How badly harms does AI do to the environment? Both Google and Mistral have published their own self-assessments on the environmental impact of AI queries, so there are several answers to that question.
In July, Mistral, publishing its own AI model, announced a self-assessment of the environmental impact of training in terms of the amount of carbon dioxide (CO2) produced, the amount of water consumed, and the amount of material consumed. Google took a slightly different approach by revealing the amount of electricity and water that Gemini Query consumes, as well as revealing the amount of CO2 it produces.
Of course, there is a warning. Each report was self-generated and not run by an external auditor. Also, training the model consumes much more resources than guesses. It also assigns a chatbot every time a daily task user queries. Still, the report provides context for how much AI taxes the environment, despite excluding the impact of AI training and speculation by Openai and other competitors.
On Thursday, Google said that the estimates of the resources consumed by the “median” Gemini Query consume 0.24WH of energy and 0.26ml (5 drops) of water, producing the equivalent of 0.03 grams of carbon dioxide. Mistral's report was slightly different. For “LE Chat” responses that generate pages of text (400 tokens), Mistral consumes 50 milliliters of water, generating 1.14 grams of carbon dioxide equivalent, and 0.2 milligrams of non-resurrection resources equivalent.
Google said the “comparative model” is usually a little more generous and only looks at the effects of active TPU consumption and GPU consumption. In this way, the median Gemini text prompt uses energy of 0.10Wh, consumes 0.12ml of water, and releases what is equivalent to 0.02 grams of carbon dioxide.
Google has not released an assessment of the impact of training on Gemini models. Mistral Did: In January 2025, the training of two large models produced the equivalent of 20.4 kilotons of carbon dioxide, consumed 281,000 cubic meters of water and 650 kilograms of resources. This is an Olympic-sized water intake pool of around 112 Olympic sizes. Using EPA estimates that the average car produces 4.6 meters of carbon dioxide per year, this also produces 4,435 units of CO2 annually.
Environmental impact assessments assume that energy is actually generated through means of generating carbon dioxide, such as coal. Like solar, “clean” energy reduces its value.
Similarly, the amount of water “consumed” is usually intended for the use of evaporative cooling. Here, heat is transferred from the chip or server (which may be cooled with water) to what is known as the evaporator cooler. Evaporative coolers transfer heat efficiently, just as your body cools after exercise. When you sweat, the moisture evaporates. This is an endothermic reaction that draws heat from the body. Evaporative coolers perform the same function, not only absorbing heat from the server farm, but also bringing that water back to the atmosphere.
Google said it is using a holistic approach to energy management, including more efficient models, models such as more efficient models, flashlights, custom-made TPUs, efficient data centers, and efficient idling of unused CPUs, and more. Clean energy generation such as planned reactors can also help reduce the number of impacts.
“Today, as AI becomes increasingly integrated into all layers of our economy, it is important for developers, policymakers, businesses, governments and citizens to better understand the environmental footprint of this transformative technology,” adds Mistral's own report. “At Mistral AI, we believe we share collective responsibility with each actor in the value chain to address and mitigate the environmental impact of innovation.”
How much water and electricity does ChatGpt consume?
Reports from Mistral and Google have not been replicated by other companies. Epochai estimates that ChatGPT's average GPT-4O queries consume approximately 0.3Wh of energy based on an estimate of the type of server used by OpenAI.
However, the amount of resources AI consumes is quite different, and even AI energy scores are at best rudimentary.
“In reality, when is the energy grid connected to the data center where it is sent to requests, including the type and size of the model, the type of output it generates, and the myriad variables beyond control, and when is it processed? MIT Technology Review A study was found. Estimates were found that 15 queries, 10 images and 3 5-second videos consume 2.9kWh of power per day.
Still, Mistral's research authors point to the path to a “scoring system” where buyers and users can use these studies to select the AI models with the lowest environmental impact. He also called on other AI model makers to track their leads.
Whether AI is “bad” for the environment is still debate, but reports from Google and Mistral provide the foundation for a more inferred argument.
