King's College London

Recently, a young entrepreneur contacted me to discuss the sustainability of her startup’s digital infrastructure. She explained how both sustainability and digital innovation sit at the core of her business model, and had created an Artificial Intelligence (AI) powered platform to support more sustainable fashion design. Using the platform, fashion designers would no longer need to physically assemble textile samples but could now run all the tests virtually with AI-generated visualisations of the clothing. This could significantly improve the sustainability of the fashion industry, considering the tremendous negative impacts of fashion design, especially fast fashion, on the planet and people.

The entrepreneur wanted to make sure her platform would lead to a real increase in sustainability, even when indirect impacts were considered. She had heard about certain environmental impacts of data centres and wondered if this challenged the sustainability of her platform. More broadly put, did “dematerializing” this design process through AI add to or subtract from sustainability?

The impact of AI: energy as a starting point

An intuitive first step would be for the entrepreneur to estimate the environmental impact of the AI model underlying her platform. She had integrated a cloud-based large language model, which was trained and hosted by a large technology company, and ran in a network of data centres. Globally, the electricity consumption by data centres has grown by around 12% per year between 2015 and 2024, with most of this growing energy demand being met by fossil fuel sources. Meanwhile, a quarter of the world’s new gas-powered energy facilities are under development in the United States, of which one-third are set to directly deliver power to data centres, on-site.

Calculating her platform’s energy consumption and emissions was thus a relevant and important exercise. But couldn’t our entrepreneur simply visit the sustainability report of the AI company and deduce her impact there? In February, the International Telecommunication Union published a set of guidelines on how the emissions and other sustainability impacts of AI should be measured. The report highlighted, unfortunately, that AI companies currently apply different calculation methods, rendering many sustainability claims incomparable. For instance, some AI companies measured the environmental impact of AI through its total electricity use in the use phase. Others also included embedded emissions in hardware manufacturing. A lack of transparency from AI companies thus rendered the task at hand very difficult.

Luckily, the entrepreneur could count on some publicly available estimates to calculate and communicate the energy consumption of her platform, which could be viewed per use of (or inference with) the AI model. Integrating an energy use counter in the interface then becomes a possibility, for example, with the AI Energy Score by Hugging Face. Once integrated in the platform, she could estimate the total energy use and show what impact a single prompt has on energy consumption. Perhaps this could incentivize consumers to use the model more sparingly.

Continuing toward a full picture: on rebounds, water and waste

This brings us to the role of consumption in the sustainability impact of AI and other digital technologies. The goal of the entrepreneur’s platform was for her consumers to decrease their resource needs and become more sustainable. But what if consumers suddenly replace one physical garment with thousands upon thousands of uses of generative AI, increasing the carbon emissions to surpass that of any real-world fashion design? Such so-called “rebound effects” are a critical issue for any digital solution and should be addressed by aiming to limit impacts on the consumption side, be it through awareness-raising or actively limiting usage on the platform¹.

Energy use is, however, only one type of impact, and the ambitious entrepreneur was interested in considering other impacts as well. In terms of water consumption, the startup’s data transfers may lead to water extraction for data centre cooling and electricity generation. A large data centre consumes as much water as a large town of 10 to 50,000 residents, which could be detrimental to already water-scarce areas. Alternatively, another often overlooked impact of AI is the resulting e-waste. The GPUs performing the calculations do not last forever, and e-waste quantities could add up to around 230 million kilograms of additional waste by 2030, equivalent to the e-waste produced by current-day Sweden. However, again due to the lack of transparency from AI firms and the high variability across data centres and regions, calculating both impacts per prompt would require the entrepreneur to perform some educated guesswork.

Seeing through the hype in AI for sustainability

In the past thirty years or so, many organisations and entrepreneurs like in our example have found themselves cleaving wave after wave of digital transformation, each wave a harbinger of new promises of higher productivity and efficiencies. One of the latest waves, the introduction of AI, has been accompanied by a discourse largely shaped by tech companies. Their narratives highlight the tremendous potential of AI for healthcare, climate change, and our quality of life. Although this positive potential is indeed large, the current exponential increase in data centre capacity is not intended to power research on protein folding for novel medicines, but mostly for the use of generative AI.

Due to the global nature of digital infrastructures and their impacts on sustainability, it is critical that those wishing to apply AI for sustainability, just like our ambitious entrepreneur, strive to rigorously estimate and communicate the sustainability impacts of AI integration. This helps avoid the negative indirect impacts from AI eclipsing its positive sustainability contributions.

---

^[1] This promotes what the German professor Tilman Santarius calls “digital sufficiency”, in which users are stimulated to use “as much digitalisation as necessary, yet as little as possible” to avoid rebound effects and thus improve sustainability. 

 

Arjen Van de Walle, is a PhD candidate at Vrije Universiteit Brussel, Belgium and  a visiting researcher at the Centre of Sustainable Business