This is a great article that even talks about edge-hosted LLM models like what you might run on your cell phone, laptop, or desktop PC.
Llama3.1 8B (8 billion parameters) only pulled 114 joules to generate a response - for comparison that's 1/10th of a second of your microwave running.
Llama3.1 405B (405 billion parameters) pulled 6706 joules to generate a response - about 8 seconds of a microwave running.
Image generating models use about 1/3rd energy as much as text models, and video generating models use about twice as much energy as text models.
The devil in those details though, is that you often will make many queries to come to your conclusion. The article has a great breakdown of how much juice you'd use to brainstorm some marketing ideas, generate videos and images, and it comes out to around 3 and a half hours of running your microwave.
Every time google search results generate a summary of your search, it's a bite of energy. Scaled across billions of queries per hour. And it's something that everyone hates!
I do think there's value in locally hosted models doing work acceleration for you, but cloud providers spinning up huge datacenters to power features that nobody wants is pretty dogshit. Not to mention, those DC's are mostly powered by natgas and belching a shitton of carbon emissions for no meaningful value.