00 · IN THREE MINUTES

The answer in three steps

  1. 1Training and serving models require chip operations, memory movement, networking and cooling.
  2. 2Water use can occur on site in cooling and off site in electricity generation and manufacturing.
  3. 3A single universal footprint per prompt is misleading without model, hardware, location and utilization data.

01 · THE MODEL RUNS IN A MACHINE

The model runs in a machine

Matrix operations are performed by processors that draw electricity and move data through memory. Servers, storage, network equipment and power conversion add overhead beyond the accelerator itself.

02 · TRAINING AND INFERENCE DIFFER

Training and inference differ

Training may concentrate enormous computation into a limited run; inference repeats whenever users request outputs. A widely used model can accumulate more lifetime energy in serving than in its initial training.

FIG. 02How the system changes state
A conceptual mechanism map. Geometry, scale and timing are explanatory unless labelled otherwise.

03 · HEAT MUST LEAVE THE BUILDING

Heat must leave the building

Fans, chillers, evaporative systems or outside air transfer waste heat. depends on cooling design and climate; electricity generation may add an indirect water footprint even where the facility itself uses little.

04 · EFFICIENCY CAN LOWER OR RAISE TOTALS

Efficiency can lower or raise totals

Better chips and algorithms reduce energy per computation, but cheaper use can increase demand. The relevant measure combines efficiency, total workload, grid carbon intensity and the timing and location of operation.

05 · TRANSPARENT ACCOUNTING NEEDS BOUNDARIES

Transparent accounting needs boundaries

Reports should separate operational electricity, embodied hardware emissions and water withdrawal from water consumption. Estimates without system boundaries can differ while all appear numerically precise.

06 · SOURCES AND EVIDENCE

Sources and evidence

Claims are linked to foundational papers, standards or the primary study behind the update.

  1. 01
    Carbon Emissions and Large Neural Network Training

    Supports a defined mechanism, measurement or evidence boundary in this article.

    PRIMARY STUDY ↗
  2. 02
    Making AI Less Thirsty

    Supports a defined mechanism, measurement or evidence boundary in this article.

    PRIMARY STUDY ↗
  3. 03
    Energy and AI

    Supports a defined mechanism, measurement or evidence boundary in this article.

    INSTITUTIONAL REPORT ↗
CHANGE LOG29 Aug 2026 · First five-language edition; mechanisms, limits, diagrams and sources checked.