The energy cost of one AI message

For a normal text exchange, the best recent public estimates suggest an energy cost of roughly a fraction of a watt-hour—not a fixed amount for every AI message. OpenAI chief executive Sam Altman has stated an average ChatGPT query uses about 0.34 watt-hours, while Epoch AI independently estimated about 0.3 Wh for a typical GPT-4o query.

Those figures are useful reference points, not direct measurements of your message. A two-line rewrite, a long document analysis and a tool-using research task may all look like one send action while requiring very different amounts of computing. Model, input length, output length, hidden reasoning, hardware and data-centre overhead all affect the total.

“One message” hides several units

A message has an input: your new prompt, system instructions and often previous conversation. It also has an output, generated one token at a time. Longer inputs need more processing; longer outputs keep the model working for more generation steps.

The visible message may trigger more than one model call. Web research can search, retrieve pages, synthesise them and check the result. An agent might write code, inspect an error and try again. Image creation uses a different computational process from text. A service average blends some mix of these jobs, while a worked estimate usually describes a particular one.

Before using a per-message figure, ask what kind of message it represents. Without that definition, the unit is convenient but weak.

Where the fraction-of-a-watt-hour estimate comes from

Altman’s 0.34 Wh number is a company statement without a published calculation. Epoch’s roughly 0.3 Wh figure is a modelled estimate with assumptions about GPT-4o, hardware, utilisation and a typical number of tokens. Agreement between them is reassuring, but it is not the same as independent access to OpenAI’s power meters.

Epoch also modelled how longer context changes the result. Its examples rise markedly for inputs of tens of thousands of tokens. That does not make those examples typical; it demonstrates why a universal constant fails.

Our deeper review of how much energy a ChatGPT query uses explains the assumptions and why ordinary text should not be used as a proxy for every AI feature.

Why the old 3 Wh estimate is a poor default

For years, articles repeated an estimate of about 3 Wh per ChatGPT query. It came from assumptions involving older hardware, a GPT-3.5-scale model and a very long output. Epoch’s comparison with earlier estimates argues that these inputs do not represent a typical modern chat.

The lesson is not that all earlier work was foolish. Estimates must choose a model, hardware generation and request shape, and those inputs age quickly. The error comes when a scenario becomes an eternal fact.

The reverse mistake is to treat the newer low average as proof that all AI use is negligible. Total demand also depends on the number of messages and the arrival of more intensive features.

A watt-hour measures energy. It does not tell you which power plant supplied it or the emissions associated with generation. Converting electricity into carbon requires a grid carbon-intensity assumption tied to a location and, ideally, a time. Company renewable-energy contracts add another accounting layer.

Water is different again. Some facilities use water directly for cooling, while electricity generation can have an indirect water footprint. Cooling design, climate and grid mix all matter. It is misleading to attach one universal water quantity to every AI message for the same reason a universal energy number fails.

AI carbon footprint explained separates operational electricity from embodied emissions in chips and buildings, which are normally outside a per-message inference estimate.

Smaller models can change the cost

For the same token counts and comparable serving conditions, a smaller model generally performs less computation than a large dense model. Quantisation, mixture-of-experts architectures and newer chips can reduce the work further. Efficient batching can share infrastructure across several requests.

But useful output matters. If a small model fails and the user repeats the prompt four times before switching to a larger one, the nominally efficient choice may not minimise the whole task. A narrow specialised model can be both smaller and better for its intended job; a difficult reasoning task may justify a larger system.

This is why small-model energy use should be considered alongside task fit, not used as a rule that the smallest option always wins.

What is normally left out

Most message estimates cover inference electricity and sometimes data-centre overhead. They usually leave out training, model development experiments, chip manufacture, data-centre construction, network equipment outside the facility and the user’s device. They may also omit idle capacity kept ready for peak demand.

There is no single mandatory boundary for every question. Operational energy is appropriate when comparing two ways to serve a response. Life-cycle impact is more appropriate when assessing a company or model over time. What matters is naming the boundary and not presenting a narrow estimate as a complete footprint.

At system scale, the IEA’s data-centre electricity outlook includes infrastructure demand and projects strong growth, with AI a major driver. It does not provide a universal message conversion.

How Treechat calculates its receipt

Treechat estimates rather than measures each answer. The calculation uses the selected model, input and output token counts, then adds a published data-centre overhead factor. It converts that energy into estimated emissions with a stated average grid assumption. The constants, rounding approach and known omissions are on our methodology page.

That receipt is best read comparatively: a shorter response from a smaller model should usually show less estimated operational energy than a long response from a larger one. It is not precise enough to certify the full life-cycle impact of either.

So the energy cost of one ordinary AI text message is plausibly around a fraction of a watt-hour under recent public estimates. For a specific message, the honest answer is a modelled range shaped by what the system actually did after you pressed send.