Library
This guide connects the disciplines; it does not replace the specialist references. The committee standards own the classes and requirements you certify against, the OEM documentation governs the build once a platform is chosen, the institutions and analysts produce the measurements and forecasts the register cites. For every entry below: what it is authoritative for, and the moment to leave this site and open it. Access labels as checked 2026-09-05; paid does not mean avoidable.
The committee references own the environmental classes and the liquid-cooling interface definitions. Use them for the numbers you will write into a specification; use Part 5 for how those numbers interact with the rest of the machine.
Authoritative for: Air (A1–A4, H1) and facility-water (W17–W45+) environmental classes, liquid-cooling guidelines, the fourteen-book Datacom series as an updated online reference — about $33/year.
Reach for it when: Any inlet, supply-water or humidity limit you intend to put in a design basis or a contract.
What this guide adds: How a class choice moves free-cooling hours, heat reuse, transient ride-through and rack qualification at the same time.
Authoritative for: Expert-authored, peer-reviewed guidance across siting, integrated design, efficiency, grid interaction, resilience, commissioning, operations and retrofit — the closest free lifecycle reference to this guide.
Reach for it when: A reviewed statement of practice you can cite to an owner or authority; a second opinion on any Part 3–5, 13 or 15 recommendation here.
What this guide adds: Silicon, networking, storage, software and finance, and the decision-and-consequence framing across all of them.
Authoritative for: Cold-plate development and qualification methods, coolant requirements, ORv3 / Open Rack HPR / ORW rack and power-shelf specifications with drawings and contributor discussion.
Reach for it when: The actual interface specification a rack, manifold, quick-disconnect or power shelf must meet.
These documents are normative. This guide explains why a requirement matters to an AI workload and where the Tier label stops describing the fault behaviour you actually get; it does not replace the text you certify against.
Authoritative for: Tier topology and operational sustainability definitions, the annual global survey and outage analysis, and a 2026 AI-facility paper series (design review, vendor evaluation, construction, L4/L5 commissioning, operations).
Reach for it when: Any statement about Tier III/IV, concurrent maintainability or fault tolerance you intend to rely on; survey PUE and outage benchmarks.
Authoritative for: Normative site, architectural, electrical, mechanical, telecom and security requirements by rating level, with an audit and certification ecosystem.
Reach for it when: A contractual facility specification and its conformity path.
Authoritative for: Consensus best practice for data-center design and for operations and maintenance, with training and the DCDC credential behind it.
Reach for it when: Detailed design and operating practice in a form an AHJ or insurer recognises.
Authoritative for: Facility infrastructure classes, availability, physical security and energy-efficiency enablement requirements used across Europe and internationally; EN 50600-4 defines the KPIs (PUE, WUE, ERF, REF) EU reporting relies on.
Reach for it when: EU or international project requirements and the exact KPI definitions behind the EED reporting obligation.
Vendor material is authoritative about the vendor's own product and nothing else. The guide compares across vendors on fixed axes; when you have chosen one, its documentation governs the build.
Authoritative for: Engineered facility reference designs for AI halls (RD100, RD48 and successors) and a broad catalog of facility trade-off calculators: build vs colo, cooling architecture, economizer PUE by location, UPS economics, distribution efficiency.
Reach for it when: A bounded, engineered configuration to start from, and a calculator that matches the exact architectural fork you are weighing.
What this guide adds: Vendor-neutral comparison, the workload and economics context, and the register's dated provenance.
Authoritative for: Configured power, cooling and rack reference designs for named accelerator generations (GB300-class and beyond).
Reach for it when: The physical-infrastructure bill of materials for a named platform.
Authoritative for: Electrical distribution, grid-interaction and modular power engineering material for high-density AI loads.
Reach for it when: Electrical product choices and the engineering behind them; separate the engineering papers from the marketing.
Authoritative for: Rack, network, storage and management integration for a named generation; the executable basis for building and accepting an NVIDIA cluster.
Reach for it when: The supported topology, BOM and acceptance requirements once the platform is NVIDIA.
Authoritative for: Cluster design, networking with multiple fabric partners, BOMs, acceptance guidance and the facility interface for MI3xx platforms.
Reach for it when: The supported implementation once the platform is AMD; the counterpart to the SuperPOD guides.
Part 8 is a map. These are the territory: firsthand production accounts, a peer-reviewed textbook, and the consortium specifications.
Authoritative for: Structured, vendor-agnostic AI networking education — scale-up and scale-out fabrics, transports, congestion control, optics, operations — with worked material and review questions.
Reach for it when: Learning AI networking properly, or checking Part 8 against a peer-reviewed treatment.
What this guide adds: The links from fabric decisions to cooling, power, procurement, commissioning and cost.
Authoritative for: A firsthand account of two 24k-GPU clusters, one RoCE and one InfiniBand, both tuned to equivalent performance on GenAI workloads — the primary evidence against any universal transport ranking.
Reach for it when: Whether Ethernet can carry your training workload, and what the tuning actually involved.
Authoritative for: The Ultra Ethernet transport and its profiles as they are actually specified, not as summarised.
Reach for it when: Protocol-level detail behind the transport comparison in Chapter 8.4.
For demand forecasts, interconnection data and flexibility programs, go to the institutions that produce them. Part 3, 4 and 15 explain what they mean for a specific site.
Authoritative for: Global data-center electricity demand scenarios, supply implications and methodology; the reference forecast most policy discussion cites.
Reach for it when: Demand-growth numbers you will be asked to defend; scenario ranges rather than a single trajectory.
Authoritative for: Energy assessment tools, the US data-center energy report, and the annual Queued Up interconnection-queue dataset the guide's queue figures come from.
Reach for it when: The primary queue statistics and the methodology caveats that go with them.
Authoritative for: Applied research and utility demonstrations of data-center load flexibility, the utility-side view of curtailable interconnection.
Reach for it when: Turning a flexibility idea into a program a utility will recognise.
Authoritative for: A phased power co-development framework and a coalition directory for getting large loads energized faster.
Reach for it when: A structured process for the utility and generation conversations in Part 3.
The guide's calculators are screening tools. These are the measured or modelled sources you underwrite against.
Authoritative for: Original supply-chain (CoWoS, HBM), rack-teardown and fabric-economics research; InferenceX serving benchmarks by model, stack, precision and SLO; ClusterMAX GPU-cloud ratings. Institutional TCO models are licensed separately.
Reach for it when: Procurement or investment-grade economics, and measured serving envelopes instead of the guide's illustrative throughput constants.
What this guide adds: A common vocabulary across facilities and finance, and the register's separation of what is measured from what is modelled. About a tenth of the register's referenced claims cite SemiAnalysis — treat it as one evidentiary source, not a cross-check of itself.
Authoritative for: A transparent 1 GW cost model with stated inputs, and a documented dataset of named clusters, hardware and milestones with explicit scope definitions.
Reach for it when: A cost-model boundary you can audit, and cluster facts with record-level provenance.
Authoritative for: Governed, reproducible performance (and optional power) measurements with rules, metadata and code links.
Reach for it when: Comparing accelerators within a benchmark's bounds before believing any vendor ratio, including the ones quoted here.
Authoritative for: An analytical model that treats hardware refresh as an optimisation variable across build, refresh and inference operations.
Reach for it when: Refresh and lifecycle economics beyond the fixed-exit-multiple screen in the project-finance calculator.
Authoritative for: Regional vacancy, pricing, pipeline and lead-time conditions by market, published on a cadence.
Reach for it when: Market-specific availability and price before a siting decision; the guide's regional pages summarise, these are the source.
If you want to understand why the machine behaves the way it does rather than what to buy, start here.
Authoritative for: The original warehouse-scale-computer argument: co-design, cost models, failure handling and energy proportionality, from the people who ran it.
Reach for it when: Enduring architecture and economics reasoning; the component numbers are dated, the reasoning is not.
Authoritative for: Systems mathematics of training and inference: arithmetic intensity, parallelism, rooflines, KV-cache and batching, with derivations and exercises (TPU-centred, GPUs included).
Reach for it when: Any time the guide asserts a bandwidth-bound or compute-bound regime and you want the derivation.
Authoritative for: Hands-on implementation of the software stack this guide describes, with lectures, assignments and profiling exercises.
Reach for it when: Competence rather than orientation on Parts 9 and 10.
Authoritative for: Structured education, supervised assessment and credentials (ATD, DCDC, CDCDP and others) with professional networks behind them.
Reach for it when: Developing a team; this guide is free pre-reading, not a credential.
The numbers here move monthly. These are the original-reporting sources the register is checked against between publishes.
Authoritative for: Daily project, facility, power and market reporting; DCD Academy courses.
Reach for it when: What happened this week; the register lags it by design.
Authoritative for: Named-analyst commentary on facilities, investment and the operators building them.
Reach for it when: Perspective on the deals and developers behind the capacity numbers.
Authoritative for: The dated status of interconnection processes, alerts and rulemakings — the only correct source for 'where is Batch Zero this week'.
Reach for it when: Any process status you would put in a schedule; the guide records the sequence, the notice is the fact.
Missing something that belongs here? Tell us through the contact form. The bibliography behind individual figures lives on the register; Appendix E maps the learning and community landscape in more depth.