Sizing AI’s environmental footprint

To some environmental commentators, AI is the real-world equivalent of Erysichthon, the mythical Greek king who was cursed with a destructive and insatiable hunger.

In their view, the technology’s resource needs are so vast that its proliferation will cause a surge in electricity demand, a draining of the planet’s fresh water supplies and a spike in global carbon emissions. Left unchecked, AI will only accelerate climate change.

At first glance, the argument looks convincing enough. The amount of power needed to keep AI models running is not trivial.

A report funded by the US Department of Energy forecasts that AI’s electricity consumption in the US could rise from tens of terawatt hours per year currently to between 165 and 326 terawatt hours by 2028, enough to power one in five US households.

Dig deeper, however, and the data reveal a more nuanced picture. Not only is AI’s environmental impact relatively small and likely to remain manageable – particularly when compared to that of other industries – but the technology also has the potential to mitigate the causes of global warming. 

Down on data centres

Data centres – the workhorses of the digital economy and AI – are the primary focus of concern, not least because of the amount of power they consume.

According to the International Energy Agency (IEA), data centres’ electricity use could increase by as much as 15% per year by 2030.

Their water use is another worry. To prevent semiconductors from overheating, data centres continuously pump water through what are often vast cooling systems. The very largest ones can use more than 5 million gallons (20 million litres) of water a day, equivalent to the daily consumption of a town with 50,000 inhabitants.

To these concerns add carbon emissions. Research from the London School of Economics finds that data centres currently produce some 400 milliontonnes of carbon per year, a figure that could more than double to one billion by the end of the decade.

Yet AI infrastructure isn’t as environmentally damaging as the headline figures suggest. The public debate is often lacking crucial context.

For one thing, its energy consumption is modest compared with that of most other industries. The exception is the US, which is unsurprising given it is home to three-quarters of the world’s computing capacity; there, more than half of the projected rise in the country’s electricity consumption through to 2030 will be driven by AI, according to the IEA.

Globally, however, the picture is very different. Worldwide, data centres account for only about 1% of all electricity use, with AI representing just a quarter of that figure. And even if data centres’ power use grows at the present pace, they will contribute only about 10% to the overall rise in electricity demand this decade, the IEA says.

Tellingly, that figure is lower than the contribution expected from electric transport, air conditioning and many other industries through to 2030 (see figure 1).

Their water consumption also looks manageable. Data centres represent far less than 1% of the world’s total water use, and AI-related workloads just a quarter of that.

Although these aggregate figures mask the supply strains that emerge whenever computing infrastructure is located in areas facing water shortages, even in those instances, problems can be avoided by smarter planning and stricter development rules.

Data centres’ carbon footprint should not be problematic, either.

They are responsible for about 0.5% of global greenhouse gas emissions, with AI again accounting for only a fraction of that amount.

Taken together, this suggests that AI’s contribution to greenhouse gas emissions and, by extension, climate change, looks containable and not the main reason why governments and regulators should closely monitor its expansion.

Relative vs absolute

Nothing makes that point more clearly than a comparison of AI’s carbon footprint to that of many of the daily activities humans take for granted. When we travel, eat and use devices to keep ourselves warm in the winter or cool in the summer, the emissions are far greater than what an AI search would produce.

Generating the same volume of carbon as the production and preparation of a sirloin steak, for example, would require undertaking 617,000 AI queries on the Google AI search engine Gemini. A business class return flight from a European city to the US west coast, meanwhile, emits as much carbon (3.8 tonnes) as 6,000 AI queries per day for 60 years.

No less important in any assessment of AI’s carbon footprint is data centres’ use of renewable power.

Data centres are heavy users of clean electricity. Currently, they account for more than 30% all power purchase agreements (PPA), special contracts underwhich energy users agree to purchase renewable electricity at pre-agreed prices for up to 20 years. As PPAs become the default choice for data centres, it's reasonable to expect AI’s footprint to shrink rather than expand.

Handprint vs footprint

Another issue that gets lost in the public debate about AI’s environmental footprint is the technology’s potential role in reducing emissions. Intelligently deployed, AI can help entire industries cut their carbon footprints. That’s the case even for high polluting, hard-to-abate sectors such as heavy industry, construction, agriculture and transport.

Take aviation. It currently accounts for around 3% of global emissions. A large share of an aircraft’s impact on the climate comes in the form of contrails – the white, cloud-like plumes of vapour that are emitted by jet engines. Yet researchers have found that, by using AI to predict and avoid the routes that produce persistent contrails, the formation of such plumes can be cut by more than half. That application alone translates into an estimated reduction in global warming potential of around 0.5%, potentially outweighing AI’s own current footprint.

AI’s inherent emission reduction capacity is what the environmental products industry refers to as the carbon handprint of a technology. The opposite of a carbon footprint, the handprint describes a technology whose widespread deployment can have a positive systemic effect on greenhouse gas emissions. AI’s handprint is potentially very large. Not only can the technology be used to develop better climate models, but it can also help companies use resources more efficiently and cut waste.

None of this is to dismiss the concerns raised about AI in their entirety.

It is important to draw attention to the amount of power AI consumes. Not least because the increasing use of autonomous AI agents, for example, risks causing a steep rise in computing and in electricity demand.

What's more, it's impossible to know for certain when energy efficiency will begin to be a feature of AI’s development. In an extreme scenario, if capital investment and the use of agentic AI accelerate, then electricity consumption could outstrip the best available forecasts.

Yet what is also clear is that the technology isn’t destined to morph into an all-consuming beast. Thanks to its use of renewables and the development of more efficient power grids, AI’s carbon footprint is likely to remain manageable. Just as importantly, and unlike the tragic king of Greek myth, its power can be deployed to protect the planet rather than harm it.