How much water and energy does it take for one query in ChatGPT?

How much resource does one ChatGPT request consume

An average ChatGPT request uses about 0.000085 gallons of water — that’s roughly 1.15 teaspoons. Processing one request requires approximately 0.34 watt-hours of electricity. To understand the scale of these figures, Altman compared them to the energy consumption of household appliances: 0.34 watt-hours is roughly equivalent to one second of operation of an electric stove or a couple of minutes of operation of an LED light bulb.

Despite these numbers, OpenAI has not yet presented detailed calculation methodologies or sources confirming these estimates. This raises certain questions and discussions in scientific and technical circles, as measuring energy consumption and water usage at the scale of data center operations is a quite complex task and depends on many factors.

The importance of the topic of AI energy consumption and water usage

The discussion about AI's energy and water consumption is becoming increasingly relevant against the backdrop of the growing role of AI in life, industry, and business. Modern large language models, such as GPT-4, require huge computational resources not only for training but also for daily operations — processing millions of user requests.

Energy and water resources — key factors for sustainable development

Data centers where AI models are hosted and operate need significant server cooling. Water is often used for cooling systems, especially in regions with hot climates. As a result, water consumption becomes a significant item of environmental costs.

According to researchers' forecasts, AI energy consumption may exceed the energy costs of cryptocurrency mining, such as Bitcoin, in the coming years. This raises concerns about the industry's impact on climate change and the depletion of natural resources.

Differences in infrastructure and their impact on energy costs

An important factor affecting overall resource costs is the location of data centers. In regions with a cold climate, server cooling requires less energy and water, whereas in hot and arid regions the situation is reversed — water becomes the key resource for cooling, and its consumption increases significantly.

Thus, the environmental footprint of AI is not uniform and depends on many factors, including data center architecture, cooling technologies used, the type of electricity, renewable or carbon-intensive, and so on.

Expert opinions

Last year, The Washington Post, together with researchers, conducted an assessment of energy costs for generating GPT-4 text. They concluded that generating a short text of 100 words requires "a little more than one bottle of water." These estimates sparked

Many users are wondering how much energy goes into generating an AI response. Altman explains: an average request requires about 0.34 watt-hours of energy — that’s roughly how much an oven consumes in a second or an energy-saving light bulb in a couple of minutes.

However, a more unexpected aspect is water consumption. Cooling servers in data centers running models like GPT-4 requires significant volumes of water, especially in regions with hot climates.

The AI industry is already facing growing criticism regarding its environmental footprint. This year, researchers predicted that AI energy consumption may exceed the costs of Bitcoin mining.

Additionally, a 2023 study by The Washington Post showed that generating short text, such as "thank you" or "ok," in GPT-4 can consume as much water as fits in a bottle.

These data raise concerns, especially in the context of the climate crisis and freshwater shortages in certain regions.

According to Altman, the cost of AI in the future will approach the cost of electricity. This promises to make AI more accessible, but simultaneously requires greater transparency in terms of sustainability and environmental impact.