Until recently, the discourse around AI ethics has primarily been occupied with human-centric topics like privacy, bias, and transparency. We have also asked if the algorithms behind AI give fair answers, if data have been handled in safe and secure ways, and if decisions taken by AI-systems could be explained. However, as complex neural networks, the core of AI-systems, emerged and computational requirements increased in all aspects of their use, a new pillar of trustworthy AI has emerged - one that is not merely ethical, but existential.
If AI is to be truly trustworthy, it must be environmentally sustainable. We are entering an era where planet-centricity must sit alongside human-centricity when AI-systems are designed, developed and used. However, a challenging paradox is that the technology straining our planetary boundaries - through unprecedented energy consumption, massive water usage for cooling, and intensive use of minerals harvested from a vulnerable nature – might also be one of our greatest hopes for accelerating the green transition.
The world needs frugal AI – AI that does not increase and even has the ability to decrease (over)consumption of the planet´s resources. Is it possible to promote development and use of frugal AI that will make the world more environmentally sustainable? Frugal AI solutions are being developed around the world, but the big question is how do we develop these solutions further and how can we make these solution scale in competition with the resource-hungry AI delivered by the global big-tech companies.
The process of making AI deserve people's trust, hence becoming trustworthy, is a normative task; that AI should be or become trustworthy. This differs from being a descriptive task, which states that something, in this case AI, is or will become trustworthy.
We will here explore some of the aspects of the balance between the resource-intensive reality of AI and its potential as a catalyst for a sustainable future, examining some of the technical and governance pathways necessary to ensure that as machines think, the planet does not pay the price.
1. The Physical Reality of Digital Intelligence
The physical reality of AI is far from the ethereal cloud metaphor often used to describe it. AI is a heavy industry. It is built on extensive use of critical minerals harvested from a vulnerable nature, cooled by millions of litres of water, and powered by massive amounts of electricity.
The Energy Hunger. Training a single large-scale AI-model can consume as much energy as many hundred homes are using in one year. But training is only the beginning. The inference phase - the billions of daily interactions users have with AI models is, by far, representing the largest part of AI’s carbon footprint. Data centres will soon account for 2% of global electricity demand - a figure projected to rise sharply as AI becomes ubiquitous across every sector in society and used by the entire population on earth, which recently reach 8 billion people and continues to grow.
The Thirst of the Machines. Cooling is the "silent" environmental cost of AI. Data centres require vast amounts of water to maintain the optimal temperature for high-performance GPUs (Graphics Processing Units). In regions already facing water scarcity, the expansion of data centres poses a direct threat to local ecosystems and human water security.
The Hunt for Critical Minerals. From the lithium and cobalt required for energy storage to the rare earth elements used in high-end semiconductors, AI’s hardware footprint involves mining in increasingly vulnerable ecosystems. The extraction of these minerals often results in biodiversity loss and significant carbon emissions, creating an "ecological debt" even before the AI systems are switched on. We also know that the world is facing destructive supply chains for these minerals – supply chains loaded with abuse of human rights and criminal activities.
2. Can AI Extend the Age of Oil?
Perhaps the most complex element of AI’s environmental impact lies in its influence on the global energy market. Recent analysis, including insights aligned with the latest International Energy Agency (IEA) reports, suggests a startling possibility: The rise of AI could extend the global demand for oil and gas, potentially pushing the "peak-oil" horizon toward 2050(3). Two primary drivers for this extension are highlighted;
Exploration and Extraction Efficiency. The major players in the oil and gas industry are among the most sophisticated adopters of AI. They use advanced machine learning to analyse old and new (seismic) data with unprecedented precision, discovering new oil and gas reserves and optimizing the extraction from mature wells that were previously considered uneconomical. By lowering the cost of production and increasing the success rate of exploration, AI effectively makes fossil fuels more resilient against the competition from new renewable energy sources.
Baseload Power Demands. As AI drives data centre demand for energy to new heights, the intermittency of wind and solar remains a challenge for the 24/7 "always-on" requirements of digital infrastructure. In the absence of massive-scale battery storage, there is a growing risk that power plants will turn to natural gas - or even extend the life of coal and oil-fired plants - to meet the immediate, surging energy needs of the AI revolution.
If AI is used primarily to squeeze the last drops of carbon from the earth more efficiently, it becomes an accelerant for climate change. However, if that same intelligence is redirected towards increasing production, distribution and use of renewable energy, the narrative changes. We – the humanity - know in which direction we need to go, but the question is how to go there?
3. Technological Pathways to Sustainable AI
New solutions that focus on reducing the environmental footprint of AI are emerging. These solutions are referred to as frugal AI solution. This happens at algorithmic level, at various architectural levels from how to build infrastructures and to the development and use of software solutions, and when it comes to the physical hardware performing the actual computations and handling storage of data.
Artificial intelligence (AI) can be described as the collection of theories, methods, algorithms and technologies that learn from data and experiences and has the ability to perform actions, physical and cognitive, perceived as intelligent.
Intelligence is about our (or a systems) ability to reason, make plans, solve tasks, think abstract, understand complex phenomena and learn fast (and from experience).
From Large to Small – Edge Intelligence. We have to move from brute force computing when developing and using AI-systems to smaller and preferably more elegant solutions. The dominating trend of "bigger is better" is being challenged by the rise of smaller models, e.g. Large Language Models (LLMs) are exchanged with Small Language Models (SLMs). By making the AI-models more knowledge-informed and optimizing algorithms to perform tasks with less data and fewer parameters, high levels of accuracy can be reached with less resources. The intelligence is moved to the edge, to smart phones and local computers rather than being conducted in massive, power-hungry data centres. In this picture, we also see that international cooperation, including joint development of high quality open-source software, must play an even more important role in the future than what is the case today.
Neuromorphic and Optical Computing, and even Quantum. Looking ahead, new hardware paradigms are emerging. Neuromorphic chips, which mimic the structure of the human brain (the most energy-efficient computer known to exist), process information only when needed, will reduce energy consumption. Similarly, optical computing using light instead of electricity to move data promises to eliminate the heat generated by traditional technology, potentially solving parts of the cooling challenge in data centres. Further ahead, with its energy efficiency, quantum computing might solve the energy challenge, at least for some specific computational task needed to develop and use AI-systems.
Data Sharing, Refinement and Reduction. Data sharing prevents redundant energy use by allowing users to build upon existing datasets rather than repeating resource-heavy collection processes and storage. Systematic data refinement ensures that only high-quality information is processed, which optimizes model performance and minimizes computational waste. Moreover, intentional data reduction lowers the overall carbon footprint of AI by decreasing the massive energy demands associated with storing and processing unnecessary volumes of information.
Circular Data Centres in an AI-powered Green Economy. As we work to reduce the need for large, energy-intensive data centres, will still need these centres. One way to go is to build the next generation of data centres as integrated energy sites. Rather than just consuming power, they must, in some way or another, produce energy and integrate itself into a circular economy. This includes using waste heat from servers to provide district heating and utilizing so called green hydrogen as backup power instead of power from classical diesel generators.
Recent and expected development of new technology gives hope, but the optimism must not postpone necessary action. Known, frugal and resource friendly, AI-solutions should be prioritized by users, and nations and the big AI-players should invest much more in research and innovation to seek environmental friendly AI solutions.
4. Governance Pathways to Sustainable AI
There is no precise definition of trustworthy AI, but inspired by EU we can say that AI is trustworthy if its precise (hereunder provide quantification of uncertainty), explainable and transparent, fair and inclusive, safe and secure, sustainable, well governed and that all this can be reach at scale. Hence, technology innovation for sustainable AI is important, but alone it is insufficient.
Its negative (and positive) environmental impact must be transparent, regulated and controlled. Governance pathways have to be developed and implemented across the globe. Moreover, organizations in private, public and civil sectors must undertake a transition from voluntary ethical pledges towards adapting a more rigorous framework of digital sustainability.
Multilateral Actions for a Sustainable Digital Future. International cooperation is essential to establish unified ethical and environmental standards that prevent a fragmented global landscape of AI regulations. By pooling resources and expertise, nations can collectively invest in green computing and energy-efficient infrastructure to mitigate the significant carbon footprint associated with digitalization in general and AI in particular. Such collaboration will also play a vital role in bridging both the green and the digital divide, ensuring that developing countries have equitable access to AI's benefits rather than being left behind in the technology race. Moreover, traditional Environmental, Social and Governance (ESG) frameworks are updated to mandate disclosures on the footprints of existing and future AI systems. To meet these requirements, data reduction must become an important element in all strategies, as minimizing the volume of information stored and processed is the most direct way to lower their reported ecological footprint. The EU AI Act has laid the groundwork for high-risk systems to be assessed for their environmental impact alongside ethics and safety . Furthermore, multilateral approaches allows for the sharing of best practices and data for solving cross-border challenges towards the age of renewables.
Standardized Metrics and Key Performance Indicators. There is a need for developing standardized metrics to create key performance indicators and labels for AIs consumption of resources (energy, water, minerals, etc.) These standards, with associated indicators, allow users and enterprises to identify "green AI" providers, creating more powerful market incentives for sustainability, e.g. data sharing/refinement/reduction. By focusing on open standards and high-quality, curated datasets, developers can improve model efficiency and achieve the superior ratings required to remain competitive. There are ongoing work to create such standardized metrics, however it is difficult due to so-called problem shifts - solving one problem often creates new problems .
Prioritize the Earth's Health alongside Human/User Needs. Designers and developers of AI-systems must be encouraged to use methods and frameworks that prioritizes the Earth's health alongside human needs. This approach shifts the focus from user experience to also include the environmental impact of technology. Before launching new AI systems, enterprises must weigh the societal benefits against the ecological costs. This philosophy also champions data sharing to prevent different users from wasting energy on redundant tasks. By collaborating on shared models, we can significantly reduce the total amount of resources used for training and use of AI. Ultimately, the goal is to ensure that technological progress works in harmony with the nature. Frugal AI must be developed and used.
A Comprehensive Tax System for Digital Resources (Data). The collection, storage, and use of digital resources, preferably data, can be taxed in the same way as other resources that place a burden on the environment and common infrastructure, for example through fees linked to data volume, energy consumption, and the carbon footprint of data centres. Organisations that collect and (commercially) exploit large amounts of data should be subject to progressive taxes based on how much data they store, how long it is retained, and how energy-intensive the processing is. This can encourage more efficient data use, incentives for data reduction and sharing, and greater adoption of energy-efficient solutions. Consumers, in turn, can be taxed indirectly through progressive charges on data traffic or electricity consumption related to particularly data-intensive activities such as high-resolution streaming or extensive cloud storage. Such a progressive model would keep basic use affordable or tax-free, while making excessive consumption more expensive. In this way, the tax system can promote more sustainable digital practices without limiting necessary access to these services.
The geopolitical situation is challenging, and most likely it will not become less demanding in the years to come. AI is deeply embedded in geopolitics, which in turn will require clear and strong governance - both to make AI itself sustainable and to leverage AI in finding sustainable solutions within and across all areas of society.
5. AI as a Catalyst for the Green Transition
We have so far discussed AI in terms of its footprint - the resources it consumes and the emissions it drives - but its handprint - the positive impact it can have across sectors - is also transformative. AIs has impact across all sectors. Here are some examples where AI-solutions either has contributed or most likely will contribute to the green transition of society.
Smarter Energy Grids towards the Age of Renewables. Switching from fossil fuels to wind and solar isn’t just a matter of building more turbines or panels - it’s a massive data challenge. Renewable energy is decentralized and variable: the sun doesn’t always shine and the wind doesn’t always blow. That’s where AI comes in. Machine learning can balance electricity supply and demand in real time, predict weather-driven fluctuations, and optimize grids from global to local scales. AI has already been shown to increase grid resilience and efficiency while reducing greenhouse gas emissions across diverse energy projects. Several recently developed national strategies across the world aim to use AI to coordinate renewable sources and reduce grid instability, which is positive and a crucial step towards lowering reliance on coal and other fossil fuels. A key question is how to handle data challenges in future AI-driven energy grids?
Precision Agriculture - Less Waste, More Food. Farms are becoming smarter and greener using AI solutions. Drones and sensors paired with machine learning functionality help farmers deliver water, fertiliser, and pesticides exactly where needed, in many cases down to the level of individual plants. This reduces harmful runoff of pollutants into waterways, improves crop yields, and cuts greenhouse gas emissions compared to more conventional farming methods .
AI-Empowered Physical Infrastructures. AI makes physical infrastructure greener by optimizing resource efficiency throughout an asset's entire lifecycle, from the potential of using generative design methods to minimize the use of materials to employing smart systems that reduce operational waste of energy. It also enhances sustainability through predictive maintenance, which identifies structural issues early to extend the lifespan of existing facilities and reduce the environmental cost of new development. By integrating real-time data and advanced analytics, AI can transform traditional infrastructure to become intelligent, resilient networks that prioritize circularity.
Reinventing Mobility – on Land, at Sea, and in the Sky. AI is dramatically reshaping how we move things - goods and people. On land, real-time traffic optimisation reduces congestion and emissions. At sea, intelligent navigation systems analyse environmental data to plan fuel-efficient routes for vessels. In the sky, AI improves air traffic flow and maintenance scheduling, cutting noise, fuel use, and delays. Across transport modes, these advancements support a more efficient and sustainable global mobility network.
Smarter Healthcare That Prevents Rather Than Reacts. AI has the potential of shifting healthcare toward prevention and sustainability. By analysing health data, AI can detect early disease markers and recommend interventions long before conditions worsen. Hence, reducing intensive treatments that consume considerable resources. In addition to improved diagnostics and more efficient treatment of individual patients, telemedicine powered by AI also has the potential of cutting the need for travel of both patients and health workers, lowering the healthcare sector’s carbon footprint. AI can also optimise healthcare logistics, from supply chains to energy usage, making healthcare systems more resilient and less wasteful.
Equity in an AI-Powered Green Economy (learning). For AI’s benefits to be truly global, access and education must be equitable. AI-driven learning systems can personalise teaching, providing high-quality, adaptive education for underprivileged children and education in regions with limited access to teachers. This lowers barriers to knowledge and equips students with skills for tomorrow’s green economy. By democratizing learning and opportunity, AI helps prevent a green divide, fostering social sustainability alongside technological progress.
From Reactive to Proactive Stewardship of the Earth. AI is turning satellite, radar, and sensor data into actionable intelligence about our planet. Rather than waiting for disasters to strike, these systems can anticipate threats and alert communities early. AI platforms can detect illegal deforestation patterns before they grow, monitor glacier melt through cloud-penetrating radar, and map flood risk days in advance, giving authorities valuable time to act. This shift from reactive to proactive stewardship means that AI can become a kind of digital nervous system for the Earth, helping to preserve ecosystems and protect vulnerable populations.
Discovering New Materials. AI is revolutionizing materials science by scanning millions of molecular combinations faster than traditional lab work ever could. This speeds up the hunt for high-capacity batteries, next-generation superconductors, and advanced carbon capture membranes - all essentials for a post-carbon economy. AI-driven discovery has already shown promise in accelerating innovation in these fields, helping reduce industrial emissions and supporting sustainability goals.
A Force Multiplier for Scientific Discovery. Artificial Intelligence (AI) has emerged as a transformative power in scientific inquiry, often referred to as AI in Science. This field encompasses the application of machine learning (ML) and related computational techniques to accelerate scientific discovery, enhance predictive accuracy, and enable the analysis of increasingly complex data in and across scientific disciplines. AI is also accelerating the scientific process itself. With respect to the green transition of society these machine-learning models can simulate complex ecological systems, forecast biodiversity loss, and model atmospheric physics with unprecedented precision. These tools enable scientists to test hypotheses and refine solutions in silico, dramatically reducing the time between discovery and real-world impact. By processing the vast data streams of the natural world, AI can provide knowledge necessary for sustainable adaptation to the new climate and nature regime and potentially offer a roadmap towards a regenerative relationship with the Earth’s resources.
6. Concluding Remarks
To build a sustainable AI-based future for all, we must, in some way or another, redefine our relationship to AI. Right now, AI is a resource-hungry giant accelerating the ecological crisis, but AI also has the potential to become a powerful tool for green stewardship of the planet´s resources.
AI’s massive energy appetite can delay our exit from the fossil fuel era. To avoid this, we must pivot towards a situation where development and use of AI stay within planetary boundaries. Among other things, this means moving away from using massive, resource-intensive global systems in favour of smaller, smarter and more environmentally friendly systems. The future needs frugal AI!
Ultimately, the true test of AI might be its ability to help humanity decouple (economic) growth from environmental degradation.
(1) International Energy Agency (IEA) (2025). Energy and AI.
(2) F&D Magazine, International Monetary Fund. Inside the AI-led resource race.
(3) International Energy Agency (IEA) (2025). World Energy Outlook 2025.
(4) The Guild (2023). Insight Paper 5. The Transition Century: The role of digital research for a green transition of society?
(5) European Parliament (2024). The EU Artificial Intelligence Act (EU AI Act).
(6) ISO/IEC TR 5469:2023. Artificial intelligence — Functional safety and AI systems.
(7) European Commission (2019). Ethics Guidelines for Trustworthy AI.
(8) Artificial Intelligence in Agriculture: 2022 Report, OECD Publishing.
