Security/Reliability** → **Commissioning** → **Day-2 Ops** → **Sustainability** → **Trends/Future** → Appendices.

Recurring cross-cutting threads tagged throughout: **POWER-BOUND**, **GOODPUT**, **DENSITY-RAMP** (40→600 kW racks), **DECISION-AND-CONSEQUENCE**.

## 4. Build phases
- **Phase 1 — Structure & research** *(current)*: master TOC + bibliography + raw per-domain research. **Gate: user signs off on the TOC.**
- **Phase 2 — Site shell**: `index.html` + assets implementing the nav rail, routing/anchors, search index, the reusable component kit (decision table, comparison matrix, callout, `<details>` disclosure, "as-of" badge), and the design system. Ship with the full TOC rendered + Part/chapter landing stubs so the skeleton is navigable before content lands.
- **Phase 3 — Content authoring** ✅ *complete*: all 166 chapters + 7 appendices authored (173 entries, ~30,800 lines, 330 decision tables), fanned out part-by-part via `tools/` + the saved authoring workflow, each `node --check`-validated and cross-reference-verified (0 dangling links). Every chapter researched from `SOURCES.md`/`research/`, with volatile numbers date-stamped + cited.
- **Phase 4 — Interactive layer**: the hero SVG journeys (grid-to-chip power path, chip-to-atmosphere heat path, "walk the data hall"), subsystem schematics, the appendix calculators (GPU TCO / $-per-GPU-hr / $-per-Mtoken, PUE/WUE, density-to-cooling-cliff, redundancy selector, site-scoring template), and the **numbers-provenance register**.
- **Phase 5 — Review, polish, publish**: completeness pass against the critique, cross-reference/de-dup audit, responsive/perf pass, optional online publish.

## 5. Site tech approach
- **Static, self-contained, no build step.** Must open from `file://`. Vanilla HTML/CSS/JS for core reading; lazy-load heavy calculators/diagrams. No server or framework required to read.
- **Information architecture** (from `research/web-design-notes.md`): persistent collapsible left nav (Part → chapter → section), breadcrumb, deep-linkable anchors, client-side full-text search with facets (by Part, thread-tag, "decision tables only").
- **Progressive disclosure**: chapter opens at altitude (summary + thesis + "key decisions" box); derivations and long decision lists in `<details>`.
- **Reusable artifacts**: interactive decision/comparison tables; consistent SVG schematics; horizontal interactive timelines (rack-density roadmap, project phase-gates); per-section "last reviewed" badges + changelog.

## 6. Authoring workflow (Phase 3)
Per chapter: (1) pull the chapter's section list from the TOC + relevant `SOURCES.md` entries + `research/domain-research.json` key numbers; (2) draft in the house voice with decision tables where the TOC marks a fork; (3) verify every number against a cited source and stamp it; (4) adversarial review for accuracy and for duplication against canonical-home chapters; (5) render into the site component kit. Keep a running changelog and a provenance-register entry for each volatile figure.

## 7. Open questions to resolve before/at the Phase 1 gate
- Audience priority: investor/decision-maker vs hands-on engineer (drives how deep the default vs `<details>` split goes).
- Geographic scope: US-first vs global-from-the-start (Appendix on regional deltas exists either way).
- Depth-first vs breadth-first authoring: write a few flagship chapters deeply, or stub every chapter shallow then deepen?