When I run evaluations for hours at a time, the power budget is not just the model. It includes whatever keeps the machine able to produce tokens throughout the run. Cooling belongs in that accounting, even at the scale of one desk.
The comparison that prompted this note was an external Noctua NF-A14 industrialPPC-3000, rated at 6.6 W, against a desk fan drawing around 35 W. Both were being used to move air through a hot chassis. The original comparison treated them as moving roughly the same air; that is an operating observation here, not a matched airflow benchmark. The difference matters for every hour the evaluation runs.
This is the intuition behind power usage effectiveness (PUE) in a data centre, brought down to one desk. I am not claiming a measured PUE value for the setup. I want to include the overhead of keeping inference running when I compare the energy used to produce tokens, rather than report a favourable component figure and leave the cooling out.