Have you heard of "rebound effect"? Sure you can say, individually, one query is not that much... but then it becomes integrated in search engines, so suddenly when there was no queries at all, now there's 500 billions per day, and it gets included in your CICD at every commit, and soon enough in your OS, etc
"Run the numbers" means "run the numbers for using agentic coding for 2 hours per day on a frontier model" not "run the numbers for a single query". The former is the worst case scenario.
Google Search's "AI", which is what you're hinting at is such a good example. Let's say there's 10 billion Google searches per day. 10 billion completions on what is going to be a very tiny, ultra finetuned model with lots of caching (including outputs).
Check out how many queries an hour of agentic coding results in. And input/completion tokens. Estimate energy usage of Opus vs something like Gemma 4 E2B. Calculate how many developers using Opus for coding 1 hour a day would equate to those 10 billion search query originated LLM calls.
You could not have provided a better example to show that without running the numbers you'll end up with assumptions that oppose reality.