Environmental footprint of AI
Artificial intelligence (AI) plays a dual role in the climate transition; whilst it has the potential to drive sustainability, it also comes with significant environmental costs. On the positive side, AI serves as a key tool in combating climate change by improving climate policy modelling, enabling systemic transformations, scaling research and innovation, and accelerating the shift to sustainable practices [4]. However, its entire life cycle—from software development to hardware waste management — presents significant environmental challenges.
We invited Elisabeth Nissen Eide to contribute with her insights and perspectives in a podcast [15]: AI and Sustainability - 30 minutter inn i fremtiden | Acast
Computational Power and Energy Consumption
The software life cycle of AI includes processes such as data collection, model training, validation, and deployment [3]. These steps are energy-intensive, requiring large-scale data centres that, with the slow expansion of renewable energy sources, increasingly depend on non-renewable resources: oil, natural gas, coal, and nuclear energy to power their operations [5].
S&P Global Ratings estimates that data centres could increase gas demand by an additional 3 to 6 billion cubic feet per day by 2030 [1]. They also report that, despite more than a decade of steady decline, U.S. coal demand is expected to rise in 2025, leading to a slower pace of coal plant retirements than previously estimated [10]. Additionally, major tech companies such as Microsoft, Google, and Amazon have signed power purchase agreements in 2024 totalling over three gigawatts (GW) of nuclear capacity to meet the growing energy demands of data centres [10]. The report highlights that the combined pressures of rising data centre consumption and ongoing energy security concerns will sustain revenues in the non-renewable sector for at least the next decade.
Corporate emissions underscore the issue as Microsoft reported in their Environmental Sustainability Report a 29.1% increase in greenhouse gas emissions since 2020 [6], while Google reported that emissions rose by 48% from 2019 levels in their environmental report [7]. Although a lot of this energy is labelled “carbon-neutral” using renewable energy credits, these offsets do not alter the actual methods of energy production, as noted by Wired [8].
Water Usage for Cooling
The cooling systems for data centres require vast amounts of freshwater, with reporting on water usage often missing from AI model transparency efforts [2]. This lack of accountability impedes innovations to ensure water sustainability. According to researchers at Cornell, training the GPT-3 language model in Microsoft’s U.S. data centres consumed an estimated 700,000 litres of clean freshwater [2]. They further emphasize that this level of consumption has significant societal implications, as the limited and uneven distribution of freshwater resources could exacerbate social conflicts and place additional strain on communities already grappling with water scarcity [2].
Efforts are being made to reduce freshwater consumption in data centres. Microsoft has developed a new data centre design that uses liquid cooling technology with a closed-loop system. This system continuously circulates water between servers and chillers to dissipate heat, eliminating the need for a constant freshwater supply [11]. Furthermore, the usage of sea water, recycled wastewater, and industrial water to cool data centres have already shown significant reductions in freshwater use.
Data centres also offer opportunities for a more circular approach to resource use. For example, the heat they generate could be repurposed for district heating in residential areas. This potential is particularly relevant in cooler regions, where excess heat can be efficiently redirected to support local energy needs.
Mining and Mineral Consumption
The hardware life cycle involves the production of essential components such as graphical processing units (GPUs) used in model training and inference, as well as the construction and operation of data centres [3]. The environmental impact of hardware is more complex and challenging to assess than that of software, as it involves resource extraction, energy consumption, and waste production. According to the International Resource Panel, the processing and extraction of raw materials—including fossil fuels, non-metallic minerals, metal ores, and biomass—account for approximately 50% of global greenhouse gas emissions and over 90% of global water stress and land-use-related biodiversity loss [13]. Additionally, the UN reports that electronic waste is increasing at a rate five times faster than documented e-waste recycling, highlighting a growing environmental challenge [14].
Additionally, the social sustainability of natural resource extraction remains a major concern, as minerals are often mined under exploitative conditions, including low wages, poor working environments, and excessive working hours. Nobel Peace Prize winner Denis Mukwege advises companies not to withdraw from factories suspected of unethical sourcing but instead to demand and enforce ethical standards to improve conditions [15].
Mitigation Strategies
Addressing AI’s environmental impact requires greater transparency and standardized reporting on energy and water usage. Providers must rethink their approach, placing sustainability at the core of innovation and adopting responsible technology development practices, such as optimizing efficiency across all aspects of operations [9]. Additionally, operational data should consider the timing and location of energy and water consumption, as geographical resource availability and weather variations can significantly affect efficiency [2].
Beyond energy consumption, reducing AI’s material footprint is also critical. Increasing the circular use of materials can help minimize the extraction of natural resources and reduce associated environmental impacts [13].
The industry is also becoming more aware of software’s environmental impact, leading to a growing emphasis on green software development. This approach prioritizes designing, development, and implementation of software that minimizes energy consumption and reduces its overall environmental footprint [12].
The UN Environment Programme (UNEP) further recommends that governments establish standardized metrics to evaluate AI’s environmental footprint, focusing on energy, water, and mineral resource consumption, along with emissions and e-waste production [3]. Tackling AI’s hidden water footprint is particularly urgent, given growing freshwater scarcity, prolonged droughts, and aging water infrastructure.
So, what can individuals and companies do? Start by assessing and documenting emissions, particularly the energy consumption of high-demand technologies like generative AI. Awareness is the first step and once companies understand their footprint, they can develop action plans to mitigate their impact. This includes evaluating when AI is necessary, choosing more energy-efficient models where possible, and prioritizing smaller, less resource-intensive alternatives [15].
Beyond AI usage, companies can set concrete sustainability requirements in their supply chains, particularly for resource mining, and promote circular economy practices by extending the lifespan of electronic devices. Transparency is key, and open discussions and clear reporting on environmental impact help drive accountability and collective action.
Sources:
- S&P Global Ratings. Data Centers: More Gas Will Be Needed To Feed U.S. Growth
- Cornell. Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models
- Artificial intelligence (AI) end-to-end: The environmental impact of the full AI life cycle needs to be comprehensively assessed
- World economic forum. What is AI's role in the climate transition and how can it drive growth?
- Heated. AI is guzzling gas
- Microsoft Environmental Sustainability Report 2024
- Google's 2024 Environmental Report
- Wired. Generative AI and Climate Change Are on a Collision Course
- Sustainable AI: Environmental Implications, Challenges and Opportunities
- Finansavisen
- Microsoft
- WHAT IS GREEN SOFTWARE?
- https://www.eea.europa.eu/en/topics/in-depth/resource-use-and-materials
- https://ewastemonitor.info/the-global-e-waste-monitor-2024/
- AI and Sustainability - 30 minutter inn i fremtiden | Acast