it's getting hot and AI data centres are using a lot of energy

1990s Have a question? Look it up in the encyclopedia

2000s Have a question? Search for it on the internet

2010s Have a question? Google it on your smartphone

2026: Have a question? Just ask ChatGPT or Claude

In the last five years alone, the way that we have searched for answers has fundamentally changed. And today, both the eye test and emerging data suggest that a growing majority of people are turning to platforms like ChatGPT and Claude whenever they have a question that needs answering - for literally any question.

There are certainly several existential concerns relating to our dependence on these platforms. Rather than actively searching for answers ourselves and engaging in critical thinking, we now have information delivered to us instantly - and, more often than not, we accept those generated answers as an absolute truth. But that's a topic for another day, and for another article.

This article is about data centres. So, for basic knowledge: when a question on these platforms is asked, it is known as a 'prompt' - essentially the instruction that guides the AI model in generating a response. And to generate the response, these systems rely on enormous amounts of computing power to process information and to analyse patterns. And this computing power needed is supplied by Data Centres - the physical infrastructure behind it all.

Person asks ChatGPT question
>
ChatGPT uses computing power from data centres to prepare reply
>
Data centres consume a lot of energy, water and space
> 
One ChatGPT response from text consumes 0.047WH of energy 

And like with any supply-and-demand curve, the rapid growth in AI usage is driving the need for greater supply. In response, new data centres are being built and deployed at extraordinary pace all around the world.

Source: international energy agency

Where are these data centres located?

As displayed in the figure below, the US currently accounts for more than 50% of the world's data centres. However, China is rapidly expanding its infrastructure, and it would not be surprising if, within the next few years, it reaches a level comparable to that of the US.

source visual capitalist

While the US and China may have larger land areas to accommodate the incoming numbers of data centres, smaller countries such as the United Kingdom face greater spatial constraints. Consequently, data centres can occupy significant amounts of land, affecting neighbouring communities.

This issue is explored further down in this article.

The technology itself is demanding more energy

Training (and then the usage) of frontier AI models requires enormous amounts of energy.

GPT-4 (the latest AI computing model) is estimated to have consumed between 50 and 70 GWh of electricity during training, which is around 40 times more than GPT-3 - the previous model. GPT-4's training alone is also estimated to have produced approximately 25,000 tonnes of CO₂e emissions and consumed around 600 million litres of water. And again, this is just for the training of the models - let alone the usage of the technology, which of course is far greater.

Once the model has been deployed and is being used publicly, it is important to note that the type of prompt will vary the energy requirements in generating the answer. In the figure below, you'll see the energy range from a short text to a high resolution image.

Types of responses and energy use

An AI generated video in high res can consume over 415 Wh per clip.

Water use follows a similar pattern to energy, increasing from around 29 mL for an image to approximately 4.1 litres for a complex video. So, when you are thinking about offloading a simple question to ChatGPT or Claude, just keep these numbers in mind.

And then, aggregating all the people around the world who are using AI platforms, the energy demands are now massive. In 2025, data centres consumed an estimated 448 TWh of electricity worldwide - enough to rank as the world's eleventh-largest electricity consumer if it were considered a country. This resulted in an estimated carbon footprint of 189 million tonnes of CO₂e and a water footprint of approximately 4.5 trillion litres.

To offset this, we would require 3.2 billion tree seedlings grown over 10 years.

Local costs, distant benefits

Communities (that also have not agreed to the construction of data centres near them) are increasingly exposed to data centres environmental impacts.

Data centres place significant demands on local water resources, as most high-density server facilities depend on water-based cooling systems. While a portion of the cooling water is returned after use, the volume and timing of water withdrawals can place considerable stress on surface water supplies, mainly in arid regions or areas already experiencing water scarcity.

protest at data centre in the US

Beyond the data centre operations, constructing data centres and manufacturing graphics processing units (GPUs), servers, batteries, and other hardware require large quantities of critical minerals such as lithium, cobalt, and rare earth elements. Extracting these materials is energy, water, and land-intensive and often takes place in regions with weaker environmental regulations, increasing the risk of environmental degradation.

And at the end of their lifespan, AI hardware contributes to the rapidly growing global e-waste stream. AI infrastructure is projected to generate up to 2.5 million metric tonnes of e-waste annually by 2030.

Without effective recycling and recovery systems, hazardous materials such as lead, cadmium, and mercury can leach into soil and water, posing long-term environmental and public health risks. The local communities are impacted, and the globe 'benefits'...

Ideas for future

Two aspects:

1/ coordinated action across entire AI ecosystem.

• transparency. Everything that is AI generated should be tagged as such so people can differentiate content online.

• standardising (and mandating) reporting for AI companies and data centres to report electricity consumption, emissions, water use, land use and hardware lifecycle impacts.

• accountability from developers regarding environmental impacts of new technology.

• better global governance. There's an AI arms race, but like with any existential threat, diplomacy and global coordination should be enacted.

2/ users

• to us, the people. If you can figure out the answer just using critical thinking, do so. Every time you request an answer from an AI platform we are consuming energy we just don't have extra.

There are a select subsection of use cases where I see why AI matters so much, and why it's so important to continue innovating and leverage it's brilliant functionalities. But, in 99% of the use cases with AI, it is merely because we want to work 'faster, more efficient, less cost intensive etc.' - and, I am just not too sure whether that is in the best long term interest of humanity.

What are your thoughts?