Foundations & How to Use This Guide

An AI data center's compute, fabric, power, cooling and software form one co-designed machine, and its governing objective is TCO per unit of useful work. Optimize the pieces separately and you can pass every local test while the whole underperforms.

Each chapter names one engineering fork, the downstream cost of each branch, and the date-stamped numbers needed to choose. Read the chapter that matches the decision in front of you.

Vocabulary collisions over power figures, capacity units, metrics, nines and network tiers cost real money downstream. Fix one shared reading of each term before any contract or sizing decision depends on it.

Standards run on two clocks: resilience, thermal and security specs revise over years, while open-hardware, AI-governance and federal regimes re-version in months. Pin the version and price the drift.

Redundancy terms define which faults the building rides through, which maintenance work it survives, and which failures silicon and software absorb. Each choice has a price in capital, space and serviceability.