Chapter 12. Environmental Costs and Hidden Infrastructure

12.1 AI has a physical footprint

AI may appear to exist only on a screen, but it relies on data centres, servers, chips, networking, cooling, buildings, electricity, water, and supply chains. Training and serving models consume resources, and demand varies substantially by model size, hardware, workload, location, utilisation, and energy source. A single universal energy figure is rarely meaningful without assumptions and boundaries.

Environmental assessment should consider both direct operational impacts and the wider lifecycle: hardware manufacturing, minerals, transport, replacement cycles, water use, and electronic waste. Claims about efficiency should be supported by transparent measurements rather than marketing language.

12.2 Efficiency and proportionality

Choose a model sized for the task; a smaller or specialised system may be sufficient.
Reduce unnecessary repeated calls and avoid generating large outputs when concise ones work.
Cache reusable results where appropriate and measure actual workload performance.
Consider whether a non-AI process, a simple script, or an existing tool solves the problem with less resource use.
Ask suppliers for credible information about energy, water, hardware lifecycle, and reporting boundaries.
Include accessibility, reliability, and total cost of ownership in efficiency decisions—not only speed.

12.3 The rebound effect

Efficiency gains do not automatically reduce total environmental impact. If lower cost causes usage to grow rapidly, total resource demand may still rise. Organisations should measure aggregate use as well as per-task efficiency and avoid unnecessary deployment. The right question is not simply “Is this model efficient?” but “Does this use create enough public or organisational value to justify its full cost?”