How bad is AI for the environment?

Though the emissions of modern innovations, like cars, computers and home heating are intimately understood, those of artificial intelligence (AI) are not. What we do know for certain is that AI’s energy consumption will rise exponentially in the coming years. Already, the infrastructures and materials needed to power AI solutions – from datacentres to chip-filled warehouses – are popping up faster than incumbent power grids can handle.

With concerns around AI’s environmental impact mounting – as tech firms remain guarded over the data – it is time to ask whether this technology will hinder or help the economy’s progress toward decarbonisation. In this instalment of Finextra’s Explainer series, we weigh up the potential positive and negative impacts of AI on the climate.

The energy consumption of datacentres

In Boxtown, south Memphis, Tennessee, sits xAI’s Colossus – the world's largest, fastest-built AI supercomputer. It is Elon Musk’s flagship datacentre, designed to train xAI's Grok chatbot and support his social media platform, X. Colossus is run on liquid coolers and massive Graphics Processing Unit (GPU) clusters, supplied by Nvidia.

Unsurprisingly, this sprawling plant consumes a vast amount of power, which is currently being produced by generators running on methane – a potent greenhouse gas, which the International Energy Agency (IEA) says is responsible for 30% of the rise in global temperatures since the Industrial Revolution.

At the time of writing, datacentres consume 1% of the world's electricity, but in the coming years this portion will rise considerably. For instance, in the United States, the cradle of AI development, datacentres accounted for 3.5% of the total electricity demand in 2024. By 2035, they will account for over double (8.6%), according to BloombergNEF. This increase will be driven by a surge in AI usage and workload – and accompanied by an increase in energy bills.

Given the size of their global footprint, the kind of energy used by datacentres matters. As it stands, however, most are like Colossus: dirty and largely free from pollution controls. So, will the rise of AI derail the economy’s transition to net zero? The reality is not so black and white.

Powered by brown and green energy

While in the short-term fossil fuels are dominating the energy supply of datacentres, clean energy might come to the fore in the long-term. Indeed, a number of agreements to buy renewable power and nuclear energy have been signed by tech firms. The future capacity needs of xAI’s Colossus, for example, are slated to be met by a combination of solar power, grid connections, and natural gas turbines – with the eventual goal of achieving 100% clean power by 2028.

Of course there are no guarantees. Macro-level trends will impact how tech firms respond to the evolving climate emergency. In the US – where the Trump administration is pushing a drill baby drill agenda – natural gas is expected to generate most of the electricity in datacentres over the coming decade. Meanwhile, China – another big player in AI development – has located most of its datacentres in its coal-heavy east, in areas such as Beijing, Shanghai, and Guangdong, where the supply of brown energy is fast and affordable.

So, though cheap solar energy is displacing coal in some instances, datacentres will likely pick up surplus capacity from idling power plants in the short-term. Put differently, the lower cost of renewables alone is not proving enough to catalyse economic decarbonisation.

The impact of LLM queries

What does the energy mix of datacentres mean for individual users of AI? What is the environmental impact of, for instance, a single large language model (LLM) search? Once again, tech companies have not been entirely forthcoming with this information.

Sam Altman, the CEO of OpenAI, has suggested that the energy needed for an average ChatGPT query would power a lightbulb for a few minutes. Official estimates are that a single text query costs 0.2-3 watt-hours, though deeper research or video production costs are far higher.

While these figures pale in comparison to flying or eating meat, it is the potential energy consumption of LLMs that is of concern. ChatGPT, for its part, is already used by several hundred million people every week. As more institutions plug LLMs and AI agents into their everyday operations, the toll on the climate will become severe.

Less discussed than the emissions – but just as concerning for the natural environment – is AI’s thirst. The UK government has predicted that by 2027 annual global water usage by AI could reach 4.2-6.6 billion cubic meters – equivalent to a significant fraction of a large country's total consumption. But, with global temperatures rising, fresh water – or blue gold, as it is known to those with limited access – is becoming a scarce commodity. In June 2025 the UK’s Environment Agency announced that by 2055 England will face a 5 billion litre shortfall in public water supplies every single day. In hotter parts of the world, the situation will be even more desperate.

Paying off the carbon debt

Needless to say, many AI providers see the technology in a far more positive light – and pivotal to the net zero effort.

In a recent report, Microsoft underlined some of the ways in which it sees AI being able to support decarbonisation, arguing that “AI-based systems can better integrate variable renewable energy into a stable electricity grid” and that they could “help reduce the cost of carbon capture by accelerating the discovery of new materials with desired properties.”

In practice, the list of potential environmental benefits of AI is far longer – including the ability to identify alternative proteins that mimic meat, improve car battery efficiency, and even spot greenhouse gas (GHG) leaks by fossil fuel projects within complex satellite data.

In an April 2025 report, the IEA argued that AI could help cut emissions by more than the datacentres produce. Google, meanwhile, believes AI has already helped it cut the energy needed for cooling in datacentres by 40%. So, does this mean AI could pay for its own pollution, and some?

For this to work, a bridge – paved with corporate action – will have to be erected between the environmental issues and the data. Even then, the carbon savings gleaned must not be used to justify further usage.

But just as AI can be used to catalyse decarbonisation efforts, it can be used to drive growth in brown industries. This effect is most pronounced in the oil and gas sector, where AI is helping firms identify ways to expand fossil fuel production, boost technically recoverable reserves, and cut the costs of offshore projects. In other words, it is triggering a fracking boom. President and CEO of Saudi Aramco, Amin Nasser, has said his company is now realising strong productivity gains, after having invested tens of billions of dollars in AI computing and hiring thousands of data scientists.

Progress vs. preservation

In light of this nuanced picture, how should governments, individuals, and investors respond? Will increasingly green markets eventually compel tech firms to reconsider their footprint, or should there be a moratorium on datacentre development, until clearer rules are rendered?

Laurence Tubiana, one of the architects of the Paris climate agreement, has proposed an AI tax to raise the funds needed for effective climate action. Others have called for the EU’s upcoming AI bill to classify fossil fuels as a high-risk application area. Whatever the strategy, it must be underpinned by proper disclosure by key AI players, like OpenAI, Nvidia, Microsoft, Google and Meta.

The green transition will be highly complex, and technology will undoubtedly have a role to play. Ultimately, a balancing act must be struck between meeting the technology demands of industry and ensuring progress does not knock economies off their decarbonisation course.