In recent years, ChatGpt has exploded in popularity Nearly 200 million users Each day, we pump over 1 billion prompts into the app. These prompts may appear to complete requests from thin air.
But behind the scenes, artificial intelligence (AI) chatbots use enormous amounts of energy. In 2023, the data center used to train and process AI was responsible 4.4% of electricity usage In the US. All over the world, these centers are composed Approximately 1.5% Global energy consumption. These numbers are expected to rise at least sharply. Doubled by 2030 As demand for AI increases.
“Just three years ago, we didn't even have ChatGpt yet,” he said. Alex de Vrise GaoVrije Universiteit Amsterdam's sustainability researcher and founder of emerging technologies DigiconomistA platform specializing in publishing the unintended results of digital trends. “And now we're talking about technology that is responsible for almost half the electricity consumption of data centers around the world.”
But why are AI chatbots becoming more energy-intensive? The answer lies on the large scale of AI chatbots. In particular, AI has two parts that use the most energy. Training and reasoning states. Mosharaf Choudhrya computer scientist at the University of Michigan.
Related: Why does electricity give humming noise?
Training an AI chatbot gives a huge dataset to a large language model (LLMS) so that AI can learn, recognize and make predictions. In general, there is a “big belief” in AI training, De Vries-Gao said.
“So what happens when you're trying to train is that these days models are getting so big that they don't fit into a single GPU. [graphics processing unit];They don't fit on a single server,” Chowdhury told Live Science.
To give you a sense of scale, 2023 Research De Vries-Gao estimated that a single NVIDIA DGX A100 server would require up to 6.5 kilowatts of power. LLM training typically requires multiple servers, each with an average of 8 GPUs, running over weeks or months. Overall, this consumes a mountain of energy. It is estimated that Openai's GPT-4 training used 50 gigawatt-hour energy equivalent to powering San Francisco for three days.
Inference also consumes a lot of energy. This is where the AI chatbot draws conclusions from what it has learned and generates output from the request. Although it requires quite a few computational resources to run after LLM is trained, inference is energy intensive due to the large number of requests made to the AI chatbot.
As of July 2025, Openai situation ChATGPT users send over 2.5 billion prompts every day. This means using multiple servers to generate instantaneous responses for these requests. It doesn't even consider other widely used chatbots, including Google's Gemini. It will become the default option immediately When users access Google search.
“So, even in reasoning, you can't really save energy,” Choudhry said. “It's not really large data. So the model is already large, but there are a huge number of people using it.”
Researchers like Chowdhury and De Vries-Gao are currently working to understand how to better quantify these energy demands and reduce them. For example, Chowdhury holds an An ML Energy Leaderboard This tracks the inference energy consumption of open source models.
However, the specific energy demands of other generation AI platforms are largely unknown. Large companies like Google, Microsoft and Meta keep these numbers secret or provide statistics that give little insight into the actual environmental impact of these applications, De Vries-Gao said. This makes it difficult to determine how much energy AI actually uses, what energy needs are in the coming years, and whether the world can keep up.
However, those using these chatbots can promote better transparency. This not only helps users make more energy-aware choices through their own AI use, but also helps them drive more robust policies that keep businesses accountable.
“One of the very fundamental issues with digital applications is that the impact is never transparent,” says De Vries-Gao. “The ball is with policymakers to encourage disclosure so that users can start something.”
