AI, energy, and environmental tradeoffs

Home Learn AI, energy, and environmental tradeoffs

In short

AI has a physical footprint because its computing equipment uses electricity, cooling, and materials. The impact of a use case depends on the system, how often it runs, where it operates, and what useful work it replaces or adds. A universal number for one prompt cannot answer all of those questions.

Why it matters

A person can generate many drafts because it feels cheap at the screen. At a larger scale, repeated processing affects infrastructure and budgets. At the same time, AI may help someone reduce waste or improve a process. Those potential benefits and costs need to be compared for a real job rather than assumed from the word AI.

How it works

Training builds or adapts a model; inference is the computation when someone uses it. Data centers supply servers, networking, and cooling for those tasks. Electricity use varies with the equipment, utilization, workload, and facility. Water use depends on cooling methods, electricity generation, location, and the boundaries of the measurement. Manufacturing and replacing hardware add another part of the footprint. The International Energy Agency’s 2025 Energy and AI report examines data-center demand and possible uses of AI in energy systems. Its companion chapter explains why different kinds of AI demand and infrastructure change the result.

For a particular task, define the outcome first. Is the system producing a useful guide, performing repeated low-value revisions, or helping avoid a material mistake? Compare a small, well-specified run with a workflow that repeats generation without a stopping rule. Look at the whole process, including storage, human review, and any physical result. A claimed efficiency improvement needs a measured baseline and a defined boundary; merely moving work to a remote service can hide the cost from the user.

Ask a provider or team for information that matches the intended workload: expected volume, time spent, infrastructure location when known, energy source, cooling approach, and how usage is measured. There may be no reliable figure for a particular prompt. Record uncertainty instead of inventing one. The IEA’s analysis is an aggregate estimate and scenario exercise, not a meter for an individual question. Adjusting the workflow to avoid needless loops is useful even before a perfect footprint estimate exists.

Example

Illustrative example: this scenario is invented.

Maya, 34, prepares descriptions for a seasonal shop display. She could ask an assistant to rewrite each description dozens of times. Instead, she supplies the actual product details, requests one draft per item, and reviews a short set of changes. She notes whether the process saves her time without increasing discarded printed signs. The meaningful comparison is the finished display and the work it took, not a guessed amount of energy per sentence.

What it doesn’t mean

The absence of a precise per-prompt figure does not mean the impact is zero. A global data-center forecast cannot tell you the footprint of one account. A possible environmental benefit in one use case does not erase infrastructure demand elsewhere. Describe the task, scale, and measurement boundary before claiming a net effect.

Try it

Choose an imaginary public brochure task. In ten minutes, compare a one-draft workflow with a loop that asks for many versions without criteria. Ask any assistant to list what information you would need to compare their resource use fairly. Note which facts are unknown and one practical stopping rule you could add now.

Keep going

Sources

Share this page