How much power and water does AI use? Google, Mistral weigh in

Here is a 189-word summary of the news article: Google and Mistral have published self-assessments of the environmental impact of their AI models. Google found that a "median" Gemini query consumes 0.24Wh of energy, 0.26ml of water, and produces 0.03g of CO2 emissions - equivalent to 9 seconds of TV watching. Mistral's estimates for a 400-token "Le Chat" response are higher: 50ml of water, 1.14g of CO2, and 0.2mg of non-renewable resources. The reports have caveats, as they are self-generated and don't include the impacts of training the models, which can be much higher. Mistral's training of its Large 2 model in 2025 produced 20.4 kilotons of CO2, consumed 281,000 cubic meters of water, and used 650 kg of resources - equivalent to the annual emissions of 4,435 cars. The reports aim to provide a framework for understanding AI's environmental footprint and encourage other providers to follow suit. While the exact figures vary, the assessments suggest AI can have a significant environmental impact that needs to be accounted for and mitigated as the technology becomes more widespread.