About this guide
An end-to-end, vendor-neutral atlas for siting, building, financing, commissioning, and operating AI data centers — 166 chapters, free, and actively maintained.
Why I built it
I work on large AI data center projects. The first thing this field teaches you is that no one is an expert in all of it — the industry is too new. Everyone at the table arrived from somewhere else: utilities, hyperscale construction, HPC, mechanical and electrical engineering, real estate, project finance, silicon. All of it real, deep expertise — and each discipline brings its own vocabulary, assumptions, and dogma.
Which is how a room full of genuine subject-matter experts can still plan a bad AI data center. I ran into it constantly: misconceptions that survived until they became change orders, schedules built around the wrong critical path, every discipline optimizing locally while the project slipped around them. I have watched projects stall, slide, and fail at every stage for the same underlying reason — the team never shared one picture of the machine it was building.
The stakes make that intolerable. This build-out is deploying capital at a scale with almost no industrial precedent — McKinsey's April 2025 base case puts the data-center investment required by 2030 at $6.7 trillion, ~$5.2T of it AI-capable (a forecast of required spend, not committed capex). Capital like that does not forgive teams that talk past each other. Getting these projects built, delivered, and operated as integrated systems is the whole game.
I started this guide to educate myself — to be a better advisor to the projects I work on. Partway in, one thing became obvious: excellent sources exist for almost every piece — ASHRAE for thermal envelopes, Uptime for resilience, OCP for rack and power interfaces, the OEM reference architectures, IEA and LBNL for energy, the analysts for economics — but nobody had assembled them into one picture a project team could read end to end, in the order decisions actually get made. The seams between those excellent sources are where expensive mistakes happen. So the guide became public and free, and it points to the specialist references rather than pretending to replace them: the library lists them and says when to reach for each.
I hope you get as much out of it as I have. Help me keep it correct — every chapter ends with Suggest an edit — and if you want to stay ahead of the changes, the updates track the significant moves as they happen. This is frontier engineering; the numbers move monthly.
How it's made
I build the guide with the latest Claude and Codex models: AI agents do the research sweeps, drafting, cross-checking, and adversarial review; I direct the work and make the publication decisions. That editorial accountability does not constitute professional certification or project approval; review status is stated explicitly where it exists. The field spans grid interconnection, power electronics, cooling, silicon, networking, storage, software, security, and finance, and it moves monthly — AI assistance is what lets one accountable author keep 166 chapters at this depth. Where readers with domain expertise correct or sharpen the material, they are credited in the corrections log.
Why you can trust it
AI-assisted and human-owned works because the accountability is machinery, not promises:
- The numbers register — the guide's volatile figures are collected as 1,133 date-stamped register references and tracked over time as values are revised.
- The corrections log — reader-caught errors, public, dated and credited. Fixes from our own review passes are not listed there. When a figure changes, the change shows in that figure's revision history on the register; significant changes also go in the updates.
- Publish gates — content ships from a database through automatic checks before it goes live: a figure that disagrees with another statement of the same number blocks the release, stale figures are flagged for review, and every cross-reference must resolve. Nothing is hand-edited onto the live site.
- Adversarial review — before releases, separate-model adversarial reviewers (including a second lab's model) attack the site's numbers, math, and claims; verified findings become fixes and update entries. This is an editorial technique, not engineering assurance: no named discipline reviewer has signed off any chapter, and the guide says so wherever a reader might otherwise assume it.
The guide is an independent, AI-assisted editorial synthesis of cited public and licensed sources. Unless a named external review is shown, it has not been approved by an engineer of record, authority having jurisdiction, utility, OEM, regulator, standards body, or other professional authority. See the project-use boundary in the Terms.
Contact
Errors, sources I should read, or anything else: use the contact form or email editors@aidatacenterguide.com. For a specific passage, use Suggest an edit at the end of any chapter — accepted corrections are credited in the log.
— Jacob Fehn