The methodology is inherently flawed by assuming all infrastructure, training, etc is going to exist with or without individual queries, while trying to answer a different question of the impact of AI on the environment. It’s like arguing the environmental impact of solar electricity is 0 because the panels would exist either way.

Thus the results inherently fail to analyze the underlying question.

A more realistic estimate is to take their total spending assuming X% of their expenses are electricity directly or indirectly because the environmental impact isn’t adds up. Even that ignores the energy costs on 3rd party servers when they download their training data.