If you run it on your own hardware then you know exactly how much energy it's using. Some of the models can even run on the average computer, but the quality is not great.

The problem we have right now is where everyone is trying to use generative AI for everything, all the time, even for basic takes such as googling for a fact, simple math that could be done on a calculator, etc. They're also often using the latest and greatest AI models, which are powerful and spend a lot of processing power each time. In order to run the servers to respond to all that, companies use a large amount of power, and then use water to cool the servers. That's the water usage from what I understand.

So if you use a simpler / more efficient model, and only use it for tasks where it's better than conventional tools, you'll be doing much better power usage wise.

The real power consumption cost is from the model training. The marginal cost of running a query is small. Unless you use LLMs heavily, you probably use more power in many other ways, like if you forget to put your computer to sleep.

More training means better models, generally. If you want to minimize that, look for smaller models on HuggingFace that are more tightly focused on whatever task you have at hand.

Or, you could just do most of your queries when your local electricity mix has more renewables.