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

Image: Unsplash””link_wrapped_content””>

To what extent does AI harm the planet? Google and Mistral both published their own self assessments of the environmental impact of a AI query.

Mistral, a company that publishes AI models, published in July a self evaluation of the environmental impact on training and querying their model. This included the amount and type of water and material consumed. Google took a different approach and published the amount of water and power that a Gemini search consumes as well as the amount CO2 that it produces.

There are, of course, caveats. Each report was generated by the user, and not by an external auditor. The training of a model uses a lot more resources than the inferencing or the tasks that users assign to a chatbot every time they ask it. The reports still provide some context on how much AI taxes our environment, even though the reports exclude the effects of AI inferencing and training by OpenAI and others.

Google announced on Thursday that it will be launching a new AI platform. According to its estimate of the resources consumed by an “average” Gemini query, it consumes 0.24Wh and 0.26 milliliters (5 drops) of water and generates the carbon dioxide equivalent of 9.2 seconds of watching television. Mistral’s Report slightly different: For “Le Chat” responses generating a full page of text (400 tokens), Mistral uses 50 milliliters water, produces 1.14 grams carbon dioxide and consumes 0.2 milligrams non-renewable resource equivalent.

Google stated that “comparative model” are typically a bit more lenient and only consider the impact of active TPU consumption and GPU consumption. The median Gemini text prompt consumes 0.10Wh, 0.12ml water, and emits equivalent to 0.02 grams carbon dioxide.

Google has not released any assessments of the impact on its Gemini models. Mistral did it: In January 2025 the training of its Large 2 model produced equivalent to 20,4 kilotons carbon dioxide, used 281,000 cubic meters water, and consumed about 650 kilograms in resources. This is equivalent to 112 Olympic-sized pools of water. Using the EPA estimate that a car produces 4.6 metric tonnes of carbon dioxide per year, this works out to 4,435 cars producing CO2 each year.

Environmental impact assessments assume energy is produced by means that produce carbon dioxide. Solar energy, for example, reduces this value.

The amount of water “consumed”in most cases, assumes that heat is transferred to an evaporative coolant from the chip or the server (which may also be cooled with water). The evaporative cooling system transfers heat in a similar way to how your body cools down after a workout. As you sweat, moisture evaporates. This endothermic reaction pulls heat out of your body. The same function is performed by an evaporative cooling system, which removes heat from a server-farm while also evaporating the water.

Google said that it uses a holistic approach toward managing energy, such as more efficient models, optimized inferencing though models like Flash-Lite, custom-built TPUs, efficient data centers, and efficient idling of CPUs that aren’t being used. Clean energy generation — The impact numbers can also be reduced by a planned nuclear power plant .

Mistral’s report states: “Today, with AI becoming increasingly integrated into all layers of our economy, it’s crucial for developers, policymakers and enterprises to better understand the impact of this transformative technology.” “At Mistral AI we believe we share a responsibility with each actor in the value chain, to address and minimize the environmental impacts of innovations.

How much electricity and water does ChatGPT use?

Other companies have not duplicated the reports from Mistral or Google. EpochAI estimates that the average GPT-4o request on ChatGPT uses about 0.3Wh, based on its estimates of the types OpenAI servers.

The amount of resources AI uses can vary significantly, and even within the same AI system. AI energy scores at best are rudimentary.

In reality, the size and type of the model you’re using, the output that you’re generating and countless other variables outside of your control, such as which energy grid is linked to the data center where your request is sent and what time of day the request is processed, can make one query thousands times more energy-intensive. Study revealed. Its estimates of 15 questions a day plus ten images plus three 5-second video would consume 2.9kWh, it found.

Nevertheless, Mistral’s authors of the study note that their own estimates point towards a “scoring” system where buyers and users can use these studies to choose AI models with a minimal environmental impact. It also urged other AI model makers follow its lead.

The debate continues over whether AI is “bad” or not for the environment, but the reports by Google and Mistral are a good starting point for a more rational discussion.

Mark Hachman, Senior Editor, PCWorld.