The year 2025 marks a symbolic and operational turning point in how artificial intelligence is perceived. Long portrayed as primarily a software-based technology, AI is now recognized as a heavy-duty physical infrastructure that consumes energy, water, metals, and land. According to the World Economic Forum, more than 75% of computing capacity deployed in 2025 will directly or indirectly serve AI applications, particularly generative and agent-based ones1. This widespread adoption brings to light environmental externalities that were previously marginalized, placing AI at the center of debates on digital sustainability.
Energy: Structural Growth That Is Hard to Contain
In 2025, energy consumption related to AI continues to rise. The International Energy Agency estimates that data centers—an increasing proportion of which are dedicated to AI—now account for nearly 4% of global electricity consumption, up from 2% in 20202. Training large models remains extremely energy-intensive, but it is primarily inference—that is, large-scale daily use—that accounts for the bulk of consumption. The proliferation of queries, autonomous agents, and real-time systems is transforming AI into a continuous load that is difficult to smooth out, putting strain on local power grids.
Water: From Blind Spot to Center Stage
The water footprint of AI emerged as a major issue in 2025. Data center cooling systems consume large volumes of freshwater, sometimes in regions already facing water stress. Research published by the University of California indicates that a data center specializing in AI can consume several hundred thousand cubic meters of water per year, depending on the cooling technologies used3. This strain on water resources fuels local tensions and raises questions of environmental justice, particularly when the economic benefits of AI accrue to actors far removed from the affected regions.
Carbon footprint: heavily dependent on infrastructure choices
The carbon footprint of AI in 2025 appears to vary widely. A study by MIT shows that the carbon footprint of the same computational workload can vary by a factor of ten depending on the local energy mix4. While some players are investing heavily in renewable energy or low-carbon electricity contracts, the overall growth in usage partially offsets these gains. The paradox is clear: AI contributes to energy optimization in many sectors, but its own development remains carbon-intensive on a large scale.
Data Centers and Local Communities: Growing Tensions
The concentration of AI infrastructure is exacerbating regional imbalances. By 2025, several regions in Europe and North America will see conflicts arise over the use of electricity, water, and land. The European Commission notes that in some areas, data centers account for more than 10% of local electricity consumption5. AI thus becomes a key issue in regional planning, forcing public authorities to balance economic attractiveness, energy security, and social acceptability.
Businesses: From Performance to Responsibility
For businesses, 2025 marks a shift in approach. AI remains a key driver of competitiveness, but its environmental impact is becoming a strategic consideration. According to PwC, 64% of large European companies now say they factor AI’s environmental footprint into their technology investment decisions6. This shift is reflected in more selective choices, a reduction in the number of models, resource sharing, and balancing maximum performance against operational efficiency.
Key drivers of technological simplicity
In light of these findings, solutions began to emerge in 2025.
- the development of more compact and specialized models,
- optimizing inference, the primary source of long-term power consumption,
- improving the energy efficiency of AI chips,
- increased use of renewable energy and heat recovery.
According to Google and Microsoft, certain optimizations already reduce energy consumption per query by 30 to 40% in specific use cases7. However, these advances are still insufficient to fully offset the growth in usage.
Gradually increasing regulation and institutional pressure
The year 2025 also sees the emergence of a more rigorous institutional framework. The OECD and the World Economic Forum are calling for environmental indicators to be systematically integrated into AI governance8. In Europe, while the AI Act focuses primarily on societal and security risks, environmental concerns are gradually entering regulatory discussions. The trend is clear: the ecological footprint is becoming a criterion for technological legitimacy.
2026: What Environmental Pathways for AI?
By 2026, projections paint a mixed picture. The International Energy Agency forecasts an additional 15% to 20% increase in electricity demand related to AI if its use continues to grow at the same pace9. However, several research firms estimate that efficiency gains in inference could reduce the energy footprint per interaction by 30 to 50% by 202610. In terms of water use, investments are shifting toward closed-loop cooling systems and regions less exposed to water stress11. Regarding carbon, the partial relocation of computational workloads and the integration of environmental indicators into AI strategies are expected to become standard practice for large organizations. The central question will no longer be merely how much energy AI consumes, but how, where, and for what purposes it utilizes resources.
2025–2026: Toward an Environmentally Responsible AI
The 2025 environmental assessment is not a final verdict but a turning point. AI is neither inherently sustainable nor irreparably harmful; it reflects the industrial, economic, and political choices that shape it. As AI establishes itself as critical infrastructure, its sustainability becomes a prerequisite for its social and political acceptance. The challenge in the coming years will not be to slow down innovation, but to steer artificial intelligence onto a path compatible with planetary boundaries, by making clear trade-offs between performance, usage, and environmental responsibility.
Learn more
To learn about a practical and positive application of artificial intelligence in the protection of natural ecosystems, you can read our article on the use of AI for the preservation of marine biodiversity in Polynesia. This analysis shows how machine learning models are being used to track, understand, and protect whale populations in the face of human and environmental pressures: Ocean IA: How Artificial Intelligence Contributes to Whale Conservation in Polynesia
References
1. World Economic Forum. (2024). Global AI Infrastructure Outlook.
https://www.weforum.org
2. International Energy Agency. (2024). Electricity 2024: Data Centers and AI.
https://www.iea.org
3. University of California, Riverside. (2023). Water consumption of AI systems.
https://www.nature.com
4. MIT. (2024). Carbon intensity of AI workloads.
https://www.mit.edu
5. European Commission. (2024). Energy consumption of data centers in the EU.
https://digital-strategy.ec.europa.eu
6. PwC. (2024). AI and environmental responsibility in Europe.
https://www.pwc.com /a>
7. Google Sustainability. (2024). Energy-efficient AI infrastructure.
https://sustainability.google
8. OECD. (2024). Sustainable and Trustworthy AI.
https://www.oecd.org
9. International Energy Agency. (2025). Electricity and AI Outlook.
https://www.iea.org
10. McKinsey & Company. (2025). The next frontier of sustainable AI.
https://www.mckinsey.com
11. World Resources Institute. (2024). Water stress and data center expansion.
https://www.wri.org