AI and the environment: which way will the balance tip?

Every question put to an AI system runs on chips in a data centre somewhere, and those chips need electricity, cooling and raw materials. At the same time, AI is helping to run power grids, forecast the weather and cut waste in buildings and factories. So is AI good or bad for the environment? The honest answer is that it can be both, and the outcome is still open.

The costs are real

The International Energy Agency estimates that data centres used about 485 terawatt-hours of electricity in 2025, slightly over 1.5 per cent of world demand, up 17 per cent in a single year. Use at AI-focused data centres rose by about half. By 2030 total demand is expected to reach around 950 terawatt-hours, roughly double. In the EU, data centres used 68 terawatt-hours in 2024, a figure the European Commission expects to rise to 114 by 2030.

The large technology companies show the strain. Google's electricity use rose 37 per cent in 2025, and its supply-chain emissions rose 25 per cent, mainly from AI hardware and construction. The company itself says AI is growing faster than the grid is decarbonising. Microsoft has reported a similar trend.

There is more than electricity. Data centres use water for cooling, and much more water is used indirectly in generating their power. Making chips is energy-intensive and uses potent greenhouse gases and specialised chemicals; the semiconductor industry's energy use more than doubled between 2015 and 2023. And servers become electronic waste: estimates for AI vary widely, from a few hundred thousand to several million tonnes by 2030, depending on how long the equipment is kept in use.

The efficiency is real too

AI is also becoming much more efficient. Google reports that the energy used for a typical text prompt to its Gemini assistant fell 33-fold in a single year, to about 0.24 watt-hours, about as much as a microwave oven uses in one second. The IEA notes that the computing power needed for a given level of language-model performance halves roughly every eight months. But efficiency alone does not settle the matter: when something becomes cheaper, we tend to use more of it, which is why total electricity use keeps rising.

What AI can save

The larger opportunity lies elsewhere in the economy. The IEA estimates that the wider use of existing AI applications could reduce emissions by up to 1,400 million tonnes a year by 2035, about three to four times the emissions it expects from data centres. Researchers at the Grantham Institute of the London School of Economics estimate that AI could help cut 3.2 to 5.4 billion tonnes a year by 2035 in just three sectors, power, food and transport, more than AI's own footprint.

Some of the examples are already working. Europe's weather centre, ECMWF, has used an AI forecasting system since February 2025 that improves some forecasts by up to 20 per cent while using around a thousand times less energy per forecast. The IEA says AI could unlock up to 175 gigawatts of extra transmission capacity on existing power lines, and save around 300 terawatt-hours of electricity in buildings. In agriculture, field trials show variable-rate technology reducing fertiliser use by up to about a third and smart sprayers halving pesticide use.

Notice that most of these gains come from specialised, predictive AI and optimisation, not from chatbots. The evidence that generative AI by itself reduces emissions at scale is still thin.

So which wins?

We believe AI can come out clearly ahead, and the best available research supports that view: the potential savings are several times larger than AI's own footprint. But it will not happen automatically. The savings are mostly modelled and still to be realised, while the costs are measured and rising today. Some researchers also warn that AI makes every industry more productive, including fossil fuels, which could add emissions.

The balance tips towards AI under clear conditions: data centres powered by clean electricity, continued gains in efficiency, and AI aimed at the tasks that matter most for the environment, such as energy, buildings, industry, transport, agriculture and science. Policy helps, and the EU is preparing a rating scheme for data centres, with the first labels expected in 2027.

What a company can do

Every business that uses AI is part of this balance. Use the smallest tool that does the job: a UNESCO and University College London study found that smaller, task-specific models and shorter prompts can cut energy use by up to 90 per cent. Avoid heavy 'reasoning' modes and image or video generation when they are not needed. Choose providers that run on renewable electricity and are open about their figures. Keep hardware in use for longer. And above all, point AI at real savings: fewer journeys, better heating and cooling control, less waste and less overstock.

Used thoughtfully, AI can help a business become both more productive and greener. That is the outcome worth working towards.

Published 1 October 2026 · © 2026 Habsburg Digital Ltd