Sources checked
A request, facility, and global system have different boundaries
Different questions require different boundaries: one request, a facility, and a global data-centre system are not interchangeable units. The report provides context for wider demand, rather than a universal energy number for any prompt.
Scenarios are not precise predictions
An original comparison should state what is counted and which uncertainty matters. Avoid converting a scenario into a precise prediction or comparing workloads with different hardware and utilization assumptions.
What the big numbers usually bundle
A forecast for a country’s data centres bundles AI with ordinary cloud, and it bundles training with serving. It assumes a build-out that may not happen, and a utilization that may not hold.
The paper or the agency note usually says so in a paragraph the headline drops. That paragraph is the result.
The headline number is the illustration.
Per-query energy is even easier to misuse. It depends on the model, the hardware, the batch size, and whether the facility’s overhead is included.
Two honest estimates can differ by an order of magnitude because they drew the box differently. If you cite one, cite the box.
What a reader can do with the figure
Use it to ask a better question of a supplier: what is your power purchase, what is your cooling, what load did you assume. Do not use it to rank two chatbots by morality.
You do not have their facility data. A company sustainability report is a company document.
Attribute it, and keep the year.
What to record before you rely on it
When you quote a figure, write the source, the year, and whether it is a forecast or a meter reading. Include the sentence that states the assumption.
If you cannot find that sentence, do not quote the number.
Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.
When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.
Name a person who owns that failure.
Common questions
Does a larger context window use more energy?
Longer requests do more compute, all else equal. How much, at whose facility, is not answered by the window size on a pricing page.
Avoid a per-token energy claim you did not compute.
Are training and serving the same problem?
They draw power at different times and places. A training run is a project.
Serving is a continuous load. A forecast that adds them is a scenario, not a single appliance’s rating.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





