Your next AI prompt comes with an energy bill: The growing environmental cost of chatbot use

As much of the Northern Hemisphere experiences increasingly intense summer temperatures and policymakers continue to examine the environmental impact of emerging technologies, attention is turning to an unexpected contributor to energy consumption: artificial intelligence. A new analysis from cybersecurity company Surfshark highlights the cumulative energy demands of generative AI systems such as ChatGPT. While an individual query consumes relatively little power, the sheer scale of global usage means that billions of daily prompts collectively require substantial energy resources.

The findings raise an important question: can society enjoy the productivity benefits of AI while also managing its environmental footprint?

Most users think of AI interactions as intangible. Typing a question into ChatGPT feels no different from performing a web search or sending an email. Behind the scenes, however, every prompt requires computing resources housed in large-scale data centres.

According to Surfshark’s analysis, a typical ChatGPT query consumes approximately 2 watt-hours of energy. That equates to running a 40-watt mini cooling fan for around three minutes or charging a smartphone with a 5-watt charger for roughly 24 minutes.  In isolation, these numbers seem trivial. The challenge arises when those queries are multiplied by billions.

OpenAI has reported that ChatGPT handles around 2.5 billion queries each day. Surfshark estimates that, at this scale, the energy consumption associated with those queries could power approximately 200,000 air-conditioning units continuously for 24 hours. According to the researchers, that would be sufficient to cool entire cities such as Miami, Lyon or Canberra for a day.  The comparison provides a striking illustration of how seemingly insignificant individual actions can accumulate into substantial infrastructure demands.

Energy consumption is only part of the story. Because many electricity grids remain partly dependent on fossil fuels, AI use also creates associated carbon emissions. Surfshark estimates that a single ChatGPT prompt generates approximately 4.32 grams of carbon dioxide equivalent.

Again, the figure appears small on a per-query basis. However, multiplied across billions of interactions, the environmental impact becomes more significant. The researchers estimate that if every person in a large industrialised country submitted a single ChatGPT request on the same day, the resulting emissions could reach thousands of tonnes of carbon dioxide.

There is also the issue of water consumption. Data centres require extensive cooling systems to maintain stable operating temperatures, and AI-driven workloads can increase pressure on cooling infrastructure. Surfshark’s analysts note that the training and operation of large AI models require huge numbers of servers running continuously, with cooling systems consuming substantial volumes of water.

Why AI consumes so much energy

The environmental challenge associated with AI is not limited to answering everyday user prompts. Experts often point out that model training represents the largest portion of AI’s energy footprint. Creating large language models requires extensive computational resources operating for prolonged periods. Once deployed, the models must then serve millions of users simultaneously.  Complicating matters further, estimates of AI energy consumption vary considerably across studies.

Surfshark notes that estimates for a ChatGPT query range from around 0.3 watt-hours to nearly 3 watt-hours depending on methodology, hardware efficiency and model type. More advanced reasoning models may require significantly greater computational effort than simpler AI systems.

The uncertainty reflects a broader transparency problem. Technology companies rarely disclose detailed energy consumption data for their AI infrastructure, leaving researchers to estimate environmental impacts using available technical information.

The issue becomes even more significant when viewed against projected growth in AI adoption. Surfshark estimates that the number of AI users worldwide reached approximately one billion in the first half of 2026, representing a dramatic increase from the previous year.

Researchers at IEEE Spectrum have similarly highlighted the scale challenge facing the industry. Based on reported usage figures, billions of daily interactions already require vast amounts of electricity, and future AI agents operating autonomously may multiply those demands even further.  The result is growing pressure on technology companies to improve efficiency while maintaining model performance.

Despite the environmental concerns, experts caution against viewing AI as inherently unsustainable. Tomas Ivanaitis, Head of Data & AI at Surfshark, argues that responsible AI use is more about efficiency than avoidance. His recommendations focus on reducing unnecessary computational workloads while continuing to benefit from the technology.

Among the suggestions are formulating prompts carefully rather than repeatedly asking variations of the same question. In addition, using AI when it genuinely adds value rather than for tasks that can be completed more efficiently without assistance.

Another model involves deploying smaller, specialised AI models for routine business applications rather than using the largest available models for every task. This recommendation may prove particularly important for organisations. Just as businesses would not use a high-performance industrial machine for a simple household task, not every AI challenge requires the most powerful model available.