• A prompt on ChatGpt consumes more than a search on Google. But some science communicators urge us to analyse AI consumption data within a broader context.
• Looking at the issue from a different angle would help us understand that the real problem lies in where energy is produced and how transparent information on consumption actually is.

Artificial intelligence (AI) is becoming a fixture in our daily lives. It helps us write texts, solve problems, generate images and videos. Even make decisions. But every interaction – every single request made to ChatGpt or other artificial intelligence software – carries an invisible cost, one measured not in money but in kilowatt-hours. How much does AI really consume? And above all: is it reasonable to be alarmed by its environmental impact, as suggested by the headlines and investigations that have been circulating in recent months?

To answer that, we need to bring order to official data, independent analyses and corporate statements. Only then can we understand whether the new race towards artificial intelligence truly represents a threat to the climate – or whether, instead, it is fuelling a narrative that serves someone else's interests.

Water and electricity consumption: the AI threat

Where does artificial intelligence's energy consumption come from? At the root lie data centres: enormous digital infrastructures that process and store data to generate real-time responses. Every interaction with a system like ChatGpt requires considerable computing power, which translates into high electricity consumption and, often, a significant need for water to cool the servers.

A single request to ChatGpt consumes approximately 3 watt-hours of electricity according to the International Energy Agency (IEA), that is ten times more than a simple Google search. A figure that, taken in isolation, might seem irrelevant. But when multiplied by the millions of users who query these tools every day, the picture changes radically. The IEA itself warns that some of the largest data centres currently in the planning phase could end up consuming the equivalent of the energy used by 5 million homes. And this is only the beginning.

Several articles have put AI in the dock for its growing energy impact. For instance, according to a joint investigation by the journalistic consortium SourceMaterial and British outlet The Guardian, the leading technology companies – including Amazon, Microsoft and Google – are planning to build data centres in some of the world's most arid regions. Among these, Aragon in Spain stands out: in this region, Amazon's new facilities could consume more electricity than the entire Spanish autonomous community and require up to 750,000 cubic metres of water per year for cooling – an amount equal to the annual water needs of an Aragonese town of around 13,000 inhabitants.

Artificial intelligence: an exaggerated alarm?

There is no shortage, however, of voices urging us to scale back the alarm. Environmental science communicator Hannah Ritchie, for example, calls into question the interpretation of the IEA's data, according to which a request to ChatGpt consumes ten times more energy than a Google search. Ritchie does not dispute the number itself – around 3 watt-hours per prompt – but puts its impact into perspective: at an individual level, she argues, it is a negligible amount. To put it in context, Ritchie explains, those 3 Wh represent just 0.00007 per cent of a person's annual electricity consumption in the United Kingdom. Considering that the average daily consumption of a British citizen is around 12,000 Wh, a single question to ChatGPT accounts for a microscopic fraction of that. It is a matter of context: figures that, when read without a sense of proportion, can appear far more worrying than they really are.

Alex de Vries, one of the first researchers to attempt an estimate of AI's energy footprint, also urges caution. In his analysis The growing energy footprint of artificial intelligence, he calculated the theoretical consumption of servers produced by Nvidia – the sector's leading company – assuming that those sold in 2023 operate at full capacity. The result: consumption of between 5 and 10 terawatt-hours per year, a fraction of the approximately 460 TWh consumed collectively by all the world's data centres.

What is also cause for concern, however, is the speed at which artificial intelligence is spreading. But here too, those who take a more rational approach urge us to look at the numbers as a whole. According to the IEA – the very same source often cited in the most alarmist headlines – by 2030 data centres will account for 3 per cent of the increase in global electricity demand. Sectors such as heavy industry, electric vehicles or air conditioning will have a far more significant impact. Yet, as Ritchie again points out, this part of the discussion tends to disappear in the noise of public debate.

On artificial intelligence, then, the point lies elsewhere

The analyses by Hannah Ritchie and the sources she cites highlight two central issues. The first concerns the scale of the problem: distinguishing between global impact and local impact is fundamental. In some countries, data centres weigh far more heavily on electricity demand. In Ireland, for example, they already account for 17 per cent of national electricity consumption, while in the United States and several European countries the share hovers around 3–4 per cent. It is in these contexts that worrying phenomena can be observed: from the reactivation of coal-fired power plants to the planning of new nuclear facilities, with major tech companies continuing to expand their infrastructures at a rapid pace. The IEA also highlights how the spatial concentration of energy demand in data centres is far higher than in other types of buildings. It is precisely this density – more than the aggregate figure at a global level – that generates significant pressure on local electricity grids.

The second issue raised concerns less the quantity and more the quality of the energy used. In other words: what matters is not only how much electricity artificial intelligence consumes, but where it comes from. Excessive alarmism, Hannah Ritchie points out, risks handing big tech a convenient alibi, allowing these companies to justify their failure to meet climate targets by appealing to the unstoppable growth of AI. In this narrative, the recourse to fossil fuels – and to nuclear energy – ends up appearing inevitable. One example is the agreement signed between Microsoft and a facility in the area of Three Mile Island, grimly notorious for the most serious nuclear accident ever to occur in the United States.

BloombergNef, too, in its New energy outlook 2025 report, confirms this trend: the increase in energy demand linked to AI could slow the decline of global emissions, as a significant share of the electricity required will continue to come from fossil sources at least until 2035.

Fewer sensationalist headlines and more transparency

A further critical issue concerns the speed at which artificial intelligence is growing. Data centres can be built in a matter of months, while energy production facilities – particularly those using renewable sources – require years to be designed, approved and built. The concrete risk is that this discrepancy will create bottlenecks in electricity availability, pushing towards greater use of fossil fuels to bridge the gap.

To make the picture even more complex, a lack of transparency plays a significant role. Major technology companies do not systematically disclose comprehensive data on the specific energy consumption associated with artificial intelligence systems, nor on the actual CO2 footprint linked to their services. Furthermore, the location of data centres is often a trade secret. The aforementioned Guardian investigation with Source Material identified, for example, 632 centres either active or under development managed by Amazon, Microsoft and Google: this map shows how these companies plan to expand the number of data centres they operate worldwide by 78 per cent compared to current levels. An expansion that risks placing strain on existing infrastructure: this represents an information gap that makes it difficult to accurately assess the environmental impact of AI and limits the ability to develop infrastructure policies adequate to meet demand.

The environmental impact of artificial intelligence cannot be reduced to the question "But how much does ChatGpt pollute?", and must instead be approached with greater rigour. The numbers, read in the right context, do not justify alarmism, but neither do they warrant blind optimism. At the same time, it is important to recognise that AI systems are not only part of the problem, but can also contribute to the solution: advanced models can optimise energy consumption, improve the management of electricity grids and make industrial processes more efficient. In this sense, the responsible use of artificial intelligence can become an accelerator of the ecological transition, provided it is guided by environmental and social sustainability criteria.

To achieve this, however, accurate analysis, data transparency and coherent political choices are needed. Only then will it be possible to seriously evaluate the role of artificial intelligence in the ecological transition. Without offering big tech companies pretexts to slow down their climate commitments, and without turning AI into yet another scapegoat that diverts attention from the real issue: decarbonisation.