AI energy consumption: turn the IEA report into a workload budget

The IEA’s April 2025 Energy and AI report examines the relationship between AI development and electricity systems. Its executive summary argues that affordable, reliable and sustainable electricity is a determinant of AI development. This is a dated analytical baseline, not a live meter of electricity use in October 2026. 1

The useful business question is narrower than whether AI consumes energy. A team needs to understand how its service demand translates into computation, where that computation runs and what practical changes reduce resource use without losing necessary quality. Global analysis provides context; application decisions still require a documented workload boundary and evidence from the chosen deployment.

Electricity consumption, power demand and associated emissions answer different questions. Energy measures consumption over time; power describes the rate or capacity required. Emissions depend additionally on the electricity system and the accounting boundary. Treating these quantities as interchangeable makes both procurement and sustainability claims difficult to verify.

Ask suppliers what their figures include: accelerator use, the whole server, cooling or a broader facility allocation. Identify whether a number is directly measured, allocated or modeled. AI data centers covers infrastructure delivery; the energy assessment should establish a consistent boundary before comparing two providers or deciding that a smaller reported figure means a better environmental outcome.

One prompt may produce a short answer; another may trigger retrieval, repeated model calls and tool execution. A universal energy-per-prompt figure hides those differences. Use a functional unit such as an accepted support resolution or a correctly reviewed document, then count the work needed to reach that result.

Record model choice, input and output sizes, retries, cache use and human correction. These variables help explain changes even when direct energy telemetry is unavailable. Enterprise agents addresses broader deployment value; here the question is whether resource demand falls per completed task rather than merely per individual call while the number of calls quietly rises.

Recommended experiments include using a smaller model for routine cases, limiting unnecessary context, caching stable responses and removing redundant agent loops. Compare each change against the same quality threshold. An apparently efficient system that creates more corrections or repeated requests may simply move the resource cost elsewhere.

Keep total demand visible as well as intensity. If cheaper execution encourages substantially more usage, aggregate consumption can rise even while each accepted task becomes more efficient. This is an evaluation consideration, not a measured outcome asserted for a particular vendor. Report the two measures separately and retain the baseline used for comparison.

The IEA stresses that electricity availability matters for AI’s development. Translate that into questions about the actual hosting region, service continuity and supplier evidence rather than assuming that a global energy trend determines every site’s situation. Contractual claims and physical delivery need separate scrutiny. 1

A pilot should finish with a workload baseline, an explicit measurement boundary and an owner for reviewing rising demand. Advisory can help define those tests without inventing unavailable telemetry. The practical conclusion is disciplined measurement: distinguish forecasts from observations, label estimates honestly and approve efficiency changes only when useful output remains acceptable.