AI electricity use: assess requests, workflows and infrastructure separately

How much electricity does an AI request use? The question sounds specific. Its answer alone will rarely be sufficient for a business decision. A short text, a lengthy image analysis and an automatically repeated process can all sit behind the same label. The number of requests also determines the total consumption of an application.

When hospitals or companies explain their use of AI, they should first define the level of analysis. Are they discussing an individual computation, an entire workflow or the infrastructure shared by many applications? These levels are connected. Each requires different information and supports different conclusions. Making them explicit makes a discussion about value and resources substantially easier to follow.

Hannah Ritchie's conversation with Scott Galloway on 20 August 2026 prompted this reflection. She discusses the importance of local electricity supply for AI development. I develop a question of my own for leaders: how can the energy discussion be organised so that it supports a defensible decision about a particular application? [1]

Start by defining the unit

In 2024, Luccioni and colleagues compared 88 models across ten tasks and 30 datasets. Their unbatched A100 experiments show task- and model-dependent consumption. They provide no universal electricity figure for today’s services. [2]

I derive a first requirement for any consumption figure: its description must accompany it. What task was completed? How extensive were the input and output? Was consumption measured or estimated? Which parts of processing does the number cover? An apparently precise figure without those details is harder to assess than a justified range with clear boundaries.

The same applies to comparisons with familiar everyday activities. They can make an order of magnitude understandable. For a decision, however, both sides of the comparison should answer the same question. Equating a short individual response with an extensive research process changes the comparison. The computation alone may also have a narrower scope than the complete service. Choosing the comparison unit therefore deserves as much attention as the number itself.

Count a complete process

Consider a fictional example. An administrative department in a hospital group introduces a search service for approved internal documents. Employees ask about purchasing procedures. The system retrieves relevant passages, generates an answer and retries after a technical interruption. Newly received documents are processed in the background. The visible questions alone provide an incomplete account of this process.

For planning, I would begin with a simple description of volumes. It includes expected use, repetitions, background processing and storage of the required data. The technical implementation determines which components actually arise. The aim is a sufficiently complete account against which the supplier can explain its figures. Missing values remain visible gaps until they can be clarified.

I would then place two views alongside each other: resources used per usable result and total resources used by the application over the chosen period. They can develop differently. A more efficient process may be used much more frequently. Resources per result then fall while total consumption increases. Whether the additional use creates corresponding value is a separate decision about the application's purpose.

In this example, “usable” means that employees can reliably resolve their legitimate question. A shorter answer that more often requires another search can lengthen the complete process. The quality requirement should therefore be established before comparing efficiency. Medical applications would additionally require their particular professional requirements to be clarified; the administrative example cannot be transferred to them without assessment.

Consider infrastructure separately

The International Energy Agency’s report of 16 April 2026 puts worldwide data-centre consumption in 2025 at 485 terawatt-hours. Its central projection reaches 950 terawatt-hours for 2030. Both figures cover all data centres. The report also identifies local supply and grid-connection constraints. [3]

For organisational leadership, this raises a different question from the one about an individual application: which dependencies arise from the service's location and electricity supply? A global total provides only a limited answer to this local question. Conversely, a constraint at one location cannot yield a general consumption figure for all AI use. Each piece of information should retain its particular reference point.

An organisation deciding on its own infrastructure therefore needs different information from one buying an existing service. For an internal operation, connection capacity, utilisation over time and available supply form part of the planning. For a purchased service, the first issue is which reliable information is available about the service used and how the supplier communicates changes. The decision remains tied to the actual scope of the undertaking.

Connect electricity and climate impact transparently

An amount of energy and a measure of climate impact have different units. Conversion requires an assumption about electricity generation. For an organisation's own comparison, I would also require the origin, reference period and boundaries of that assumption to be traceable.

I would be particularly careful with an assertion that an environmental burden has been avoided. If an application is intended to replace another activity, the activity that actually disappears must be described. When the service is additional, its value can still be justified; the comparison changes. A defensible account records these assumptions before assigning a number to the saving. Later observations can then be compared with the original expectation.

Uncertainty can also be presented usefully. A supplier might provide a measurement for part of the processing and an estimate for other components. Each should remain identifiable. A combined total can help planning when its uncertainty is retained. Additional decimal places cannot fill a gap in the explanation of what the calculation covers.

Make a decision that can be checked later

For an initial discussion, I would prepare one page containing four items: the application's purpose, the complete process, expected usage and supportable resource requirements. Add the open question whose resolution would most influence the decision. A supplier's figure may be missing. The team may still be uncertain how often an automatically triggered process is actually needed.

As a founder, I am interested in the connection between technical design and responsible use. A clear account of volumes can reveal where unnecessary work arises and where additional consumption serves a defensible purpose. It brings the energy discussion closer to the decisions an organisation can actually make and subsequently examine.