What AI Is Made Of

Evidence, assumptions, and a reproducible model

From a gigawatt to the things it takes.

This is an investigation of a hypothetical data centre with one gigawatt of commissioned IT load. Each run adds a layer: the plant, its equipment, material evidence, and finally an equipment material balance.

The research trail

  1. RUN 01

    Establish the physical evidence

    Gather sources and trace claims about power, cooling, equipment and material flows. Keep assumptions separate from sourced observations.

    Read the findings →
  2. RUN 01.5

    Specify the reference plant

    Define the load, design conditions, architecture, power and heat ledger, and equipment quantities. These remain the inherited physical reference for the site.

    Read the findings →
  3. RUN 02

    Reconcile scoped materials

    Connect equipment quantities to component evidence. Preserve open scopes and avoid treating unknown material content as zero.

    Read the findings →
  4. RUN 03

    Deepen the forensics

    Examine product documents, material fractions and engineering bounds. These scoped records remain available separately from the Run 04 estimates.

    Read the findings →
  5. RUN 04

    Estimate and reconcile

    Build an explicit equipment/material scenario from the retained findings and source records. Export low, base and high cases, document parameters and compare every available historical endpoint. Use the comparison records to understand differences between model versions.

    Read the findings →

How to read the results

Why 428 and 642 transformers?

The reference design has 428 IT step-down transformers plus 214 facility step-down transformers: 642 step-down units in total. The 24 upstream high-voltage transformers are a separate class. Transformers inside PDUs are also outside this step-down subtotal. The illustrated bricks represent classes, not one brick per physical unit.

Audit the evidence

These public copies preserve the selected research records and include source hashes. Historical reports describe their original run, while the active site uses Run 04. The full development repository remains private.

The former Run 01.5 protocol link now leads here, covering the full Run 01–04 process. For the original protocol, see the archived Run 01.5 protocol.