“This Tool Estimates the Energy Consumption of Your Chatbot Conversations”
Have you ever wondered how much electricity is consumed when you ask an AI model a question or say thank you? Hugging Face‘s very well-known engineer Julien Delavande has created a special and excellent tool to find the answer to this question.
AI models consume electricity every time they are used. These models run on GPUs and other special chips, which consume a lot of energy to work at a large scale. Although it is not easy to know the exact power consumption of a model, it is believed that in the coming years, the demand for electricity will increase significantly due to the increasing use of AI.
The need for more electricity for AI has forced some companies to adopt paths that are not good for the environment. In such a situation, tools like Delavande’s serve to draw attention to this issue and may force some people to think.
Delavande and his colleagues say that “saving even a small amount of energy can have a huge impact over millions of questions and answers — even the choice of model or the length of its response can have an impact on the environment.”
Delavande’s tool works with the “Chat UI,” an open-source interface for models like Meta’s Llama 3.3 70B and Google’s Gemma 3. The tool shows how much energy is being spent interacting with a model — in real time. This tool indicates how much energy is being consumed during a conversation, and it also shows how long common household items — like a microwave or LED — can run on that much energy.
For example, asking a Llama 3.3 70B to write a typical email uses about 0.1841 watt-hours of energy — the same as running a microwave for 0.12 seconds or using a toaster for 0.02 seconds.
While this tool only provides an estimate and no claims are made about its accuracy, it is a reminder that everything has a cost — AI chatbots are no different.
Delavande and his team believe, “Projects like AI energy score and research on AI energy consumption are a step towards bringing transparency to the open-source community. There may come a day when energy consumption information will be as common as nutrition labels on food packages!”
