AI’s green promise hides a growing environmental cost

Artificial intelligence (AI) promises efficiency gains and accelerated sustainability innovation in the clean technology sector. However, a new study warns that this transformation comes with a growing environmental cost that could undermine its own benefits if left unmanaged.

A systematic review, titled "Navigating the Environmental Paradox of AI: A Decision Framework for Clean Technology Practitioners" and published in the journal Clean Technologies, analyses how AI's environmental trade-offs. Drawing on 73 peer-reviewed studies from 2021 to 2025, the research introduces a structured framework to evaluate both the environmental benefits and costs of AI deployment across its lifecycle.

AI's environmental promise: Efficiency, optimization, and innovation

The research identifies five major ways in which AI contributes positively to environmental sustainability, highlighting its growing role as a core enabler of clean technology systems.

• Energy optimization: AI systems improve efficiency across industrial processes, buildings, and energy grids. The study finds that AI-driven systems can reduce industrial energy consumption by 20 to 40 percent and improve building energy efficiency by up to 35 percent. Smart grids powered by machine learning also enable better integration of renewable energy, reducing waste and improving grid stability.
• Production enhancement: AI applications such as precision agriculture significantly reduce resource use. The analysis shows that fertilizer and water consumption can drop by 20 to 30 percent while maintaining crop yields. In manufacturing, predictive maintenance and process optimization cut material waste and energy consumption, making industrial systems more sustainable.
• Green innovation: Machine learning models are increasingly used to discover new materials for batteries, carbon capture, and renewable energy technologies. These tools shorten research timelines and improve the efficiency of design processes, potentially speeding up the transition to low-carbon technologies.
• Resource conservation: AI supports circular economy practices. Predictive analytics helps extend product lifecycles, reduce waste, and optimize supply chains. Transportation emissions, for example, can be reduced by up to 30 percent through AI-driven logistics optimization.
• Precision applications in sectors such as healthcare and infrastructure reduce unnecessary resource use. AI systems can lower medical waste by minimizing redundant procedures and detect leaks in water or energy systems before they escalate into larger losses.

Together, these benefits position AI as a powerful tool for environmental optimization. However, the study makes clear that these gains do not come without significant trade-offs.

The hidden costs: Energy, water, and resource burdens

Alongside its benefits, AI introduces a range of environmental costs that are expanding rapidly as adoption scales. The most significant burden is energy consumption. Training large AI models requires massive computational power. The study reports that training a single advanced model can consume thousands of megawatt-hours of electricity, while large-scale AI platforms require hundreds of megawatt-hours daily for ongoing operations. Global data center electricity demand is projected to exceed 1,000 terawatt-hours by 2030, rivaling the energy consumption of entire countries.

Closely linked is water usage, driven by data center cooling requirements. Depending on the technology and climate, AI infrastructure can consume between 1.8 and 12 liters of water per kilowatt-hour of electricity. Training a single large model can indirectly use over 700,000 liters of water. By 2030, annual data center water consumption in the United States alone could exceed 1 trillion liters.

The study also highlights the growing issue of electronic waste. AI relies on specialized hardware such as GPUs, which have short lifecycles of two to three years. This rapid turnover contributes to rising e-waste volumes, with global recycling rates remaining below 22 percent.

Apart from operational impacts, infrastructure expansion adds further environmental strain. Data centers require large amounts of land, steel, and concrete, contributing to habitat disruption and increased emissions during construction.

Lastly, supply chain extraction for AI hardware involves mining rare earth materials such as lithium and cobalt. These processes are associated with deforestation, water pollution, and high carbon emissions, often concentrated in regions with weaker environmental regulations.

The study estimates that the information and communications technology sector already accounts for nearly 4 percent of global greenhouse gas emissions, with AI expected to significantly increase this share in the coming years.

The S-Curve reality: Why AI's impact changes over time

AI's environmental impact follows a predictable S-curve pattern over time. In the initial phase (0 to 2 years), AI deployment typically reduces emissions. Efficiency gains from optimization outweigh the relatively small infrastructure footprint, resulting in net environmental benefits.

This is followed by a rebound phase (2 to 5 years), where environmental costs begin to rise. As AI adoption expands, demand for data centers, computing power, and hardware increases. This leads to higher energy use, water consumption, and emissions, often offsetting earlier gains. The study identifies this as a form of the "digital rebound effect," where efficiency improvements lead to increased overall consumption.

In the long-term phase (5+ years), AI has the potential to return to net positive environmental impact, but only under specific conditions. These include widespread adoption of renewable energy, improved infrastructure efficiency, and stabilization of hardware growth.

Notably, the study warns that many current AI deployments are entering or already in the rebound phase, where environmental impacts are at their highest.

Geography matters: A 60-fold difference in outcomes

The research finds that location plays a decisive role in determining AI's environmental footprint. In regions powered by renewable energy, such as Norway or Iceland, AI systems can operate with minimal carbon emissions. In contrast, deployments in coal-dependent regions can generate up to 40 times more emissions for the same workload.

When combined with water scarcity and climate conditions, the variation can reach up to 60 times between optimal and challenging locations. The study categorizes regions into three groups:

• Optimal regions: High renewable energy, abundant water, and cool climates. These areas can achieve net environmental benefits quickly.
• Challenging regions: Fossil fuel dependence, water scarcity, and high temperatures. AI deployment here may never become environmentally positive.
• Transition regions: Mixed energy sources and moderate resources, where outcomes depend on policy and infrastructure choices.

This geographic disparity underscores the need for location-specific deployment strategies rather than one-size-fits-all approaches.

A decision framework for sustainable AI deployment

To address these complexities, the authors introduce an Environmental Asset-Cost Framework, which evaluates AI across five benefit categories and five cost categories while incorporating time and geography as key variables.

The framework provides actionable guidance for practitioners, including:

• Prioritizing AI deployment in regions with low-carbon energy and sufficient water resources
• Pairing AI infrastructure with new renewable energy capacity
• Adopting alternative cooling technologies to reduce water use
• Extending hardware lifecycles to limit e-waste
• Implementing energy caps to control rebound effects

Sustainable AI deployment requires planning beyond short-term gains, with a focus on long-term lifecycle impacts, the study asserts.