The Rising Environmental Impact of Data Centres
Within the most advanced data centres around the globe, energy-intensive calculations take place around the clock. Chatbots and various generative AI tools are engaged in a wide array of tasks, ranging from the mundane—such as curating social media images—to the critical, including providing relationship advice, analysing medical images for cancer diagnosis, generating code for developers, and detecting financial fraud for banks.
As the popularity of AI tools has surged in recent years, so too have the associated environmental costs. Currently, data centres are responsible for an annual consumption of 414 terawatt-hours of electricity, which accounts for approximately 1.5% of global electricity use, as reported by the International Energy Agency. This figure has experienced a notable increase, growing by 12% annually over the past five years, and is projected to escalate by 17% by 2025. By 2030, electricity demand from data centres is expected to more than double. A significant portion of this rising demand is being met through fossil fuels, raising concerns among experts about the impact on local water resources used for cooling data centres, particularly in drought-stricken areas.
Individual AI Usage and Its Environmental Footprint
While the energy consumed by AI for generating text or images may seem minimal on an individual basis, experts emphasise that it is the responsibility of tech companies to mitigate the resource consumption of AI. This can be achieved through the development of more energy-efficient algorithms and hardware. However, there are straightforward steps individuals can take to ensure their use of AI has a minimal environmental impact. These include carefully considering the necessity of AI for a particular task and tailoring instructions to reduce computational demand.
“Individual decisions are not meaningless, and some have more impact than people realise,” asserts Ivana Drobnjak, a computer scientist at University College London.
The Energy Consumption of Chatbots
Estimating the energy expenditure for processing a single query via a chatbot is complex. For instance, Google estimates that its Gemini AI consumes approximately 0.24 watt-hours to respond to an average-length text query—equivalent to the electricity required to watch television for less than nine seconds. Additionally, it uses about 0.26 millilitres of water and emits approximately 0.03 grams of carbon dioxide (for comparison, driving a petrol car for one mile emits around 400 grams). While these figures are small on an individual level, they accumulate significantly for users and businesses heavily reliant on AI tools.
The substantial energy consumption of AI models is partly attributed to the processors powering them, such as graphics processing units (GPUs), which consume significantly more energy than central processing units (CPUs) that handle simpler tasks like web browsing and email. This issue is further compounded by the underlying models of many popular generative AI tools, including large language models (LLMs) that drive chatbots and AI assistants.
Understanding the Computational Demand of AI Models
These models are based on a specific design known as the “transformer” architecture, which enables LLMs to be trained on vast amounts of linguistic patterns found in texts. This training allows them to compute hundreds of billions or even trillions of parameters, which can then be utilised to generate new strings of text based on predicted word sequences.
A transformer-based LLM requires significant computational effort, as for each new word generated in response to a user’s query, it processes the inquiry and the preceding words through the model, performing billions of calculations each time. Tech companies assert that LLMs have become more energy-efficient over time; Google’s calculations indicate that the 0.24 watt-hours consumed by Gemini for a medium-length text request represents a 33-fold reduction in energy consumption compared to the model’s performance the previous year.
Strategies for Reducing AI Resource Consumption
Nevertheless, even modest amounts of energy can accumulate rapidly given the scale of AI usage. Based on 2025 estimates from OpenAI, Drobnjak projected in May that around 3.2 billion queries were being sent to the ChatGPT chatbot each day. Users are increasingly asking AI tools to process and generate vast quantities of text, images, and videos, often engaging in lengthy conversations with chatbots. Furthermore, a growing number of individuals are creating their own “AI agents” that, in turn, submit queries to AI chatbots.
So, what can users do to minimise the resources consumed while using AI? Experts offer several recommendations.
Reconsidering AI for Simple Tasks
As a preliminary measure to save energy, users should carefully consider whether AI is truly necessary for a specific task. “Asking ChatGPT ‘What should I wear today?’ or ‘What’s the weather forecast?’ is akin to taking a Concorde to the supermarket,” comments Günter Klambauer, an AI expert from Johannes Kepler University in Austria.
The same applies to web search engines that employ AI to automatically generate responses alongside traditional search results, such as Google’s AI summaries or Bing’s Copilot search function. “If you are simply looking for a specific article, turning off that feature could lead to significant energy savings,” notes Udit Gupta, an expert in electrical engineering and computer science at Cornell Tech in New York. Selecting “Web results only” in the browser or appending “-ai” to search queries may provide a solution.
Opting for Smaller Models
Individuals and businesses that frequently utilise AI tools for specific tasks, such as translation or summarisation, might consider switching to a smaller, task-specific LLM. These models are trained more specifically and require fewer calculations, resulting in lower energy consumption than their more generalised counterparts when performing the same tasks.
In a 2025 study published by UNESCO, Drobnjak evaluated the advantages of employing smaller models—such as “opus-mt-en-es” for English-to-Spanish translations and other models for summarisation and query response—over Meta’s Llama 3.1 model. Although these smaller models tend to be less intuitive than more popular AI models, they are freely available on the Hugging Face AI platform. The study revealed that the smaller models consumed between 15 and 50 times less energy while producing higher quality results for their intended tasks.
Minimising Conversational AI Usage
Drobnjak’s research demonstrated that abandoning larger models could lead to a reduction of up to 90% in overall energy consumption, making this the most effective energy-saving strategy. As Gupta states, “There is no need to use a billion-parameter model to edit an email.”
Given that LLMs perform numerous calculations for each consecutive word generated, it is helpful to choose models that produce less text overall. Mosharaf Chowdhury, an AI systems expert at the University of Michigan who has been measuring the energy consumption of publicly available LLMs, has found that models that are inherently “more conversational” tend to consume more energy. For instance, a version of the Qwen model developed by Alibaba Cloud consumes significantly more energy when it operates in its “reasoning problem-solving mode,” generating approximately ten times more words in response to a question compared to its “text conversation mode.” Consequently, some experts recommend using reasoning mode only for complex questions and relying on the standard chatbot mode for other cases.
Practical Tips for Energy Efficiency
Simply asking AI chatbots to “be concise” or setting a word limit can also contribute to energy savings. In the UNESCO study, Drobnjak and her colleagues found they could reduce the energy consumption of the Llama model by 50% by instructing it to halve its output. In contrast, keeping the initial request brief resulted in a less significant savings of only 5% for a request that was half the length of the original query.
“The size of the response is what most significantly determines energy expenditure,” asserts Drobnjak. She has collaborated with the city of San Francisco to develop energy-saving tips for AI users, which include being as specific as possible and adding instructions such as “a maximum of five points.”
Optimising Image and Video Generation
Similar recommendations apply to image and video generation, which can consume orders of magnitude more energy than text generation, as they involve iterating millions of pixels multiple times, processing the entire image afresh each time, explains Drobnjak. These tools are immensely popular, with nearly 40% of adolescents aged 13 to 17 surveyed in a recent Pew Research Center study reporting using AI to create or edit images or videos.
Drobnjak advises generating images or videos only when necessary and at the required resolution. “One option is to start with a low resolution and, if the algorithm is on the right track, gradually increase the resolution,” she says. She also highlights that editing existing images always requires fewer computational resources than generating new ones from scratch.
Moreover, when generating multiple images, it is more efficient to do so in a single session or in batches rather than making multiple requests separately.
Small Actions, Big Impact
These measures may appear to be a drop in the ocean, and to some extent, they are, according to experts. However, every little count. While we await tech companies, scientists, and policymakers to discover ways to lessen the global environmental impact of AI, “the person who knows how