A carbon-flavored query: How to reduce AI’s ecological footprint

No matter how “disembodied” artificial intelligence (AI) may seem, it leaves behind a very tangible footprint  —  energy and water. The more we use AI, the more serious its impact becomes  —  not only due to training new models but also due to the constant operation of data centers. Many people don’t even realize that behind every simple request lies a complex system of servers, algorithms, and energy flows, resulting in a significant environmental footprint.

How AI’s environmental footprint is created

Most of the AI ​​load falls on physical infrastructure: data centers, power grids, cooling systems, and supporting equipment. Data centers are buildings housing servers, data storage, switches, and routers, where computing occurs around the clock. According to the International Energy Agency (IEA), approximately 60% of a modern data center’s energy is spent on the servers themselves, up to 30% on cooling systems, and the remaining 10–20% on the network, storage, and auxiliary systems.

In 2024, data centers consumed approximately 415 TWh of electricity, representing approximately 1.5% of global consumption. According to IEA projections, by 2030, data center electricity consumption will more than double to approximately 945 TWh  —  roughly the same as the current consumption of all of Japan.

Server cooling uses water and evaporative cooling systems, which require significant amounts of water. MIT News estimates that approximately 2 liters of water are required for every kilowatt-hour consumed by a data center.

Besides operational consumption, there are so-called “embodied emissions”  —  CO₂ emitted during the production of servers, chips, cooling systems, and the construction and transportation of infrastructure. These emissions often remain under the radar and are not included in standard estimates of data center energy consumption.

How much does a single request weigh?

If we move down from the infrastructure level to the user level, the environmental footprint of each individual request can also be measured, albeit in tiny fractions. According to Sam Altman, CEO of OpenAI, one average request to ChatGPT consumes about 0.34 Wh of energy and 0.000085 gallons of water  —  roughly 1/15th of a teaspoon. At first glance, this seems like a small amount. But when there are billions of such requests, their cumulative impact becomes quite significant.

According to the study How Hungry is AI?, for more complex models like GPT-4o, a single long request can consume up to 0.43 Wh, and with mass use, this translates into enormous amounts of energy and water. The researchers also note that generating images, video, and audio requires several times more resources than a simple text response. Thus, 100 million requests per day (a completely realistic volume for popular chatbots) is equivalent to the energy consumption of an entire metropolis for several hours or the use of hundreds of thousands of liters of water.

Can AI be made greener, and who is responsible?

For several years, major IT corporations have been fighting to make their data centers and models more environmentally friendly. While previously the focus was primarily on offsetting emissions, companies are now moving toward systemic measures  —  from optimizing model architecture to switching to renewable energy sources.

Google

Google is currently considered a leader in green AI. According to their 2025 report, over 12 months, the company was able to reduce energy consumption by 33 times and its carbon footprint by 44 times, while improving the quality of its responses. Google also aims to transition to completely carbon-free energy 24/7 and offset 120% of its water consumption. They openly publish their energy consumption calculation methodology to set a standard of transparency for the entire industry.

Microsoft

Microsoft also declares its intention to make its AI services as sustainable as possible. In its 2024 report, the company emphasizes that AI development is not a reason to postpone environmental commitments. The company’s portfolio already includes 135 renewable power purchase agreements (PPAs), supporting the transition to zero emissions in data center operations.

Furthermore, Microsoft is conducting research with Pacific Northwest National Laboratory: using AI, they have discovered new battery materials that can reduce dependence on lithium. In this way, artificial intelligence helps itself become cleaner.

Meta

According to its 2025 Sustainability Report, Meta (an organization designated as extremist in Russia by a court ruling and banned on March 21, 2022) plans to achieve net-zero emissions and positive water use status by 2030. However, the company is simultaneously investing hundreds of billions of dollars in new data centers in the US, which will house its most powerful AI servers. This poses a key challenge for the industry: how to reconcile technological growth with planetary conservation. Meta claims that these investments will enable the construction of more energy-efficient AI systems in the future, but the actual balance between growth and sustainability remains unclear.

What each user can do

Even with highly optimized data centers, the bulk of the workload is generated by users. As neural networks become a part of everyday life, completely abandoning AI is virtually impossible  —  it is far wiser to use it responsibly. Any user can implement specific practices, such as:

• Formulating queries more clearly and precisely: specifying the format (“one sentence,” “three-item list,” “brief”) to avoid receiving an overly long response;
• Archiving and clearing history: deleting old chats and files reduces the load on storage;
• Avoid unnecessary regeneration: each “try again” click is a new computation chain;
• Use text formats whenever possible. A short paragraph requires hundreds of times less energy than an image;
• Choose providers that publish sustainability reports and use clean energy;
• Take a break from AI: consciously limit entertaining queries, use AI only where it provides tangible benefits.

Familiar environmental practices, such as saving electricity, should also be extended to the digital environment. It’s important to understand that artificial intelligence relies on powerful computing centers, which require colossal amounts of energy. Consciously and precisely formulating AI tasks is a simple yet effective way to reduce its environmental footprint.

illuminem Voices is a democratic space presenting the thoughts and opinions of leading Sustainability & Energy writers, their opinions do not necessarily represent those of illuminem.

Track the real‑world impact behind the sustainability headlines. illuminem’s Data Hub™ offers transparent performance data and climate targets of companies driving the transition.