17  Sustainable AI

Resources: Slides, Videos, Exercises

DALL·E 3 Prompt: 3D illustration on a light background of a sustainable AI network interconnected with a myriad of eco-friendly energy sources. The AI actively manages and optimizes its energy from sources like solar arrays, wind turbines, and hydro dams, emphasizing power efficiency and performance. Deep neural networks spread throughout, receiving energy from these sustainable resources.

Purpose

How do environmental considerations influence the design and implementation of machine learning systems, and what principles emerge from examining AI through an ecological perspective?

Machine learning systems inherently require significant computational resources, raising critical concerns about their environmental impact. Addressing these concerns requires a deep understanding of how architectural decisions affect energy consumption, resource utilization, and ecological sustainability. Designers and engineers must consider the relationships between computational demands, resource utilization, and environmental consequences across various system components. A systematic exploration of these considerations helps identify key architectural principles and design strategies that harmonize performance objectives with ecological stewardship.

Learning Objectives

• Define the key environmental impacts of AI systems.
• Identify the ethical considerations surrounding sustainable AI.
• Analyze strategies for reducing AI’s carbon footprint.
• Describe the role of energy-efficient design in sustainable AI.
• Discuss the importance of policy and regulation for sustainable AI.
• Recognize key challenges in the AI hardware and software lifecycle.

17.1 Overview

Machine learning has become an essential driver of technological progress, powering advancements across industries and scientific domains. However, as AI models grow in complexity and scale, the computational demands required to train and deploy them have increased significantly, raising critical concerns about sustainability. The environmental impact of AI extends beyond energy consumption, encompassing carbon emissions, resource extraction, and electronic waste. As a result, it is imperative to examine AI systems through the lens of sustainability and assess the trade-offs between performance and ecological responsibility.

Developing large-scale AI models, such as state-of-the-art language and vision models, requires substantial computational power. Training a single large model can consume thousands of megawatt-hours of electricity, equivalent to powering hundreds of households for a month. Much of this energy is supplied by data centers, which rely heavily on nonrenewable energy sources, contributing to global carbon emissions. Estimates indicate that AI-related emissions are comparable to those of entire industrial sectors, highlighting the urgency of transitioning to more energy-efficient models and renewable-powered infrastructure.

Beyond energy consumption, AI systems also impact the environment through hardware manufacturing and resource utilization. Training and inference workloads depend on specialized processors, such as GPUs and TPUs, which require rare earth metals whose extraction and processing generate significant pollution. Additionally, the growing demand for AI applications accelerates electronic waste production, as hardware rapidly becomes obsolete. Even small-scale AI systems, such as those deployed on edge devices, contribute to sustainability challenges, necessitating careful consideration of their lifecycle impact.

This chapter examines the sustainability challenges associated with AI systems and explores emerging solutions to mitigate their environmental footprint. It discusses strategies for improving algorithmic efficiency, optimizing training infrastructure, and designing energy-efficient hardware. Additionally, it considers the role of renewable energy sources, regulatory frameworks, and industry best practices in promoting sustainable AI development. By addressing these challenges, the field can advance toward more ecologically responsible AI systems while maintaining technological progress.

17.2 Ethical Responsibility

17.2.1 Long-Term Viability

The long-term sustainability of AI is increasingly challenged by the exponential growth of computational demands required to train and deploy machine learning models. Over the past decade, AI systems have scaled at an unprecedented rate, with compute requirements increasing 350,000× from 2012 to 2019 (Schwartz et al. 2020). This trend shows no signs of slowing down, as advancements in deep learning continue to prioritize larger models with more parameters, larger training datasets, and higher computational complexity. However, sustaining this trajectory poses significant sustainability challenges, particularly as the efficiency gains from hardware improvements fail to keep pace with the rising demands of AI workloads.

Historically, computational efficiency improved with advances in semiconductor technology. Moore’s Law, which predicted that the number of transistors on a chip would double approximately every two years, led to continuous improvements in processing power and energy efficiency. However, Moore’s Law is now reaching fundamental physical limits, making further transistor scaling increasingly difficult and costly. Dennard scaling, which once ensured that smaller transistors would operate at lower power levels, has also ended, leading to stagnation in energy efficiency improvements per transistor. As a result, while AI models continue to scale in size and capability, the hardware running these models is no longer improving at the same exponential rate. This growing divergence between computational demand and hardware efficiency creates an unsustainable trajectory in which AI consumes ever-increasing amounts of energy.

The training of complex AI systems like large deep learning models demands startlingly high levels of computing power with profound energy implications. Consider OpenAI’s state-of-the-art language model GPT-3 as a prime example. This system pushes the frontiers of text generation through algorithms trained on massive datasets, with training estimated to require 1,300 megawatt-hours (MWh) of electricity—roughly equivalent to the monthly energy consumption of 1,450 average U.S. households (Maslej et al. 2023). In recent years, these generative AI models have gained increasing popularity, leading to more models being trained with ever-growing parameter counts.

Research shows that increasing model size, dataset size, and compute used for training improves performance smoothly with no signs of saturation (Kaplan et al. 2020), as evidenced in Figure 17.1 where test loss decreases as each of these three factors increases. Beyond training, AI-powered applications such as large-scale recommender systems and generative models require continuous inference at scale, consuming significant energy even after training completes. As AI adoption grows across industries from finance to healthcare to entertainment, the cumulative energy burden of AI workloads continues to rise, raising concerns about the environmental impact of widespread deployment.

Figure 17.1: Performance improves with compute, dataset set, and model size. Source: Kaplan et al. (2020).

Kaplan, Jared, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. “Scaling Laws for Neural Language Models.” ArXiv Preprint abs/2001.08361 (January). http://arxiv.org/abs/2001.08361v1.

Beyond electricity consumption, the sustainability challenges of AI extend to hardware resource demands. High-performance computing (HPC) clusters and AI accelerators rely on specialized hardware, including GPUs, TPUs, and FPGAs, all of which require rare earth metals and complex manufacturing processes. The production of AI chips is energy-intensive, involving multiple fabrication steps that contribute to Scope 3 emissions, which account for the majority of the carbon footprint in semiconductor manufacturing. As model sizes continue to grow, the demand for AI hardware increases, further exacerbating the environmental impact of semiconductor production and disposal.

The long-term sustainability of AI requires a shift in how machine learning systems are designed, optimized, and deployed. As compute demands outpace efficiency improvements, addressing AI’s environmental impact will require rethinking system architecture, energy-aware computing, and lifecycle management. Without intervention, the unchecked growth of AI models will continue to place unsustainable pressures on energy grids, data centers, and natural resources, underscoring the need for a more systematic approach to sustainable AI development.

The environmental impact of AI is not just a technical issue but also an ethical and social one. As AI becomes more integrated into our lives and industries, its sustainability becomes increasingly critical.

17.2.2 Ethical Issues

The environmental impact of AI raises fundamental ethical questions regarding the responsibility of developers, organizations, and policymakers to mitigate its carbon footprint. As AI systems continue to scale, their energy consumption and resource demands have far-reaching implications, necessitating a proactive approach to sustainability. Developers and companies that build and deploy AI systems must consider not only performance and efficiency but also the broader environmental consequences of their design choices.

A key ethical challenge lies in balancing technological progress with ecological responsibility. The pursuit of increasingly large models often prioritizes accuracy and capability over energy efficiency, leading to substantial environmental costs. While optimizing for sustainability may introduce trade-offs—such as increased development time or minor reductions in accuracy—it is an ethical imperative to integrate environmental considerations into AI system design. This requires shifting industry norms toward sustainable computing practices, such as energy-aware training techniques, low-power hardware designs, and carbon-conscious deployment strategies (Patterson et al. 2021).

Beyond sustainability, AI development also raises broader ethical concerns related to transparency, fairness, and accountability. Figure 17.2 illustrates the ethical challenges associated with AI development, linking different types of concerns—such as inscrutable evidence, unfair outcomes, and traceability—to issues like opacity, bias, and automation bias. These concerns extend to sustainability, as the environmental trade-offs of AI development are often opaque and difficult to quantify. The lack of traceability in energy consumption and carbon emissions can lead to unjustified actions, where companies prioritize performance gains without fully understanding or disclosing the environmental costs.

Figure 17.2: Ethical challenges in AI development. Source: COE

Addressing these concerns also demands greater transparency and accountability from AI companies. Large technology firms operate extensive cloud infrastructures that power modern AI applications, yet their environmental impact is often opaque. Organizations must take active steps to measure, report, and reduce their carbon footprint across the entire AI lifecycle, from hardware manufacturing to model training and inference. Voluntary self-regulation is an important first step, but policy interventions and industry-wide standards may be necessary to ensure long-term sustainability. Reported metrics such as energy consumption, carbon emissions, and efficiency benchmarks could serve as mechanisms to hold organizations accountable.

Furthermore, ethical AI development must encourage open discourse on environmental trade-offs. Researchers should be empowered to advocate for sustainability within their institutions and organizations, ensuring that environmental concerns are factored into AI development priorities. The broader AI community has already begun addressing these issues, as exemplified by the open letter advocating a pause on large-scale AI experiments, which highlights concerns about unchecked expansion. By fostering a culture of transparency and ethical responsibility, the AI industry can work toward aligning technological advancement with ecological sustainability.

AI has the potential to reshape industries and societies, but its long-term viability depends on how responsibly it is developed. Ethical AI development is not only about preventing harm to individuals and communities but also about ensuring that AI-driven innovation does not come at the cost of environmental degradation. As stewards of these powerful technologies, developers and organizations have a profound duty to integrate sustainability into AI’s future trajectory.

17.2.3 Case Study: DeepMind’s Energy Efficiency

Google’s data centers form the backbone of services such as Search, Gmail, and YouTube, handling billions of queries daily. These data centers operate at massive scales, consuming vast amounts of electricity, particularly for cooling infrastructure that ensures optimal server performance. Improving the energy efficiency of data centers has long been a priority, but conventional engineering approaches faced diminishing returns due to the complexity of the cooling systems and the highly dynamic nature of environmental conditions. To address these challenges, Google collaborated with DeepMind to develop a machine learning-driven optimization system that could automate and enhance energy management at scale.

Building on more than a decade of efforts to optimize data center design, energy-efficient hardware, and renewable energy integration, DeepMind’s AI approach targeted one of the most energy-intensive aspects of data centers: cooling systems. Traditional cooling relies on manually set heuristics that account for factors such as server heat output, external weather conditions, and architectural constraints. However, these systems exhibit nonlinear interactions, meaning that simple rule-based optimizations often fail to capture the full complexity of their operations. The result was suboptimal cooling efficiency, leading to unnecessary energy waste.

DeepMind’s team trained a neural network model using Google’s historical sensor data, which included real-time temperature readings, power consumption levels, cooling pump activity, and other operational parameters. The model learned the intricate relationships between these factors and could dynamically predict the most efficient cooling configurations. Unlike traditional approaches, which relied on human engineers periodically adjusting system settings, the AI model continuously adapted in real time to changing environmental and workload conditions.

The results were unprecedented efficiency gains. When deployed in live data center environments, DeepMind’s AI-driven cooling system reduced cooling energy consumption by 40%, leading to an overall 15% improvement in Power Usage Effectiveness (PUE)—a key metric for data center energy efficiency that measures the ratio of total energy consumption to the energy used purely for computing tasks (Barroso, Hölzle, and Ranganathan 2019). Notably, these improvements were achieved without any additional hardware modifications, demonstrating the potential of software-driven optimizations to significantly reduce AI’s carbon footprint.

Beyond a single data center, DeepMind’s AI model provided a generalizable framework that could be adapted to different facility designs and climate conditions, offering a scalable solution for optimizing power consumption across global data center networks. This case study exemplifies how AI can be leveraged not just as a consumer of computational resources but as a tool for sustainability, driving substantial efficiency improvements in the infrastructure that supports machine learning itself.

The integration of data-driven decision-making, real-time adaptation, and scalable AI models demonstrates the growing role of intelligent resource management in sustainable AI system design. This breakthrough exemplifies how machine learning can be applied to optimize the very infrastructure that powers it, ensuring a more energy-efficient future for large-scale AI deployments.

17.3 AI Carbon Footprint

The carbon footprint of artificial intelligence is a critical aspect of its overall environmental impact. As AI adoption continues to expand, so does its energy consumption and associated greenhouse gas emissions. Training and deploying AI models require vast computational resources, often powered by energy-intensive data centers that contribute significantly to global carbon emissions. However, the carbon footprint of AI extends beyond electricity usage, encompassing hardware manufacturing, data storage, and end-user interactions—all of which contribute to emissions across an AI system’s lifecycle.

Quantifying the carbon impact of AI is complex, as it depends on multiple factors, including the size of the model, the duration of training, the hardware used, and the energy sources powering data centers. Large-scale AI models, such as GPT-3, require thousands of megawatt-hours (MWh) of electricity, equivalent to the energy consumption of entire communities. The energy required for inference—the phase in which trained models generate outputs—is also substantial, particularly for widely deployed AI services such as real-time translation, image generation, and personalized recommendations. Unlike traditional software, which has a relatively static energy footprint, AI models consume energy continuously, leading to an ongoing sustainability challenge.

Beyond direct energy use, the carbon footprint of AI must also account for indirect emissions from hardware production and supply chains. Manufacturing AI accelerators such as GPUs, TPUs, and custom chips involves energy-intensive fabrication processes that rely on rare earth metals and complex supply chains. The full life cycle emissions of AI systems—encompassing data centers, hardware manufacturing, and global AI deployments—must be considered to develop more sustainable AI practices.

Understanding AI’s carbon footprint requires breaking down where emissions come from, how they are measured, and what strategies can be employed to mitigate them. We explore the following:

• Carbon emissions and energy consumption trends in AI—quantifying AI’s energy demand and providing real-world comparisons.
• Scopes of carbon emissions (Scope 1, 2, and 3)—differentiating between direct, indirect, and supply chain-related emissions.
• The energy cost of training vs. inference—analyzing how different phases of AI impact sustainability.

By dissecting these components, we can better assess the true environmental impact of AI systems and identify opportunities to reduce their footprint through more efficient design, energy-conscious deployment, and sustainable infrastructure choices.

17.3.1 Emissions & Consumption

Artificial intelligence systems require vast computational resources, making them one of the most energy-intensive workloads in modern computing. The energy consumed by AI systems extends beyond the training of large models to include ongoing inference workloads, data storage, and communication across distributed computing infrastructure. As AI adoption scales across industries, understanding its energy consumption patterns and carbon emissions is critical for designing more sustainable machine learning infrastructure.

Data centers play a central role in AI’s energy demands, consuming vast amounts of electricity to power compute servers, storage, and cooling systems. Without access to renewable energy, these facilities rely heavily on nonrenewable sources such as coal and natural gas, contributing significantly to global carbon emissions. Current estimates suggest that data centers produce up to 2% of total global CO₂ emissions—a figure that is closing in on the airline industry’s footprint (Liu et al. 2020). The energy burden of AI is expected to grow exponentially due to three key factors: increasing data center capacity, rising AI training workloads, and surging inference demands (Patterson, Gonzalez, Holzle, et al. 2022). Without intervention, these trends risk making AI’s environmental footprint unsustainably large (Thompson, Spanuth, and Matthews 2023).

Liu, Yanan, Xiaoxia Wei, Jinyu Xiao, Zhijie Liu, Yang Xu, and Yun Tian. 2020. “Energy Consumption and Emission Mitigation Prediction Based on Data Center Traffic and PUE for Global Data Centers.” Global Energy Interconnection 3 (3): 272–82. https://doi.org/10.1016/j.gloei.2020.07.008.

Thompson, Neil, Tobias Spanuth, and Hyrum Anderson Matthews. 2023. “The Computational Limits of Deep Learning and the Future of AI.” Communications of the ACM 66 (3): 48–57. https://doi.org/10.1145/3580309.

Energy Demands in Data Centers

AI workloads are among the most compute-intensive operations in modern data centers. Companies such as Meta operate hyperscale data centers spanning multiple football fields in size, housing hundreds of thousands of AI-optimized servers. The training of large language models (LLMs) such as GPT-4 required over 25,000 Nvidia A100 GPUs running continuously for 90 to 100 days (Choi and Yoon 2024), consuming thousands of megawatt-hours (MWh) of electricity. These facilities rely on high-performance AI accelerators like NVIDIA DGX H100 units, each of which can draw up to 10.2 kW at peak power (Choquette 2023).

AI’s rapid adoption is driving a significant increase in data center energy consumption. As shown in Figure 17.3, the energy demand of AI workloads is projected to substantially increase total data center energy use, especially after 2024. While efficiency gains have historically offset rising power needs, these gains are decelerating, amplifying AI’s environmental impact.

Figure 17.3: Projected Data Center and AI Power Demand with Power Efficiency Gains (2015–2030). Source: Masanet et al. (2020), Cisco, IEA, Goldman Sachs Global Investment Research.

Cooling is another major factor in AI’s energy footprint. Large-scale AI training and inference workloads generate massive amounts of heat, necessitating advanced cooling solutions to prevent hardware failures. Estimates indicate that 30-40% of a data center’s total electricity usage goes into cooling alone (Dayarathna, Wen, and Fan 2016). Companies have begun adopting alternative cooling methods to reduce this demand. For example, Microsoft’s data center in Ireland leverages a nearby fjord, using over half a million gallons of seawater daily to dissipate heat. However, as AI models scale in complexity, cooling demands continue to grow, making sustainable AI infrastructure design a pressing challenge.

AI vs. Other Industries

The environmental impact of AI workloads has emerged as a significant concern, with carbon emissions approaching levels comparable to established carbon-intensive sectors. Research demonstrates that training a single large AI model generates carbon emissions equivalent to multiple passenger vehicles over their complete lifecycle (Strubell, Ganesh, and McCallum 2019b). To contextualize AI’s environmental footprint, Figure 17.5 compares the carbon emissions of large-scale machine learning tasks to transcontinental flights, illustrating the substantial energy demands of training and inference workloads. It shows a comparison from lowest to highest carbon footprints, starting with a roundtrip flight between NY and SF, human life average per year, American life average per year, US car including fuel over a lifetime, and a Transformer model with neural architecture search, which has the highest footprint. These comparisons underscore the need for more sustainable AI practices to mitigate the industry’s carbon impact.

Figure 17.4: Carbon footprint of NLP model in lbs of carbon dioxide.

The training phase of large natural language processing models produces carbon dioxide emissions comparable to hundreds of transcontinental flights. When examining the broader industry impact, AI’s aggregate computational carbon footprint is approaching parity with the commercial aviation sector. Furthermore, as AI applications scale to serve billions of users globally, the cumulative emissions from continuous inference operations may ultimately exceed those generated during training.

Figure 17.5 provides a detailed analysis of carbon emissions across various large-scale machine learning tasks at Meta, illustrating the substantial environmental impact of different AI applications and architectures. This quantitative assessment of AI’s carbon footprint underscores the pressing need to develop more sustainable approaches to machine learning development and deployment. Understanding these environmental costs is crucial for implementing effective mitigation strategies and advancing the field responsibly.

Figure 17.5: Carbon footprint of large-scale ML tasks. Source: Wu et al. (2022).

17.3.2 Updated Analysis

Moreover, AI’s impact extends beyond energy consumption during operation. The full lifecycle emissions of AI include hardware manufacturing, supply chain emissions, and end-of-life disposal, making AI a significant contributor to environmental degradation. AI models not only require electricity to train and infer, but they also depend on a complex infrastructure of semiconductor fabrication, rare earth metal mining¹, and electronic waste disposal. The next section breaks down AI’s carbon emissions into Scope 1 (direct emissions), Scope 2 (indirect emissions from electricity), and Scope 3 (supply chain and lifecycle emissions) to provide a more detailed view of its environmental impact.

¹ The production of AI chips requires rare earth elements such as neodymium and dysprosium, the extraction of which has significant environmental consequences.

17.3.3 Carbon Emission Scopes

AI is expected to see an annual growth rate of 37.3% between 2023 and 2030. Yet, applying the same growth rate to operational computing could multiply annual AI energy needs up to 1,000 times by 2030. So, while model optimization tackles one facet, responsible innovation must also consider total lifecycle costs at global deployment scales that were unfathomable just years ago but now pose infrastructure and sustainability challenges ahead.

Scope 1

Scope 1 emissions refer to direct greenhouse gas emissions produced by AI data centers and computing facilities. These emissions result primarily from on-site power generation, including backup diesel generators used to ensure reliability in large cloud environments, as well as facility cooling systems. Although many AI data centers predominantly rely on grid electricity, those with their own power plants or fossil-fuel-dependent backup systems contribute significantly to direct emissions, especially in regions where renewable energy sources are less prevalent (Masanet et al. 2020a).

———. 2020a. “Recalibrating Global Data Center Energy-Use Estimates.” Science 367 (6481): 984–86. https://doi.org/10.1126/science.aba3758.

Scope 2

Scope 2 emissions encompass indirect emissions from electricity purchased to power AI infrastructure. The majority of AI’s operational energy consumption falls under Scope 2, as cloud providers and enterprise computing facilities require massive electrical inputs for GPUs, TPUs, and high-density servers. The carbon intensity associated with Scope 2 emissions varies geographically based on regional energy mixes. Regions dominated by coal and natural gas electricity generation create significantly higher AI-related emissions compared to regions utilizing renewable sources such as wind, hydro, or solar. This geographic variability motivates companies to strategically position data centers in areas with cleaner energy sources and adopt carbon-aware scheduling strategies to reduce emissions (Alvim et al. 2022).

Scope 3

Scope 3 emissions constitute the largest and most complex category, capturing indirect emissions across the entire AI supply chain and lifecycle. These emissions originate from manufacturing, transportation, and disposal of AI hardware, particularly semiconductors and memory modules. Semiconductor manufacturing is particularly energy-intensive, involving complex processes such as chemical etching, rare-earth metal extraction, and extreme ultraviolet (EUV) lithography, all of which produce substantial carbon outputs. Indeed, manufacturing a single high-performance AI accelerator can generate emissions equivalent to several years of operational energy use (Gupta, Kim, et al. 2022).

Gupta, Udit, Young Geun Kim, Sylvia Lee, Jordan Tse, Hsien-Hsin S Lee, Gu-Yeon Wei, David Brooks, and Carole-Jean Wu. 2022. “Chasing Carbon: The Elusive Environmental Footprint of Computing.” IEEE Micro 42 (6): 68–78. https://doi.org/10.1109/MM.2022.3186575.

Beyond manufacturing, Scope 3 emissions include the downstream impact of AI once deployed. AI services such as search engines, social media platforms, and cloud-based recommendation systems operate at enormous scale, requiring continuous inference across millions or even billions of user interactions. The cumulative electricity demand of inference workloads can ultimately surpass the energy used for training, further amplifying AI’s carbon impact. End-user devices, including smartphones, IoT devices, and edge computing platforms, also contribute to Scope 3 emissions, as their AI-enabled functionality depends on sustained computation. Companies such as Meta and Google report that Scope 3 emissions from AI-powered services make up the largest share of their total environmental footprint, due to the sheer scale at which AI operates.

These massive facilities provide the infrastructure for training complex neural networks on vast datasets. For instance, based on leaked information, OpenAI’s language model GPT-4 was trained on Azure data centers packing over 25,000 Nvidia A100 GPUs, used continuously for over 90 to 100 days.

The GHG Protocol framework, illustrated in Figure 17.6, provides a structured way to visualize the sources of AI-related carbon emissions. Scope 1 emissions arise from direct company operations, such as data center power generation and company-owned infrastructure. Scope 2 covers electricity purchased from the grid, the primary source of emissions for cloud computing workloads. Scope 3 extends beyond an organization’s direct control, including emissions from hardware manufacturing, transportation, and even the end-user energy consumption of AI-powered services. Understanding this breakdown allows for more targeted sustainability strategies, ensuring that efforts to reduce AI’s environmental impact are not solely focused on energy efficiency but also address the broader supply chain and lifecycle emissions that contribute significantly to the industry’s carbon footprint.

Figure 17.6: The GHG Protocol framework categorizes emissions into Scope 1, 2, and 3, helping organizations assess their direct and indirect carbon impact. Source: Circularise.

17.3.4 Training vs. Inference Impact

The energy consumption of AI systems is often closely associated with the training phase, where substantial computational resources are utilized to develop large-scale machine learning models. However, while training demands significant power, it represents a one-time cost per model version. In contrast, inference—the ongoing application of trained models to new data—happens continuously at a massive scale and often becomes the dominant contributor to energy consumption over time (Patterson et al. 2021). As AI-powered services, such as real-time translation, recommender systems, and generative AI applications expand globally, inference workloads increasingly drive AI’s overall carbon footprint.

Training Energy Demands

Training state-of-the-art AI models demands enormous computational resources. For example, models like GPT-4 were trained using over 25,000 Nvidia A100 GPUs operating continuously for approximately three months within cloud-based data centers (Choi and Yoon 2024). OpenAI’s dedicated supercomputer infrastructure, built specifically for large-scale AI training, contains 285,000 CPU cores, 10,000 GPUs, and network bandwidth exceeding 400 gigabits per second per server, illustrating the vast scale and associated energy consumption of AI training infrastructures (Patterson et al. 2021).

Choi, Sebin, and Sungmin Yoon. 2024. “GPT-Based Data-Driven Urban Building Energy Modeling (GPT-UBEM): Concept, Methodology, and Case Studies.” Energy and Buildings 325 (December): 115042. https://doi.org/10.1016/j.enbuild.2024.115042.

Choquette, Jack. 2023. “NVIDIA Hopper H100 GPU: Scaling Performance.” IEEE Micro 43 (3): 9–17. https://doi.org/10.1109/mm.2023.3256796.

Dayarathna, Miyuru, Yonggang Wen, and Rui Fan. 2016. “Data Center Energy Consumption Modeling: A Survey.” IEEE Communications Surveys &Amp; Tutorials 18 (1): 732–94. https://doi.org/10.1109/comst.2015.2481183.

High-performance AI accelerators, such as NVIDIA DGX H100 systems, are specifically designed for these training workloads. Each DGX H100 unit can draw up to 10.2 kW at peak load, with clusters often consisting of thousands of nodes running continuously (Choquette 2023). The intensive computational loads result in significant heat dissipation, necessitating substantial cooling infrastructure. Cooling alone can account for 30-40% of total data center energy consumption (Dayarathna, Wen, and Fan 2016).

While significant, these energy costs occur once per trained model. The primary sustainability challenge emerges during model deployment, where inference workloads continuously serve millions or billions of users.

Inference Energy Costs

Inference workloads execute every time an AI model responds to queries, classifies images, or makes predictions. Unlike training, inference scales dynamically and continuously across applications such as search engines, recommendation systems, and generative AI models. Although each individual inference request consumes far less energy compared to training, the cumulative energy usage from billions of daily AI interactions quickly surpasses training-related consumption (Patterson et al. 2021).

Patterson, David, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. “Carbon Emissions and Large Neural Network Training.” arXiv Preprint arXiv:2104.10350, April. http://arxiv.org/abs/2104.10350v3.

For example, AI-driven search engines handle billions of queries per day, recommendation systems provide personalized content continuously, and generative AI services such as ChatGPT or DALL-E have substantial per-query computational costs. The inference energy footprint is especially pronounced in transformer-based models due to high memory and computational bandwidth requirements.

As shown in Figure 17.7, the market for inference workloads in data centers is projected to grow significantly from $4-5 billion in 2017 to $9-10 billion by 2025, more than doubling in size. Similarly, edge inference workloads are expected to increase from less than $0.1 billion to $4-4.5 billion in the same period. This growth substantially outpaces the expansion of training workloads in both environments, highlighting how the economic footprint of inference is rapidly outgrowing that of training operations.

Figure 17.7: Market size for inference and training hardware. Source: McKinsey.

Unlike traditional software applications with fixed energy footprints, inference workloads dynamically scale with user demand. AI services like Alexa, Siri, and Google Assistant rely on continuous cloud-based inference, processing millions of voice queries per minute, necessitating uninterrupted operation of energy-intensive data center infrastructure.

Edge AI Impact

Inference does not always happen in large data centers—edge AI is emerging as a viable alternative to reduce cloud dependency. Instead of routing every AI request to centralized cloud servers, some AI models can be deployed directly on user devices or at edge computing nodes. This approach reduces data transmission energy costs and lowers the dependency on high-power cloud inference.

However, running inference at the edge does not eliminate energy concerns—especially when AI is deployed at scale. Autonomous vehicles, for instance, require millisecond-latency AI inference, meaning cloud processing is impractical. Instead, vehicles are now being equipped with onboard AI accelerators that function as “data centers on wheels (Sudhakar, Sze, and Karaman 2023). These embedded computing systems process real-time sensor data equivalent to small data centers, consuming significant power even without relying on cloud inference.

Sudhakar, Soumya, Vivienne Sze, and Sertac Karaman. 2023. “Data Centers on Wheels: Emissions from Computing Onboard Autonomous Vehicles.” IEEE Micro 43 (1): 29–39. https://doi.org/10.1109/mm.2022.3219803.

Similarly, consumer devices such as smartphones, wearables, and IoT sensors individually consume relatively little power but collectively contribute significantly to global energy use due to their sheer numbers. Therefore, the efficiency benefits of edge computing must be balanced against the extensive scale of device deployment.

17.4 Beyond Carbon

While reducing AI’s carbon emissions is critical, the environmental impact extends far beyond energy consumption. The manufacturing of AI hardware involves significant resource extraction, hazardous chemical usage, and water consumption that often receive less attention despite their ecological significance.

Modern semiconductor fabrication plants (fabs) that produce AI chips require millions of gallons of water daily and use over 250 hazardous substances in their processes (Mills and Le Hunte 1997). In regions already facing water stress, such as Taiwan, Arizona, and Singapore, this intensive water usage threatens local ecosystems and communities.

The industry also relies heavily on scarce materials like gallium, indium, arsenic, and helium, which are essential for AI accelerators and high-speed communication chips (Chen 2006; Davies 2011). These materials face both geopolitical supply risks and depletion concerns.

We will explore these critical but often overlooked aspects of AI’s environmental impact, including water consumption, hazardous waste production, rare material scarcity, and biodiversity disruption. Understanding these broader ecological impacts is essential for developing truly sustainable AI infrastructure.

17.4.1 Water Usage

Semiconductor fabrication is an exceptionally water-intensive process, requiring vast quantities of ultrapure water ² for cleaning, cooling, and chemical processing. The scale of water consumption in modern fabs is comparable to that of entire urban populations. For example, TSMC’s latest fab in Arizona is projected to consume 8.9 million gallons of water per day, accounting for nearly 3% of the city’s total water production. This demand places significant strain on local water resources, particularly in water-scarce regions such as Taiwan, Arizona, and Singapore, where semiconductor manufacturing is concentrated. Semiconductor companies have recognized this challenge and are actively investing in recycling technologies and more efficient water management practices. STMicroelectronics, for example, recycles and reuses approximately 41% of its water, significantly reducing its environmental footprint (see Figure 17.8 showing the typical semiconductor fab water cycle).

² Ultrapure water (UPW): Water that has been purified to stringent standards, typically containing less than 1 part per billion of impurities. UPW is essential for semiconductor fabrication, as even trace contaminants can impair chip performance and yield.

Figure 17.8: Typical water cycle in semiconductor manufacturing, illustrating water extraction, ultrapure water generation, manufacturing use, recycling processes, and wastewater treatment before discharge or reuse. Source: ST Sustainability Report.

The primary use of ultrapure water in semiconductor fabrication is for flushing contaminants from wafers at various production stages. Water also serves as a coolant and carrier fluid in thermal oxidation, chemical deposition, and planarization processes. A single 300mm silicon wafer requires over 8,300 liters of water, with more than two-thirds of this being ultrapure water (Cope 2009). During peak summer months, the cumulative daily water consumption of major fabs rivals that of cities with populations exceeding half a million people.

³ Saltwater Intrusion: The process by which seawater enters freshwater aquifers due to groundwater overuse, leading to water quality degradation.

The impact of this massive water usage extends beyond consumption. Excessive water withdrawal from local aquifers lowers groundwater levels, leading to issues such as land subsidence and saltwater intrusion³. In Hsinchu, Taiwan, one of the world’s largest semiconductor hubs, extensive water extraction by fabs has led to falling water tables and encroaching seawater contamination, affecting both agriculture and drinking water supplies.

Figure 17.9 contextualizes the daily water footprint of data centers compared to other industrial uses, illustrating the immense water demand of high-tech infrastructure.

Figure 17.9: Daily Water Footprint of Datacenters in comparison with other water uses. Source: Google’s Data Center Cooling

While some semiconductor manufacturers implement water recycling systems, the effectiveness of these measures varies. Intel reports that 97% of its direct water consumption is attributed to fabrication processes (Cooper et al. 2011), and while water reuse is increasing, the sheer scale of water withdrawals remains a critical sustainability challenge.

Cooper, Tom, Suzanne Fallender, Joyann Pafumi, Jon Dettling, Sebastien Humbert, and Lindsay Lessard. 2011. “A Semiconductor Company’s Examination of Its Water Footprint Approach.” In Proceedings of the 2011 IEEE International Symposium on Sustainable Systems and Technology, 1–6. IEEE; IEEE. https://doi.org/10.1109/issst.2011.5936865.

Beyond depletion, water discharge from semiconductor fabs introduces contamination risks if not properly managed. Wastewater from fabrication contains metals, acids, and chemical residues that must be thoroughly treated before release. Although modern fabs employ advanced purification systems, the extraction of contaminants still generates hazardous byproducts, which, if not carefully disposed of, pose risks to local ecosystems.

The growing demand for semiconductor manufacturing, driven by AI acceleration and computing infrastructure expansion, makes water management a crucial factor in sustainable AI development. Ensuring the long-term viability of semiconductor production requires not only reducing direct water consumption but also enhancing wastewater treatment and developing alternative cooling technologies that minimize reliance on fresh water sources.

17.4.2 Hazardous Chemicals

Semiconductor fabrication is heavily reliant on highly hazardous chemicals, which play an essential role in processes such as etching, doping, and wafer cleaning. The manufacturing of AI hardware, including GPUs, TPUs, and other specialized accelerators, requires the use of strong acids, volatile solvents, and toxic gases, all of which pose significant health and environmental risks if not properly managed. The scale of chemical usage in fabs is immense, with thousands of metric tons of hazardous substances consumed annually (Kim et al. 2018).

Kim, Sunju, Chungsik Yoon, Seunghon Ham, Jihoon Park, Ohun Kwon, Donguk Park, Sangjun Choi, Seungwon Kim, Kwonchul Ha, and Won Kim. 2018. “Chemical Use in the Semiconductor Manufacturing Industry.” International Journal of Occupational and Environmental Health 24 (3-4): 109–18. https://doi.org/10.1080/10773525.2018.1519957.

Among the most critical chemical categories used in fabrication are strong acids, which facilitate wafer etching and oxide removal. Hydrofluoric acid, sulfuric acid, nitric acid, and hydrochloric acid are commonly employed in the cleaning and patterning stages of chip production. While effective for these processes, these acids are highly corrosive and toxic, capable of causing severe chemical burns and respiratory damage if mishandled. Large semiconductor fabs require specialized containment, filtration, and neutralization systems to prevent accidental exposure and environmental contamination.

Solvents are another critical component in chip manufacturing, primarily used for dissolving photoresists and cleaning wafers. Key solvents include xylene, methanol, and methyl isobutyl ketone (MIBK), which, despite their utility, present air pollution and worker safety risks. These solvents are volatile organic compounds (VOCs)⁴ that can evaporate into the atmosphere, contributing to indoor and outdoor air pollution. If not properly contained, VOC exposure can result in neurological damage, respiratory issues, and long-term health effects for workers in semiconductor fabs.

⁴ Volatile organic compounds (VOCs): Organic chemicals that easily evaporate into the air, posing health risks and contributing to air pollution. VOCs are commonly used in semiconductor manufacturing for cleaning, etching, and photoresist removal.

Toxic gases are among the most dangerous substances used in AI chip manufacturing. Gases such as arsine (AsH₃), phosphine (PH₃), diborane (B₂H₆), and germane (GeH₄) are used in doping and chemical vapor deposition processes, essential for fine-tuning semiconductor properties. These gases are highly toxic and even fatal at low concentrations, requiring extensive handling precautions, gas scrubbers, and emergency response protocols. Any leaks or accidental releases in fabs can lead to severe health hazards for workers and surrounding communities.

While modern fabs employ strict safety controls, protective equipment, and chemical treatment systems, incidents still occur, leading to chemical spills, gas leaks, and contamination risks. The challenge of effectively managing hazardous chemicals is heightened by the ever-increasing complexity of AI accelerators, which require more advanced fabrication techniques and new chemical formulations.

Beyond direct safety concerns, the long-term environmental impact of hazardous chemical use remains a major sustainability issue. Semiconductor fabs generate large volumes of chemical waste, which, if improperly handled, can contaminate groundwater, soil, and local ecosystems. Regulations in many countries require fabs to neutralize and treat waste before disposal, but compliance and enforcement vary globally, leading to differing levels of environmental protection.

To mitigate these risks, fabs must continue advancing green chemistry initiatives, exploring alternative etchants, solvents, and gas formulations that reduce toxicity while maintaining fabrication efficiency. Additionally, process optimizations that minimize chemical waste, improve containment, and enhance recycling efforts will be essential to reducing the environmental footprint of AI hardware production.

17.4.3 Resource Depletion

While silicon is abundant and readily available, the fabrication of AI accelerators, GPUs, and specialized AI chips depends on scarce and geopolitically sensitive materials that are far more difficult to source. AI hardware manufacturing requires a range of rare metals, noble gases, and semiconductor compounds, many of which face supply constraints, geopolitical risks, and environmental extraction costs. As AI models become larger and more computationally intensive, the demand for these materials continues to rise, raising concerns about long-term availability and sustainability.

Although silicon forms the foundation of semiconductor devices, high-performance AI chips depend on rare elements such as gallium, indium, and arsenic, which are essential for high-speed, low-power electronic components (Chen 2006). Gallium and indium, for example, are widely used in compound semiconductors, particularly for 5G communications, optoelectronics, and AI accelerators. The United States Geological Survey (USGS) has classified indium as a critical material, with global supplies expected to last fewer than 15 years at the current rate of consumption (Davies 2011).

Chen, H.-W. 2006. “Gallium, Indium, and Arsenic Pollution of Groundwater from a Semiconductor Manufacturing Area of Taiwan.” Bulletin of Environmental Contamination and Toxicology 77 (2): 289–96. https://doi.org/10.1007/s00128-006-1062-3.

⁵ Extreme ultraviolet (EUV) lithography: A cutting-edge semiconductor manufacturing technique that uses EUV light to etch nanoscale features on silicon wafers. EUV lithography is essential for producing advanced AI chips with smaller transistors and higher performance.

Another major concern is helium, a noble gas critical for semiconductor cooling, plasma etching, and EUV lithography⁵ used in next-generation chip production. Helium is unique in that once released into the atmosphere, it escapes Earth’s gravity and is lost forever, making it a non-renewable resource (Davies 2011). The semiconductor industry is one of the largest consumers of helium, and supply shortages have already led to price spikes and disruptions in fabrication processes. As AI hardware manufacturing scales, the demand for helium will continue to grow, necessitating more sustainable extraction and recycling practices.

Beyond raw material availability, the geopolitical control of rare earth elements poses additional challenges. China currently dominates over 90% of the world’s rare earth element (REE) refining capacity, including materials essential for AI chips, such as neodymium (for high-performance magnets in AI accelerators) and yttrium (for high-temperature superconductors) (Jha 2014). This concentration of supply creates supply chain vulnerabilities, as trade restrictions or geopolitical tensions could severely impact AI hardware production.

Jha, A. R. 2014. Rare Earth Materials: Properties and Applications. CRC Press. https://doi.org/10.1201/b17045.

Table 17.1 highlights the key materials essential for AI semiconductor manufacturing, their applications, and supply concerns.

Table 17.1: Rare materials that are widely used in the semiconductor industry that are facing resource depletion.

Material	Application in AI Semiconductor Manufacturing	Supply Concerns
Silicon (Si)	Primary substrate for chips, wafers, transistors	Processing constraints; geopolitical risks
Gallium (Ga)	GaN-based power amplifiers, high-frequency components	Limited availability; byproduct of aluminum and zinc production
Germanium (Ge)	High-speed transistors, photodetectors, optical interconnects	Scarcity; geographically concentrated
Indium (In)	Indium Tin Oxide (ITO), optoelectronics	Limited reserves; recycling dependency
Tantalum (Ta)	Capacitors, stable integrated components	Conflict mineral; vulnerable supply chains
Rare Earth Elements (REEs)	Magnets, sensors, high-performance electronics	High geopolitical risks; environmental extraction concerns
Cobalt (Co)	Batteries for edge computing devices	Human rights issues; geographical concentration (Congo)
Tungsten (W)	Interconnects, barriers, heat sinks	Limited production sites; geopolitical concerns
Copper (Cu)	Interconnects, barriers, heat sinks	Limited high-purity sources; geopolitical concerns
Helium (He)	Semiconductor cooling, plasma etching, EUV lithography	Non-renewable; irretrievable atmospheric loss; limited extraction capacity
Indium (In)	ITO layers, optoelectronic components	Limited global reserves; geopolitical concentration
Cobalt (Co)	Batteries for edge computing devices	Geographical concentration; human rights concerns
Tungsten (W)	Interconnects, heat sinks	Limited production sites; geopolitical concerns
Copper (Cu)	Conductive pathways, wiring	Geopolitical dependencies; limited recycling capacity

The rapid growth of AI and semiconductor demand has accelerated the depletion of these critical resources, creating an urgent need for material recycling, substitution strategies, and more sustainable extraction methods. Some efforts are underway to explore alternative semiconductor materials that reduce dependency on rare elements, but these solutions require significant advancement before they become viable alternatives at scale.

17.4.4 Waste Generation

Semiconductor fabrication produces significant volumes of hazardous waste, including gaseous emissions, VOCs, chemical-laden wastewater, and solid toxic byproducts. The production of AI accelerators, GPUs, and other high-performance chips involves multiple stages of chemical processing, etching, and cleaning, each generating waste materials that must be carefully treated to prevent environmental contamination.

Fabs release gaseous waste from various processing steps, particularly chemical vapor deposition (CVD), plasma etching, and ion implantation. This includes toxic and corrosive gases such as arsine (AsH₃), phosphine (PH₃), and germane (GeH₄), which require advanced scrubber systems to neutralize before release into the atmosphere. If not properly filtered, these gases pose severe health hazards and contribute to air pollution and acid rain formation (Grossman 2007).

Grossman, Elizabeth. 2007. High Tech Trash: Digital Devices, Hidden Toxics, and Human Health. Island press.

VOCs are another major waste category, emitted from photoresist processing, cleaning solvents, and lithographic coatings. Chemicals such as xylene, acetone, and methanol readily evaporate into the air, where they contribute to ground-level ozone formation and indoor air quality hazards for fab workers. In regions where semiconductor production is concentrated, such as Taiwan and South Korea, regulators have imposed strict VOC emission controls to mitigate their environmental impact.

Semiconductor fabs also generate large volumes of spent acids and metal-laden wastewater, requiring extensive treatment before discharge. Strong acids such as sulfuric acid, hydrofluoric acid, and nitric acid are used to etch silicon wafers, removing excess materials during fabrication. When these acids become contaminated with heavy metals, fluorides, and chemical residues, they must undergo neutralization and filtration before disposal. Improper handling of wastewater has led to groundwater contamination incidents, highlighting the importance of robust waste management systems (Prakash et al. 2023).

The solid waste produced in AI hardware manufacturing includes sludge, filter cakes, and chemical residues collected from fab exhaust and wastewater treatment systems. These byproducts often contain concentrated heavy metals, rare earth elements, and semiconductor process chemicals, making them hazardous for conventional landfill disposal. In some cases, fabs incinerate toxic waste, generating additional environmental concerns related to airborne pollutants and toxic ash disposal.

Beyond the waste generated during manufacturing, the end-of-life disposal of AI hardware presents another sustainability challenge. AI accelerators, GPUs, and server hardware have short refresh cycles, with data center equipment typically replaced every 3-5 years. This results in millions of tons of e-waste annually, much of which contains toxic heavy metals such as lead, cadmium, and mercury. Despite growing efforts to improve electronics recycling, current systems capture only 17.4% of global e-waste, leaving the majority to be discarded in landfills or improperly processed (Singh and Ogunseitan 2022).

Addressing the hazardous waste impact of AI requires advancements in both semiconductor manufacturing and e-waste recycling. Companies are exploring closed-loop recycling for rare metals, improved chemical treatment processes, and alternative materials with lower toxicity. However, as AI models continue to drive demand for higher-performance chips and larger-scale computing infrastructure, the industry’s ability to manage its waste footprint will be a key factor in achieving sustainable AI development.

17.4.5 Biodiversity Impact

The environmental footprint of AI hardware extends beyond carbon emissions, resource depletion, and hazardous waste. The construction and operation of semiconductor fabrication facilities (fabs), data centers, and supporting infrastructure directly impact natural ecosystems, contributing to habitat destruction, water stress, and pollution. These environmental changes have far-reaching consequences for wildlife, plant ecosystems, and aquatic biodiversity, highlighting the need for sustainable AI development that considers broader ecological effects.

Semiconductor fabs and data centers require large tracts of land, often leading to deforestation and destruction of natural habitats. These facilities are typically built in industrial parks or near urban centers, but as demand for AI hardware increases, fabs are expanding into previously undeveloped regions, encroaching on forests, wetlands, and agricultural land.

The physical expansion of AI infrastructure disrupts wildlife migration patterns, as roads, pipelines, transmission towers, and supply chains fragment natural landscapes. Species that rely on large, connected ecosystems for survival—such as migratory birds, large mammals, and pollinators—face increased barriers to movement, reducing genetic diversity and population stability. In regions with dense semiconductor manufacturing, such as Taiwan and South Korea, habitat loss has already been linked to declining biodiversity in affected areas (Hsu et al. 2016).

Hsu, Liang-Ching, Ching-Yi Huang, Yen-Hsun Chuang, Ho-Wen Chen, Ya-Ting Chan, Heng Yi Teah, Tsan-Yao Chen, Chiung-Fen Chang, Yu-Ting Liu, and Yu-Min Tzou. 2016. “Accumulation of Heavy Metals and Trace Elements in Fluvial Sediments Received Effluents from Traditional and Semiconductor Industries.” Scientific Reports 6 (1): 34250. https://doi.org/10.1038/srep34250.

The massive water consumption of semiconductor fabs poses serious risks to aquatic ecosystems, particularly in water-stressed regions. Excessive groundwater extraction for AI chip production can lower water tables, affecting local rivers, lakes, and wetlands. In Hsinchu, Taiwan, where fabs draw millions of gallons of water daily, seawater intrusion has been reported in local aquifers, altering water chemistry and making it unsuitable for native fish species and vegetation.

Beyond depletion, wastewater discharge from fabs introduces chemical contaminants into natural water systems. While many facilities implement advanced filtration and recycling, even trace amounts of heavy metals, fluorides, and solvents can accumulate in water bodies, bioaccumulating in fish and disrupting aquatic ecosystems. Additionally, thermal pollution from data centers, which release heated water back into lakes and rivers, can raise temperatures beyond tolerable levels for native species, affecting oxygen levels and reproductive cycles (LeRoy Poff, Brinson, and Day 2002).

LeRoy Poff, N, MM Brinson, and JW Day. 2002. “Aquatic Ecosystems & Global Climate Change.” Pew Center on Global Climate Change.

Semiconductor fabs emit a variety of airborne pollutants, including VOCs, acid mists, and metal particulates, which can travel significant distances before settling in the environment. These emissions contribute to air pollution and acid deposition, which damage plant life, soil quality, and nearby agricultural systems.

Airborne chemical deposition has been linked to tree decline, reduced crop yields, and soil acidification, particularly near industrial semiconductor hubs. In areas with high VOC emissions, plant growth can be stunted by prolonged exposure, affecting ecosystem resilience and food chains. Additionally, accidental chemical spills or gas leaks from fabs pose severe risks to both local wildlife and human populations, requiring strict regulatory enforcement to minimize long-term ecological damage (Wald and Jones 1987).

Wald, Peter H., and Jeffrey R. Jones. 1987. “Semiconductor Manufacturing: An Introduction to Processes and Hazards.” American Journal of Industrial Medicine 11 (2): 203–21. https://doi.org/10.1002/ajim.4700110209.

The environmental consequences of AI hardware manufacturing demonstrate the urgent need for sustainable semiconductor production, including reduced land use, improved water recycling, and stricter emissions controls. Without intervention, the accelerating demand for AI chips could further strain global biodiversity, emphasizing the importance of balancing technological progress with ecological responsibility.

17.5 Semiconductor Life Cycle

The environmental footprint of AI systems extends beyond energy consumption during model training and inference. A comprehensive assessment of AI’s sustainability must consider the entire lifecycle—from the extraction of raw materials used in hardware manufacturing to the eventual disposal of obsolete computing infrastructure. Life Cycle Analysis (LCA) provides a systematic approach to quantifying the cumulative environmental impact of AI across its four key phases: design, manufacture, use, and disposal.

By applying LCA to AI systems, researchers and policymakers can pinpoint critical intervention points to reduce emissions, improve resource efficiency, and implement sustainable practices. This approach provides a holistic understanding of AI’s ecological costs, extending sustainability considerations beyond operational power consumption to include hardware supply chains and electronic waste management.

Figure 17.10 illustrates the four primary stages of an AI system’s lifecycle, each contributing to its total environmental footprint.

Figure 17.10: AI System Life Cycle Analysis is divided into four key phases: Design, Manufacture, Use, Disposal.

The following sections will analyze each lifecycle phase in detail, exploring its specific environmental impacts and sustainability challenges.

17.5.1 Design Phase

The design phase of an AI system encompasses the research, development, and optimization of machine learning models before deployment. This stage involves iterating on model architectures, adjusting hyperparameters, and running training experiments to improve performance. These processes are computationally intensive, requiring extensive use of hardware resources and energy. The environmental cost of AI model design is often underestimated, but repeated training runs, algorithm refinements, and exploratory experimentation contribute significantly to the overall sustainability impact of AI systems.

Developing an AI model requires running multiple experiments to determine the most effective architecture. Neural architecture search (NAS), for instance, automates the process of selecting the best model structure by evaluating hundreds or even thousands of configurations, each requiring a separate training cycle. Similarly, hyperparameter tuning involves modifying parameters such as learning rates, batch sizes, and optimization strategies to enhance model performance, often through exhaustive search techniques. Pre-training and fine-tuning further add to the computational demands, as models undergo multiple training iterations on different datasets before deployment. The iterative nature of this process results in high energy consumption, with studies indicating that hyperparameter tuning alone can account for up to 80% of training-related emissions (Strubell, Ganesh, and McCallum 2019b).

The scale of energy consumption in the design phase becomes evident when examining large AI models. OpenAI’s GPT-3, for example, required an estimated 1,300 megawatt-hours (MWh) of electricity for training, a figure comparable to the energy consumption of 1,450 U.S. homes over an entire month (Maslej et al. 2023). However, this estimate only reflects the final training run and does not account for the extensive trial-and-error processes that preceded model selection. In deep reinforcement learning applications, such as DeepMind’s AlphaZero, models undergo repeated training cycles to improve decision-making policies, further amplifying energy demands.

The carbon footprint of AI model design varies significantly depending on the computational resources required and the energy sources powering the data centers where training occurs. A widely cited study found that training a single large-scale natural language processing (NLP) model could produce emissions equivalent to the lifetime carbon footprint of five cars (Strubell, Ganesh, and McCallum 2019b). The impact is even more pronounced when training is conducted in data centers reliant on fossil fuels. For instance, models trained in coal-powered facilities in Virginia (USA) generate far higher emissions than those trained in regions powered by hydroelectric or nuclear energy. Hardware selection also plays a crucial role; training on energy-efficient tensor processing units (TPUs) can significantly reduce emissions compared to using traditional graphics processing units (GPUs).

Table 17.2 summarizes the estimated carbon emissions associated with training various AI models, illustrating the correlation between model complexity and environmental impact.

Table 17.2: Estimated carbon emissions associated with training various AI models, based on computational requirements and energy consumption. Source: Adapted from (Patterson, Gonzalez, Holzle, et al. 2022; Strubell, Ganesh, and McCallum 2019b).

AI Model	Training Compute (FLOPs)	Estimated \(\textrm{CO}_2\) Emissions (kg)	Equivalent Car Miles Driven
GPT-3	\(3.1 \times 10^{23}\)	502,000 kg	1.2 million miles
T5-11B	\(2.3 \times 10^{22}\)	85,000 kg	210,000 miles
BERT (Base)	\(3.3 \times 10^{18}\)	650 kg	1,500 miles
ResNet-50	\(2.0 \times 10^{17}\)	35 kg	80 miles

Addressing the sustainability challenges of the design phase requires innovations in training efficiency and computational resource management. Researchers have explored techniques such as sparse training, low-precision arithmetic, and weight-sharing methods to reduce the number of required computations without sacrificing model performance. The use of pre-trained models has also gained traction as a means of minimizing resource consumption. Instead of training models from scratch, researchers can fine-tune smaller versions of pre-trained networks, leveraging existing knowledge to achieve similar results with lower computational costs.

Optimizing model search algorithms further contributes to sustainability. Traditional neural architecture search methods require evaluating a large number of candidate architectures, but recent advances in energy-aware NAS approaches prioritize efficiency by reducing the number of training iterations needed to identify optimal configurations. Companies have also begun implementing carbon-aware computing strategies by scheduling training jobs during periods of lower grid carbon intensity or shifting workloads to data centers with cleaner energy sources (Gupta, Elgamal, et al. 2022).

The design phase sets the foundation for the entire AI lifecycle, influencing energy demands in both the training and inference stages. As AI models grow in complexity, their development processes must be reevaluated to ensure that sustainability considerations are integrated at every stage. The decisions made during model design not only determine computational efficiency but also shape the long-term environmental footprint of AI technologies.

17.5.2 Manufacturing Phase

The manufacturing phase of AI systems is one of the most resource-intensive aspects of their lifecycle, involving the fabrication of specialized semiconductor hardware such as GPUs, TPUs, FPGAs, and other AI accelerators. The production of these chips requires large-scale industrial processes, including raw material extraction, wafer fabrication, lithography, doping, and packaging—all of which contribute significantly to environmental impact (@ Bhamra et al. 2024). This phase not only involves high energy consumption but also generates hazardous waste, relies on scarce materials, and has long-term consequences for resource depletion.

Fabrication Materials

The foundation of AI hardware lies in semiconductors, primarily silicon-based integrated circuits that power AI accelerators. However, modern AI chips rely on more than just silicon; they require specialty materials such as gallium, indium, arsenic, and helium, each of which carries unique environmental extraction costs. These materials are often classified as critical elements due to their scarcity, geopolitical sensitivity, and high energy costs associated with mining and refining (Bhamra et al. 2024).

Bhamra, Ran, Adrian Small, Christian Hicks, and Olimpia Pilch. 2024. “Impact Pathways: Geopolitics, Risk and Ethics in Critical Minerals Supply Chains.” International Journal of Operations &Amp; Production Management, September. https://doi.org/10.1108/ijopm-03-2024-0228.

Cope, Gord. 2009. “Pure Water, Semiconductors and the Recession.” Global Water Intelligence 10 (10).

Silicon itself is abundant, but refining it into high-purity wafers requires extensive energy-intensive processes. The production of a single 300mm silicon wafer requires over 8,300 liters of water, along with strong acids such as hydrofluoric acid, sulfuric acid, and nitric acid used for etching and cleaning (Cope 2009). The demand for ultra-pure water in semiconductor fabrication places a significant burden on local water supplies, with leading fabs consuming millions of gallons per day.

Beyond silicon, gallium and indium are essential for high-performance compound semiconductors, such as those used in high-speed AI accelerators and 5G communications. The U.S. Geological Survey has classified indium as a critically endangered material, with global supplies estimated to last fewer than 15 years at current consumption rates (Davies 2011). Meanwhile, helium, a crucial cooling agent in chip production, is a non-renewable resource that, once released, escapes Earth’s gravity, making it permanently unrecoverable. The continued expansion of AI hardware manufacturing is accelerating the depletion of these critical elements, raising concerns about long-term sustainability.

Davies, Martin. 2011. “Endangered Elements: Critical Thinking.” In Study Skills for International Postgraduates, 111–30. Macmillan Education UK. https://doi.org/10.1007/978-0-230-34553-9\_8.

The environmental burden of semiconductor fabrication is further amplified by the use of EUV lithography, a process required for manufacturing sub-5nm chips. EUV systems consume massive amounts of energy, requiring high-powered lasers and complex optics. The International Semiconductor Roadmap estimates that each EUV tool consumes approximately one megawatt (MW) of electricity, significantly increasing the carbon footprint of cutting-edge chip production.

Manufacturing Energy Consumption

The energy required to manufacture AI hardware is substantial, with the total energy cost per chip often exceeding its entire operational lifetime energy use. The manufacturing of a single AI accelerator can emit more carbon than years of continuous use in a data center, making fabrication a key hotspot in AI’s environmental impact.

Hazardous Waste and Water Usage in Fabs

Semiconductor fabrication also generates large volumes of hazardous waste, including gaseous emissions, VOCs, chemical wastewater, and solid byproducts. The acids and solvents used in chip production produce toxic waste streams that require specialized handling to prevent contamination of surrounding ecosystems. Despite advancements in wastewater treatment, trace amounts of metals and chemical residues can still be released into rivers and lakes, affecting aquatic biodiversity and human health (Prakash et al. 2023).

Prakash, Shvetank, Tim Callahan, Joseph Bushagour, Colby Banbury, Alan V. Green, Pete Warden, Tim Ansell, and Vijay Janapa Reddi. 2023. “CFU Playground: Full-Stack Open-Source Framework for Tiny Machine Learning (TinyML) Acceleration on FPGAs.” In 2023 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), abs/2201.01863:157–67. IEEE. https://doi.org/10.1109/ispass57527.2023.00024.

⁶ Taiwan Semiconductor Manufacturing Company (TSMC) is one of the world’s largest semiconductor fabs, consuming millions of gallons of water daily in chip production, raising concerns about water scarcity.

The demand for water in semiconductor fabs has also raised concerns about regional water stress. The TSMC⁶ fab in Arizona is projected to consume 8.9 million gallons per day, a figure that accounts for nearly 3% of the city’s water supply. While some fabs have begun investing in water recycling systems, these efforts remain insufficient to offset the growing demand.

Sustainable Initiatives

Recognizing the sustainability challenges of semiconductor manufacturing, industry leaders have started implementing initiatives to reduce energy consumption, waste generation, and emissions. Companies like Intel, TSMC, and Samsung have pledged to transition towards carbon-neutral semiconductor fabrication through several key approaches. Many fabs are incorporating renewable energy sources, with facilities in Taiwan and Europe increasingly powered by hydroelectric and wind energy. Water conservation efforts have expanded through closed-loop recycling systems that reduce dependence on local water supplies. Manufacturing processes are being redesigned with eco-friendly etching and lithography techniques that minimize hazardous waste generation. Additionally, companies are developing energy-efficient chip architectures, such as low-power AI accelerators optimized for performance per watt, to reduce the environmental impact of both manufacturing and operation. Despite these efforts, the overall environmental footprint of AI chip manufacturing continues to grow as demand for AI accelerators escalates. Without significant improvements in material efficiency, recycling, and fabrication techniques, the manufacturing phase will remain a major contributor to AI’s sustainability challenges.

The manufacturing phase of AI hardware represents one of the most resource-intensive and environmentally impactful aspects of AI’s lifecycle. The extraction of critical materials, high-energy fabrication processes, and hazardous waste generation all contribute to AI’s growing carbon footprint. While industry efforts toward sustainable semiconductor manufacturing are gaining momentum, scaling these initiatives to meet rising AI demand remains a significant challenge.

Addressing the sustainability of AI hardware will require a combination of material innovation, supply chain transparency, and greater investment in circular economy models that emphasize chip recycling and reuse. As AI systems continue to advance, their long-term viability will depend not only on computational efficiency but also on reducing the environmental burden of their underlying hardware infrastructure.

17.5.3 Use Phase

The use phase of AI systems represents one of the most energy-intensive stages in their lifecycle, encompassing both training and inference workloads. As AI adoption grows across industries, the computational requirements for developing and deploying models continue to increase, leading to greater energy consumption and carbon emissions. The operational costs of AI systems extend beyond the direct electricity used in processing; they also include the power demands of data centers, cooling infrastructure, and networking equipment that support large-scale AI workloads. Understanding the sustainability challenges of this phase is critical for mitigating AI’s long-term environmental impact.

AI model training is among the most computationally expensive activities in the use phase. Training large-scale models involves running billions or even trillions of mathematical operations across specialized hardware, such as GPUs and TPUs, for extended periods. The energy consumption of training has risen sharply in recent years as AI models have grown in complexity. OpenAI’s GPT-3, for example, required approximately 1,300 megawatt-hours (MWh) of electricity, an amount equivalent to powering 1,450 U.S. homes for a month (Maslej et al. 2023). The carbon footprint of such training runs depends largely on the energy mix of the data center where they are performed. A model trained in a region relying primarily on fossil fuels, such as coal-powered data centers in Virginia, generates significantly higher emissions than one trained in a facility powered by hydroelectric or nuclear energy.

Maslej, Nestor, Loredana Fattorini, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, et al. 2023. “Artificial Intelligence Index Report 2023.” ArXiv Preprint abs/2310.03715 (October). http://arxiv.org/abs/2310.03715v1.

Patterson, David, Joseph Gonzalez, Urs Holzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. 2022. “The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink.” Computer 55 (7): 18–28. https://doi.org/10.1109/mc.2022.3148714.

Beyond training, the energy demands of AI do not end once a model is developed. The inference phase, where a trained model is used to generate predictions, is responsible for an increasingly large share of AI’s operational carbon footprint. In real-world applications, inference workloads run continuously, handling billions of requests daily across services such as search engines, recommendation systems, language models, and autonomous systems. The cumulative energy impact of inference is substantial, especially in large-scale deployments. While a single training run for a model like GPT-3 is energy-intensive, inference workloads running across millions of users can consume even more power over time. Studies have shown that inference now accounts for more than 60% of total AI-related energy consumption, exceeding the carbon footprint of training in many cases (Patterson, Gonzalez, Holzle, et al. 2022).

Data centers play a central role in enabling AI, housing the computational infrastructure required for training and inference. These facilities rely on thousands of high-performance servers, each drawing significant power to process AI workloads. The power usage effectiveness of a data center, which measures the efficiency of its energy use, directly influences AI’s carbon footprint. Many modern data centers operate with PUE values between 1.1 and 1.5, meaning that for every unit of power used for computation, an additional 10% to 50% is consumed for cooling, power conversion, and infrastructure overhead (Barroso, Hölzle, and Ranganathan 2019). Cooling systems, in particular, are a major contributor to data center energy consumption, as AI accelerators generate substantial heat during operation.

Barroso, Luiz André, Urs Hölzle, and Parthasarathy Ranganathan. 2019. The Datacenter as a Computer: Designing Warehouse-Scale Machines. Springer International Publishing. https://doi.org/10.1007/978-3-031-01761-2.

Gupta, Udit, Mariam Elgamal, Gage Hills, Gu-Yeon Wei, Hsien-Hsin S. Lee, David Brooks, and Carole-Jean Wu. 2022. “ACT: Designing Sustainable Computer Systems with an Architectural Carbon Modeling Tool.” In Proceedings of the 49th Annual International Symposium on Computer Architecture, 784–99. ACM. https://doi.org/10.1145/3470496.3527408.

The geographic location of data centers has a direct impact on their sustainability. Facilities situated in regions with renewable energy availability can significantly reduce emissions compared to those reliant on fossil fuel-based grids. Companies such as Google and Microsoft have invested in carbon-aware computing strategies, scheduling AI workloads during periods of high renewable energy production to minimize their carbon impact (Gupta, Elgamal, et al. 2022). Google’s DeepMind, for instance, developed an AI-powered cooling optimization system that reduced data center cooling energy consumption by 40%, lowering the overall carbon footprint of AI infrastructure.

The increasing energy demands of AI raise concerns about grid capacity and sustainability trade-offs. AI workloads often compete with other high-energy sectors, such as manufacturing and transportation, for limited electricity supply. In some regions, the rise of AI-driven data centers has led to increased stress on power grids, necessitating new infrastructure investments. The so-called “duck curve” problem, where renewable energy generation fluctuates throughout the day, poses additional challenges for balancing AI’s energy demands with grid availability. The shift toward distributed AI computing and edge processing is emerging as a potential solution to reduce reliance on centralized data centers, shifting some computational tasks closer to end users.

Mitigating the environmental impact of AI’s use phase requires a combination of hardware, software, and infrastructure-level optimizations. Advances in energy-efficient chip architectures, such as low-power AI accelerators and specialized inference hardware, have shown promise in reducing per-query energy consumption. AI models themselves are being optimized for efficiency through techniques such as quantization, pruning, and distillation, which allow for smaller, faster models that maintain high accuracy while requiring fewer computational resources. Meanwhile, ongoing improvements in cooling efficiency, renewable energy integration, and data center operations are essential for ensuring that AI’s growing footprint remains sustainable in the long term.

As AI adoption continues to expand, energy efficiency must become a central consideration in model deployment strategies. The use phase will remain a dominant contributor to AI’s environmental footprint, and without significant intervention, the sector’s electricity consumption could grow exponentially. Sustainable AI development requires a coordinated effort across industry, academia, and policymakers to promote responsible AI deployment while ensuring that technological advancements do not come at the expense of long-term environmental sustainability.

17.5.4 Disposal Phase

The disposal phase of AI systems is often overlooked in discussions of sustainability, yet it presents significant environmental challenges. The rapid advancement of AI hardware has led to shorter hardware lifespans, contributing to growing electronic waste (e-waste) and resource depletion. As AI accelerators, GPUs, and high-performance processors become obsolete within a few years, managing their disposal has become a pressing sustainability concern. Unlike traditional computing devices, AI hardware contains complex materials, rare earth elements, and hazardous substances that complicate recycling and waste management efforts. Without effective strategies for repurposing, recycling, or safely disposing of AI hardware, the environmental burden of AI infrastructure will continue to escalate.

The lifespan of AI hardware is relatively short, particularly in data centers where performance efficiency dictates frequent upgrades. On average, GPUs, TPUs, and AI accelerators are replaced every three to five years, as newer, more powerful models enter the market. This rapid turnover results in a constant cycle of hardware disposal, with large-scale AI deployments generating substantial e-waste. Unlike consumer electronics, which may have secondary markets for resale or reuse, AI accelerators often become unviable for commercial use once they are no longer state-of-the-art. The push for ever-faster and more efficient AI models accelerates this cycle, leading to an increasing volume of discarded high-performance computing hardware.

One of the primary environmental concerns with AI hardware disposal is the presence of hazardous materials. AI accelerators contain heavy metals such as lead, cadmium, and mercury, as well as toxic chemical compounds used in semiconductor fabrication. If not properly handled, these materials can leach into soil and water sources, causing long-term environmental and health hazards. The burning of e-waste releases toxic fumes, contributing to air pollution and exposing workers in informal recycling operations to harmful substances. Studies estimate that only 17.4% of global e-waste is properly collected and recycled, leaving the majority to end up in landfills or informal waste processing sites with inadequate environmental protections (Singh and Ogunseitan 2022).

Singh, Narendra, and Oladele A. Ogunseitan. 2022. “Disentangling the Worldwide Web of e-Waste and Climate Change Co-Benefits.” Circular Economy 1 (2): 100011. https://doi.org/10.1016/j.cec.2022.100011.

The complex composition of AI hardware presents significant challenges for recycling. Unlike traditional computing components, which are relatively straightforward to dismantle, AI accelerators incorporate specialized multi-layered circuits, exotic metal alloys, and tightly integrated memory architectures that make material recovery difficult. The disassembly and separation of valuable elements such as gold, palladium, and rare earth metals require advanced recycling technologies that are not widely available. The presence of mixed materials further complicates the process, as some components are chemically bonded or embedded in ways that make extraction inefficient.

Despite these challenges, efforts are being made to develop sustainable disposal solutions for AI hardware. Some manufacturers have begun designing AI accelerators with modular architectures, allowing for easier component replacement and extending the usable lifespan of devices. Research is also underway to improve material recovery processes, making it possible to extract and reuse critical elements such as gallium, indium, and tungsten from discarded chips. Emerging techniques such as hydrometallurgical and biometallurgical processing show promise in extracting rare metals with lower environmental impact compared to traditional smelting and refining methods.

The circular economy model offers a promising approach to mitigating the e-waste crisis associated with AI hardware. Instead of following a linear “use and discard” model, circular economy principles emphasize reuse, refurbishment, and recycling to extend the lifecycle of computing devices. Companies such as Google and Microsoft have launched initiatives to repurpose decommissioned AI hardware for secondary applications, such as running lower-priority machine learning tasks or redistributing functional components to research institutions. These efforts help reduce the overall demand for new semiconductor production while minimizing waste generation.

In addition to corporate sustainability initiatives, policy interventions and regulatory frameworks are critical in addressing the disposal phase of AI systems. Governments worldwide are beginning to implement extended producer responsibility (EPR) policies, which require technology manufacturers to take accountability for the environmental impact of their products throughout their entire lifecycle. In regions such as the European Union, strict e-waste management regulations mandate that electronic manufacturers participate in certified recycling programs and ensure the safe disposal of hazardous materials. However, enforcement remains inconsistent, and significant gaps exist in global e-waste tracking and management.

The future of AI hardware disposal will depend on advancements in recycling technology, regulatory enforcement, and industry-wide adoption of sustainable design principles. The growing urgency of AI-driven e-waste underscores the need for integrated lifecycle management strategies that account for the full environmental impact of AI infrastructure, from raw material extraction to end-of-life recovery. Without concerted efforts to improve hardware sustainability, the rapid expansion of AI will continue to exert pressure on global resources and waste management systems.

17.6 Mitigating Environmental Impact

The rapid expansion of AI has brought remarkable advancements in automation, language understanding, and decision-making, but it has also led to a significant and growing environmental impact. AI models, particularly large-scale deep learning systems, require massive computational resources for both training and inference. This results in high energy consumption, extensive carbon emissions, and resource-intensive hardware manufacturing. As AI adoption accelerates, the challenge of ensuring environmentally sustainable AI development becomes more urgent.

Addressing AI’s environmental footprint requires a multi-faceted approach, integrating energy-efficient AI models, optimized hardware, sustainable data center operations, and carbon-aware computing strategies. Additionally, AI systems must be designed with lifecycle sustainability in mind, ensuring that models remain efficient throughout their deployment, from training to inference.

A fundamental principle that must guide all efforts to mitigate AI’s environmental impact is Jevon’s Paradox. This paradox, observed by William Stanley Jevons in the 19th century⁷ (Jevons 1865), says that improvements in technological efficiency can lead to an increase in overall consumption. In the context of AI, even as we develop more energy-efficient models and hardware, the increased accessibility and adoption of AI technologies could lead to a net increase in energy consumption and resource utilization. Therefore, we must approach mitigation strategies with a keen awareness of this potential rebound effect, ensuring that efficiency gains do not inadvertently drive greater consumption. This section explores key strategies for mitigating AI’s environmental impact, beginning with sustainable AI development principles.

⁷ William Stanley Jevons observed during the Industrial Revolution that technological improvements in coal efficiency paradoxically led to increased coal consumption rather than conservation. In his 1865 book “The Coal Question”, he noted that more efficient steam engines made coal power more economical, which expanded its industrial applications and ultimately increased total coal usage. This principle became known as Jevons Paradox.

Jevons, William Stanley. 1865. The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal Mines. London: Macmillan; Co. https://www.econlib.org/library/YPDBooks/Jevons/jvnCQ.html.

This effect is illustrated in Figure 17.11. As AI systems become more efficient, the cost per unit of computation decreases, whether for language model tokens, computer vision inferences, or recommendation system predictions. In the figure, moving from point A to point B represents a drop in computation cost. However, this price reduction leads to increased usage across all AI applications, as shown by the corresponding shift from point C to point D on the horizontal axis. While there are savings from reduced costs, the total consumption of AI services increases even more rapidly, ultimately resulting in higher overall resource usage and environmental impact. This dynamic highlights the core of Jevon’s Paradox in AI: efficiency alone is not sufficient to guarantee sustainability.

Figure 17.11: Jevon’s Paradox in AI efficiency improvements and usage.

17.6.1 Sustainable Development

The design and development of AI models have historically prioritized performance, often at the expense of efficiency. However, as computational demands rise and AI systems scale, this approach is becoming increasingly unsustainable. A single training run of a large transformer-based model can emit as much carbon as five cars over their entire lifetime (Strubell, Ganesh, and McCallum 2019b). Furthermore, many AI models require frequent retraining to adapt to evolving data, compounding their energy consumption. Addressing these sustainability challenges requires a shift from brute-force computation to efficiency-driven innovation. By optimizing model architectures, reducing redundant training, and integrating sustainability principles throughout the AI lifecycle, the environmental impact of AI can be significantly reduced.

———. 2019b. “Energy and Policy Considerations for Deep Learning in NLP.” arXiv Preprint arXiv:1906.02243, June, 3645–50. https://doi.org/10.18653/v1/p19-1355.

Energy-Efficient Design

Many deep learning models rely on billions of parameters, requiring trillions of floating-point operations per second (FLOPS) during training and inference. While these large models achieve state-of-the-art performance, research indicates that much of their computational complexity is unnecessary. Many parameters contribute little to final predictions, leading to wasteful resource utilization. To mitigate this inefficiency, several optimization techniques have been developed to reduce the computational overhead of AI models while maintaining accuracy.

One of the most widely used methods for improving energy efficiency is pruning, a technique that removes unnecessary connections from a trained neural network. By systematically eliminating redundant weights, pruning reduces both the model size and the number of computations required during inference. Studies have shown that structured pruning can remove up to 90% of weights in models such as ResNet-50 while maintaining comparable accuracy. This approach enables AI models to operate efficiently on lower-power hardware, making them more suitable for deployment in resource-constrained environments.

Another technique for reducing energy consumption is quantization, which lowers the numerical precision of computations in AI models. Standard deep learning models typically use 32-bit floating-point precision, but many operations can be performed with 8-bit or even 4-bit integers without significant accuracy loss. By using lower precision, quantization reduces memory requirements, speeds up inference, and lowers power consumption. For example, NVIDIA’s TensorRT framework applies post-training quantization to deep learning models, achieving a threefold increase in inference speed while maintaining nearly identical accuracy. Similarly, Intel’s Q8BERT demonstrates that quantizing the BERT language model to 8-bit integers can reduce its size by a factor of four with minimal performance degradation (Zafrir et al. 2019).

Zafrir, Ofir, Guy Boudoukh, Peter Izsak, and Moshe Wasserblat. 2019. “Q8BERT: Quantized 8Bit BERT.” In 2019 Fifth Workshop on Energy Efficient Machine Learning and Cognitive Computing - NeurIPS Edition (EMC2-NIPS), 36–39. IEEE; IEEE. https://doi.org/10.1109/emc2-nips53020.2019.00016.

A third approach, knowledge distillation, allows large AI models to transfer their learned knowledge to smaller, more efficient models. In this process, a large teacher model trains a smaller student model to approximate its predictions, enabling the student model to achieve competitive performance with significantly fewer parameters. Google’s DistilBERT exemplifies this technique, retaining 97% of the original BERT model’s accuracy while using only 40% of its parameters. Knowledge distillation techniques enable AI practitioners to deploy lightweight models that require substantially less computational power while delivering high-quality predictions.

While these optimization techniques improve efficiency, they also introduce trade-offs. Pruning and quantization can lead to small reductions in model accuracy, requiring fine-tuning to balance performance and sustainability. Knowledge distillation, on the other hand, demands additional training cycles, meaning that energy savings are realized primarily during deployment rather than in the training phase. Furthermore, we must consider Jevon’s Paradox: will these efficiency gains lead to a proliferation of AI applications, ultimately increasing overall energy consumption? To counteract this, strategies that combine efficiency with limitations on resource usage are necessary. Nonetheless, these techniques represent essential strategies for reducing the energy footprint of AI models without compromising their effectiveness.

Lifecycle-Aware Systems

In addition to optimizing individual models, AI systems must be designed with a broader lifecycle-aware perspective. Many AI deployments operate with a short-term mindset, where models are trained, deployed, and then discarded within a few months. This frequent retraining cycle leads to excessive computational waste. By incorporating sustainability considerations into the AI development pipeline, it is possible to extend model lifespan, reduce unnecessary computation, and minimize environmental impact.

One of the most effective ways to reduce redundant computation is to limit the frequency of full model retraining. Many production AI systems do not require complete retraining from scratch; instead, they can be updated using incremental learning techniques that adapt existing models to new data. Transfer learning is a widely used approach in which a pre-trained model is fine-tuned on a new dataset, significantly reducing the computational cost compared to training a model from the ground up (Narang et al. 2021). This technique is particularly valuable for domain adaptation, where models trained on large general datasets can be customized for specific applications with minimal retraining.

Narang, Sharan, Hyung Won Chung, Yi Tay, Liam Fedus, Thibault Fevry, Michael Matena, Karishma Malkan, et al. 2021. “Do Transformer Modifications Transfer Across Implementations and Applications?” In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 21:1–67. 140. Association for Computational Linguistics. https://doi.org/10.18653/v1/2021.emnlp-main.465.

Henderson, Peter, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau. 2020. “Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning.” Journal of Machine Learning Research 21 (248): 1–43. http://arxiv.org/abs/2002.05651v2.

Another critical aspect of lifecycle-aware AI development is the integration of LCA methodologies. LCA provides a systematic framework for quantifying the environmental impact of AI systems at every stage of their lifecycle, from initial training to long-term deployment. Organizations such as MLCommons are actively developing sustainability benchmarks that measure factors such as energy efficiency per inference and carbon emissions per model training cycle (Henderson et al. 2020). By embedding LCA principles into AI workflows, developers can identify sustainability bottlenecks early in the design process and implement corrective measures before models enter production.

Beyond training efficiency and design evaluation, AI deployment strategies can further enhance sustainability. Cloud-based AI models often rely on centralized data centers, requiring significant energy for data transfer and inference. In contrast, edge computing enables AI models to run directly on end-user devices, reducing the need for constant cloud communication. Deploying AI models on specialized low-power hardware at the edge not only improves latency and privacy but also significantly decreases energy consumption (Xu et al. 2021).

Xu, Xiaolong, Fan Li, Wei Zhang, Liang He, and Ruidong Li. 2021. “Edge Intelligence: Architectures, Challenges, and Applications.” IEEE Internet of Things Journal 8 (6): 4229–49.

However, Jevon’s Paradox reminds us that optimizing individual stages might not lead to overall sustainability. For example, even if we improve the recyclability of AI hardware, increased production due to greater demand could still lead to increased resource depletion. Therefore, limiting the production of unneeded hardware is also important. By adopting a lifecycle-aware approach to AI development, practitioners can reduce the environmental impact of AI systems while promoting long-term sustainability.

Policy and Incentives

While technical optimizations play a crucial role in mitigating AI’s environmental impact, they must be reinforced by policy incentives and industry-wide commitments to sustainability. Several emerging initiatives aim to integrate sustainability principles into AI development at scale.

One promising approach is carbon-aware AI scheduling, where AI workloads are dynamically allocated based on the availability of renewable energy. Companies such as Google have developed scheduling algorithms that shift AI training jobs to times when wind or solar power is abundant, reducing reliance on fossil fuels (Patterson, Gonzalez, Le, et al. 2022). These strategies are particularly effective in large-scale data centers, where peak energy demand can be aligned with periods of low-carbon electricity generation.

Patterson, David, Joseph Gonzalez, Quoc Le, Maud Texier, and Jeff Dean. 2022. “Carbon-Aware Computing for Sustainable AI.” Communications of the ACM 65 (11): 50–58.

Benchmarks and leaderboards focused on sustainability are also gaining traction within the AI community. The ML.ENERGY Leaderboard, for example, ranks AI models based on energy efficiency and carbon footprint, encouraging researchers to optimize models not only for performance but also for sustainability. Similarly, MLCommons is working on standardized benchmarks that evaluate AI efficiency in terms of power consumption per inference, providing a transparent framework for comparing the environmental impact of different models.

Regulatory efforts are beginning to shape the future of sustainable AI. The European Union’s Sustainable Digital Markets Act⁸ has introduced guidelines for transparent AI energy reporting, requiring tech companies to disclose the carbon footprint of their AI operations. As regulatory frameworks evolve, organizations will face increasing pressure to integrate sustainability considerations into their AI development practices (Commission 2023).

⁸ Sustainable Digital Markets Act (SDMA): A legislative initiative by the European Union aimed at promoting transparency in AI energy consumption and enforcing sustainability standards in digital markets.

Commission, European. 2023. “Sustainable Digital Markets Act: Environmental Transparency in AI.”

By aligning technical optimizations with industry incentives and policy regulations, AI practitioners can ensure that sustainability becomes an integral component of AI development. The shift toward energy-efficient models, lifecycle-aware design, and transparent environmental reporting will be critical in mitigating AI’s ecological impact while continuing to drive innovation.

17.6.2 Infrastructure Optimization

The sustainability of AI systems is shaped not only by the efficiency of machine learning models but also by the infrastructure that powers them. While algorithmic improvements such as pruning, quantization, and knowledge distillation reduce computational requirements at the model level, the broader energy footprint of AI is largely dictated by how and where these computations are performed. Large-scale AI workloads are executed in cloud data centers, which house thousands of interconnected servers running continuously to support machine learning training and inference. These facilities consume enormous amounts of electricity, with some hyperscale data centers drawing over 100 megawatts of power, an amount comparable to the energy demand of a small city (Jones, Johnson, and Montgomery 2021). In addition to direct energy consumption, the cooling requirements of AI infrastructure introduce further sustainability challenges. Data centers must dissipate significant amounts of heat generated by AI accelerators, often relying on energy-intensive air conditioning or water-based cooling systems. As AI adoption continues to expand, these infrastructure-level considerations will play an increasingly central role in determining the overall environmental impact of machine learning systems.

Jones, Nicholas P., Mark Johnson, and Claire Montgomery. 2021. “The Environmental Impact of Data Centers: Challenges and Sustainable Solutions.” Energy Reports 7: 4381–92.

Addressing these challenges requires a shift toward energy-efficient AI infrastructure. The integration of renewable energy into cloud data centers, the adoption of advanced cooling strategies, and the development of carbon-aware workload scheduling can significantly reduce the carbon footprint of AI operations. By designing AI infrastructure to align with sustainability principles, it is possible to minimize the environmental cost of computation while maintaining the performance required for modern machine learning workloads. This section explores the key approaches to optimizing AI infrastructure, focusing on energy-efficient data centers, dynamic workload scheduling based on carbon intensity, and AI-driven cooling strategies. These advancements offer a pathway toward more sustainable AI deployment, ensuring that the growth of machine learning does not come at the expense of long-term environmental responsibility.

Green Data Centers

The increasing computational demands of AI have made data centers one of the largest consumers of electricity in the digital economy. Large-scale cloud data centers provide the infrastructure necessary for training and deploying machine learning models, but their energy consumption is substantial. A single hyperscale data center can consume over 100 megawatts of power, a level comparable to the electricity usage of a small city. Without intervention, the continued growth of AI workloads threatens to push the energy consumption of data centers beyond sustainable levels. The industry must adopt strategies to optimize power efficiency, integrate renewable energy sources, and improve cooling mechanisms to mitigate the environmental impact of AI infrastructure.

One of the most promising approaches to reducing data center emissions is the transition to renewable energy. Major cloud providers, including Google, Microsoft, and Amazon Web Services, have committed to powering their data centers with renewable energy, but implementation challenges remain. Unlike fossil fuel plants, which provide consistent electricity output, renewable sources such as wind and solar are intermittent, with generation levels fluctuating throughout the day. To address this variability, AI infrastructure must incorporate energy storage solutions, such as large-scale battery deployments, and implement intelligent scheduling mechanisms that shift AI workloads to times when renewable energy availability is highest. Google, for example, has set a goal to operate its data centers on 24/7 carbon-free energy by 2030, ensuring that every unit of electricity consumed is matched with renewable generation rather than relying on carbon offsets alone.

Cooling systems represent another major contributor to the energy footprint of data centers, often accounting for 30 to 40 percent of total electricity consumption. Traditional cooling methods rely on air conditioning units and mechanical chillers, both of which require significant power and water resources. To improve efficiency, data centers are adopting alternative cooling strategies that reduce energy waste. Liquid cooling, which transfers heat away from AI accelerators using specially designed coolant systems, is significantly more effective than traditional air cooling and is now being deployed in high-density computing clusters. Free-air cooling, which utilizes natural airflow instead of mechanical refrigeration, has also been adopted in temperate climates, where external conditions allow for passive cooling. Microsoft has taken this a step further by deploying underwater data centers that use the surrounding ocean as a natural cooling mechanism, reducing the need for active temperature regulation.

Beyond hardware-level optimizations, AI itself is being used to improve the energy efficiency of data center operations. DeepMind has developed machine learning algorithms capable of dynamically adjusting cooling parameters based on real-time sensor data. These AI-powered cooling systems analyze temperature, humidity, and fan speeds, making continuous adjustments to optimize energy efficiency. When deployed in Google’s data centers, DeepMind’s system achieved a 40 percent reduction in cooling energy consumption, demonstrating the potential of AI to enhance the sustainability of the infrastructure that supports machine learning workloads.

However, Jevon’s Paradox suggests that even highly efficient data centers could contribute to increased consumption if they enable a massive expansion of AI-driven services. Optimizing the energy efficiency of data centers is critical to reducing the environmental impact of AI, but efficiency alone is not enough. We must also consider strategies for limiting the growth of data center capacity. The integration of renewable energy, the adoption of advanced cooling solutions, and the use of AI-driven optimizations can significantly decrease the carbon footprint of AI infrastructure. As AI continues to scale, these innovations will play a central role in ensuring that machine learning remains aligned with sustainability goals.

Carbon-Aware Scheduling

Beyond improvements in hardware and cooling systems, optimizing when and where AI workloads are executed is another critical strategy for reducing AI’s environmental impact. The electricity used to power data centers comes from energy grids that fluctuate in carbon intensity based on the mix of power sources available at any given time. Fossil fuel-based power plants supply a significant portion of global electricity, but the share of renewable energy varies by region and time of day. Without optimization, AI workloads may be executed when carbon-intensive energy sources dominate the grid, unnecessarily increasing emissions. By implementing carbon-aware scheduling, AI computations can be dynamically shifted to times and locations where low-carbon energy is available, significantly reducing emissions without sacrificing performance.

Google has pioneered one of the most advanced implementations of carbon-aware computing in its cloud infrastructure. In 2020, the company introduced a scheduling system that delays non-urgent AI tasks until times when renewable energy sources such as solar or wind power are more abundant. This approach enables AI workloads to align with the natural variability of clean energy availability, reducing reliance on fossil fuels while maintaining high computational efficiency. Google has further extended this strategy by geographically distributing AI workloads, moving computations to data centers in regions where clean energy is more accessible. By shifting large-scale AI training jobs from fossil fuel-heavy grids to low-carbon power sources, the company has demonstrated that significant emissions reductions can be achieved through intelligent workload placement.

The potential for carbon-aware scheduling extends beyond hyperscale cloud providers. Companies that rely on AI infrastructure can integrate carbon intensity metrics into their own computing pipelines, making informed decisions about when to run machine learning jobs. Microsoft’s sustainability-aware cloud computing initiative allows organizations to select carbon-optimized virtual machines, ensuring that workloads are executed with the lowest possible emissions. Research efforts are also underway to develop open-source carbon-aware scheduling frameworks, enabling a broader range of AI practitioners to incorporate sustainability into their computing strategies.

The effectiveness of carbon-aware AI scheduling depends on accurate real-time data about grid emissions. Electricity providers and sustainability organizations have begun publishing grid carbon intensity data through publicly available APIs, allowing AI systems to dynamically respond to changes in energy supply. For instance, the Electricity Maps API provides real-time CO₂ emissions data for power grids worldwide, enabling AI infrastructure to adjust computational workloads based on carbon availability. As access to grid emissions data improves, carbon-aware computing will become a scalable and widely adoptable solution for reducing the environmental impact of AI operations.

By shifting AI computations to times and places with cleaner energy sources, carbon-aware scheduling represents a powerful tool for making AI infrastructure more sustainable. Unlike hardware-based optimizations that require physical upgrades, scheduling improvements can be implemented through software, offering an immediate and cost-effective pathway to emissions reductions. As more organizations integrate carbon-aware scheduling into their AI workflows, the cumulative impact on reducing global AI-related carbon emissions could be substantial.

While these strategies apply broadly to AI workloads, inference operations present unique sustainability challenges and opportunities. Unlike training, which represents a one-time energy cost, inference constitutes an ongoing and growing energy demand as AI applications scale worldwide. Cloud providers are increasingly adopting carbon-aware scheduling specifically for inference workloads, dynamically shifting these operations to regions powered by abundant renewable energy (Alvim et al. 2022). However, as shown in Figure 17.12, the variability of renewable energy production presents significant challenges. The European grid data illustrates how renewable sources fluctuate throughout the day—solar energy peaks at midday, while wind energy shows distinct peaks in mornings and evenings. Currently, fossil and coal-based generation methods supplement energy needs when renewables fall short.

Alvim, Mário S., Konstantinos Chatzikokolakis, Yusuke Kawamoto, and Catuscia Palamidessi. 2022. “Information Leakage Games: Exploring Information as a Utility Function.” ACM Transactions on Privacy and Security 25 (3): 1–36. https://doi.org/10.1145/3517330.

Figure 17.12: Energy sources and generation capabilities. Source: Energy Charts.

To fully leverage carbon-aware scheduling for AI inference workloads, innovation in energy storage solutions is essential for consistent renewable energy use. The base energy load is currently met with nuclear energy—a constant source that produces no direct carbon emissions but lacks the flexibility to accommodate renewable energy variability. Tech companies like Microsoft have shown interest in nuclear energy to power their data centers, as their more constant demand profile (compared to residential use) aligns well with nuclear generation characteristics.

Beyond scheduling, optimizing inference sustainability requires complementary hardware and software innovations. Model quantization techniques enable lower-precision arithmetic to significantly cut power consumption without sacrificing accuracy (Gholami 2021). Knowledge distillation methods allow compact, energy-efficient models to replicate the performance of larger, resource-intensive networks (Hinton, Vinyals, and Dean 2015). Coupled with specialized inference accelerators like Google’s TPUs, these approaches substantially reduce inference’s environmental impact.

Gholami, Amir et al. 2021. “A Survey of Quantization Methods for Efficient Neural Network Inference.” IEEE Transactions on Neural Networks and Learning Systems 32 (10): 4562–81. https://doi.org/10.1109/TNNLS.2021.3088493.

Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. 2015. “Distilling the Knowledge in a Neural Network.” arXiv Preprint arXiv:1503.02531, March. http://arxiv.org/abs/1503.02531v1.

You, Jie, Jae-Won Chung, and Mosharaf Chowdhury. 2023. “Zeus: Understanding and Optimizing GPU Energy Consumption of DNN Training.” In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23), 119–39. Boston, MA: USENIX Association. https://www.usenix.org/conference/nsdi23/presentation/you.

Chung, Jae-Won, Yile Gu, Insu Jang, Luoxi Meng, Nikhil Bansal, and Mosharaf Chowdhury. 2023. “Reducing Energy Bloat in Large Model Training.” ArXiv Preprint abs/2312.06902 (December). http://arxiv.org/abs/2312.06902v3.

Software frameworks specifically designed for energy efficiency also play a crucial role. Energy-aware AI frameworks, such as Zeus (You, Chung, and Chowdhury 2023) and Perseus (Chung et al. 2023), balance computational speed and power efficiency during both training and inference. These platforms optimize model execution by analyzing trade-offs between speed and energy consumption, facilitating widespread adoption of energy-efficient AI strategies, particularly for inference operations that must run continuously at scale.

AI-Driven Thermal Optimization

Cooling systems are one of the most energy-intensive components of AI infrastructure, often accounting for 30-40% of total data center electricity consumption. As AI workloads become more computationally demanding, the heat generated by high-performance accelerators, such as GPUs and TPUs, continues to increase. Without efficient cooling solutions, data centers must rely on power-hungry air conditioning systems or water-intensive thermal management strategies, both of which contribute to AI’s overall environmental footprint. To address this challenge, AI-driven cooling optimization has emerged as a powerful strategy for improving energy efficiency while maintaining reliable operations.

DeepMind has demonstrated the potential of AI-driven cooling by deploying machine learning models to optimize temperature control in Google’s data centers. Traditional cooling systems rely on fixed control policies, making adjustments based on predefined thresholds for temperature and airflow. However, these rule-based systems often operate inefficiently, consuming more energy than necessary. By contrast, DeepMind’s AI-powered cooling system continuously analyzes real-time sensor data, including temperature, humidity, cooling pump speeds, and fan activity, to identify the most energy-efficient configuration for a given workload. Using deep reinforcement learning, the system dynamically adjusts cooling settings to minimize energy consumption while ensuring that computing hardware remains within safe operating temperatures.

When deployed in production, DeepMind’s AI-driven cooling system achieved a 40% reduction in cooling energy usage, leading to an overall 15% reduction in total data center power consumption. This level of efficiency improvement demonstrates how AI itself can be used to mitigate the environmental impact of machine learning infrastructure. The success of DeepMind’s system has inspired further research into AI-driven cooling, with other cloud providers exploring similar machine learning-based approaches to dynamically optimize thermal management.

Beyond AI-driven control systems, advances in liquid cooling and immersion cooling are further improving the energy efficiency of AI infrastructure. Unlike traditional air cooling, which relies on the circulation of cooled air through server racks, liquid cooling transfers heat directly away from high-performance AI chips using specially designed coolants. This approach significantly reduces the energy required for heat dissipation, allowing data centers to operate at higher densities with lower power consumption. Some facilities have taken this concept even further with immersion cooling, where entire server racks are submerged in non-conductive liquid coolants. This technique eliminates the need for traditional air-based cooling systems entirely, drastically cutting down on electricity usage and water consumption.

Microsoft has also explored innovative cooling solutions, deploying underwater data centers that take advantage of natural ocean currents to dissipate heat. By placing computing infrastructure in sealed submersible enclosures, Microsoft has demonstrated that ocean-based cooling can reduce power usage while extending hardware lifespan due to the controlled and stable underwater environment. While such approaches are still experimental, they highlight the growing interest in alternative cooling technologies that can make AI infrastructure more sustainable.

AI-driven cooling and thermal management represent an immediate and scalable opportunity for reducing the environmental impact of AI infrastructure. Unlike major hardware upgrades, which require capital-intensive investment, software-based cooling optimizations can be deployed rapidly across existing data centers. By leveraging AI to enhance cooling efficiency, in combination with emerging liquid and immersion cooling technologies, the industry can significantly reduce energy consumption, lower operational costs, and contribute to the long-term sustainability of AI systems.

17.6.3 Addressing Full Environmental Footprint

As AI systems continue to scale, efforts to mitigate their environmental impact have largely focused on improving energy efficiency in model design and optimizing data center infrastructure. While these advancements are essential, they only address part of the problem. AI’s environmental impact extends far beyond operational energy use, encompassing everything from the water consumption in semiconductor manufacturing to the growing burden of electronic waste. A truly sustainable AI ecosystem must account for the full life cycle of AI hardware and software, integrating sustainability at every stage—from material sourcing to disposal.

Earlier in this chapter, we explored the LCA of AI systems, highlighting the substantial carbon emissions, water consumption, and material waste associated with AI hardware manufacturing and deployment. Many of these environmental costs are embedded in the supply chain and do not appear in operational energy reports, leading to an incomplete picture of AI’s true sustainability. Moreover, data centers remain water-intensive, with cooling systems consuming millions of gallons per day, and AI accelerators are often refreshed on short life cycles, leading to mounting e-waste.

This section builds on those discussions by examining how AI’s broader environmental footprint can be reduced. We explore strategies to mitigate AI’s supply chain impact, curb water consumption, and extend hardware longevity. Moving beyond optimizing infrastructure, this approach takes a holistic view of AI sustainability, ensuring that improvements are not just localized to energy efficiency but embedded throughout the entire AI ecosystem.

Revisiting Life Cycle Impact

AI’s environmental footprint extends far beyond electricity consumption during model training and inference. The full life cycle of AI systems—from hardware manufacturing to disposal—contributes significantly to global carbon emissions, resource depletion, and electronic waste. Earlier in this chapter, we examined the LCA of AI hardware, which revealed that emissions are not solely driven by power consumption but also by the materials and processes involved in fabricating AI accelerators, storage devices, and networking infrastructure.

One of the most striking findings from LCA studies is the embodied carbon cost of AI hardware. Unlike operational emissions, which can be reduced by shifting to cleaner energy sources, embodied emissions result from the raw material extraction, semiconductor fabrication, and supply chain logistics that precede an AI accelerator’s deployment. Research indicates that manufacturing emissions alone can account for up to 30% of an AI system’s total carbon footprint, with this number potentially growing as data centers improve their reliance on renewable energy sources.

Moreover, AI’s water consumption has often been overlooked in sustainability discussions. Semiconductor fabrication plants—where AI accelerators are produced—are among the most water-intensive industrial facilities in the world, consuming millions of gallons daily for wafer cleaning and chemical processing. Data centers, too, rely on large amounts of water for cooling, with some hyperscale facilities using as much as 450,000 gallons per day—a number that continues to rise as AI workloads become more power-dense. Given that many of the world’s chip manufacturing hubs are located in water-stressed regions, such as Taiwan and Arizona, AI’s dependence on water raises serious sustainability concerns.

Beyond emissions and water use, AI hardware also contributes to a growing e-waste problem. The rapid evolution of AI accelerators has led to short hardware refresh cycles, where GPUs and TPUs are frequently replaced with newer, more efficient versions. While improving efficiency is critical, discarding functional hardware after only a few years leads to unnecessary electronic waste and resource depletion. Many AI chips contain rare earth metals and toxic components, which, if not properly recycled, can contribute to environmental pollution.

Mitigating AI’s environmental impact requires addressing these broader challenges—not just through energy efficiency improvements but by rethinking AI’s hardware life cycle, reducing water-intensive processes, and developing sustainable recycling practices. In the following sections, we explore strategies to tackle these issues head-on, ensuring that AI’s progress aligns with long-term sustainability goals.

Mitigating Supply Chain Impact

Addressing AI’s environmental impact requires intervention at the supply chain level, where significant emissions, resource depletion, and waste generation occur before AI hardware even reaches deployment. While much of the discussion around AI sustainability focuses on energy efficiency in data centers, the embodied carbon emissions from semiconductor fabrication, raw material extraction, and hardware transportation represent a substantial and often overlooked portion of AI’s total footprint. These supply chain emissions are difficult to offset, making it essential to develop strategies that reduce their impact at the source.

One of the primary concerns is the carbon intensity of semiconductor manufacturing. Fabricating AI accelerators such as GPUs, TPUs, and custom ASICs requires extreme precision and involves processes such as EUV lithography, chemical vapor deposition, and ion implantation, each of which consumes vast amounts of electricity. Since many semiconductor manufacturing hubs operate in regions where grid electricity is still predominantly fossil-fuel-based, the energy demands of chip fabrication contribute significantly to AI’s carbon footprint. Research suggests that semiconductor fabrication alone can account for up to 30% of an AI system’s total emissions, underscoring the need for more sustainable manufacturing processes.

Beyond carbon emissions, AI’s reliance on rare earth elements and critical minerals presents additional sustainability challenges. High-performance AI hardware depends on materials such as gallium, neodymium, and cobalt, which are essential for producing efficient and powerful computing components. However, extracting these materials is highly resource-intensive and often results in toxic waste, deforestation, and habitat destruction. The environmental cost is compounded by geopolitical factors, as over 90% of the world’s rare earth refining capacity is controlled by China, creating vulnerabilities in AI’s global supply chain. Ensuring responsible sourcing of these materials is critical to reducing AI’s ecological and social impact.

Several approaches can mitigate the environmental burden of AI’s supply chain. Reducing the energy intensity of chip manufacturing is one avenue, with some semiconductor manufacturers exploring low-energy fabrication processes and renewable-powered production facilities. Another approach focuses on extending the lifespan of AI hardware, as frequent hardware refresh cycles contribute to unnecessary waste. AI accelerators are often designed for peak training performance but remain viable for inference workloads long after they are retired from high-performance computing clusters. Repurposing older AI chips for less computationally intensive tasks, rather than discarding them outright, could significantly reduce the frequency of hardware replacement.

Recycling and closed-loop supply chains also play a crucial role in making AI hardware more sustainable. Recovering and refining valuable materials from retired GPUs, TPUs, and ASICs can reduce reliance on virgin resource extraction while minimizing e-waste. Industry-wide recycling initiatives, combined with hardware design that prioritizes recyclability, could significantly improve AI’s long-term sustainability.

Prioritizing supply chain sustainability in AI is not just an environmental necessity but also an opportunity for innovation. By integrating energy-efficient fabrication, responsible material sourcing, and circular hardware design, the AI industry can take meaningful steps toward reducing its environmental impact before these systems ever reach operation. These efforts, combined with continued advances in energy-efficient AI computing, will be essential to ensuring that AI’s growth does not come at an unsustainable ecological cost.

Reducing Water and Resource Consumption

Mitigating AI’s environmental impact requires direct action to reduce its water consumption and resource intensity. AI’s reliance on semiconductor fabrication and data centers creates significant strain on water supplies and critical materials, particularly in regions already facing resource scarcity. Unlike carbon emissions, which can be offset through renewable energy, water depletion and material extraction have direct, localized consequences, making it essential to integrate sustainability measures at the design and operational levels.

One of the most effective strategies for reducing AI’s water footprint is improving water recycling in semiconductor fabrication. Leading manufacturers are implementing closed-loop water systems, which allow fabs to reuse and treat water rather than continuously consuming fresh supplies. Companies such as Intel and TSMC have already developed advanced filtration and reclamation processes that recover over 80% of the water used in chip production. Expanding these efforts across the industry is essential for minimizing the impact of AI hardware manufacturing.

Similarly, data centers can reduce water consumption by optimizing cooling systems. Many hyperscale facilities still rely on evaporative cooling, which consumes vast amounts of water. Transitioning to direct-to-chip liquid cooling or air-based cooling technologies can significantly reduce water use. In regions with water scarcity, some operators have begun using wastewater or desalinated water for cooling rather than drawing from potable sources. These methods help mitigate the environmental impact of AI infrastructure while maintaining efficient operation.

On the materials side, reducing AI’s dependency on rare earth metals and critical minerals is crucial for long-term sustainability. While some materials, such as silicon, are abundant, others—such as gallium, neodymium, and cobalt—are subject to geopolitical constraints and environmentally damaging extraction methods. Researchers are actively exploring alternative materials and low-waste manufacturing processes to reduce reliance on these limited resources. Additionally, recycling programs for AI accelerators and other computing hardware can recover valuable materials, reducing the need for virgin extraction.

Beyond individual mitigation efforts, industry-wide collaboration is necessary to develop standards for responsible water use, material sourcing, and recycling programs. Governments and regulatory bodies can also incentivize sustainable practices by enforcing water conservation mandates, responsible mining regulations, and e-waste recycling requirements. By prioritizing these mitigation strategies, the AI industry can work toward minimizing its ecological footprint while continuing to advance technological progress.

Systemic Sustainability Approaches

Mitigating AI’s environmental impact requires more than isolated optimizations—it demands a systemic shift toward sustainable AI development. Addressing the long-term sustainability of AI means integrating circular economy principles, establishing regulatory policies, and fostering industry-wide collaboration to ensure that sustainability is embedded into the AI ecosystem from the ground up.

Jevon’s Paradox highlights the limitations of focusing solely on individual efficiency improvements. We need systemic solutions that address the broader drivers of AI consumption. This includes policies that promote sustainable AI practices, incentives for responsible resource usage, and public awareness campaigns that encourage mindful AI consumption.

One of the most effective ways to achieve lasting sustainability is by aligning AI development with circular economy principles. Unlike the traditional linear model of “build, use, discard,” a circular approach prioritizes reuse, refurbishment, and recycling to extend the lifespan of AI hardware (Stahel 2016). Manufacturers and cloud providers can adopt modular hardware designs, allowing individual components—such as memory and accelerators—to be upgraded without replacing entire servers. In addition, AI hardware should be designed with recyclability in mind, ensuring that valuable materials can be extracted and reused instead of contributing to electronic waste.

Stahel, Walter R. 2016. “The Circular Economy.” Nature 531 (7595): 435–38. https://doi.org/10.1038/531435a.

Masanet, Eric, Arman Shehabi, Nuoa Lei, Sarah Smith, and Jonathan Koomey. 2020b. “Recalibrating Global Data Center Energy-Use Estimates.” Science 367 (6481): 984–86. https://doi.org/10.1126/science.aba3758.

Regulatory frameworks also play a crucial role in enforcing sustainability standards. Governments can introduce carbon transparency mandates, requiring AI infrastructure providers to report the full lifecycle emissions of their operations, including embodied carbon from manufacturing (Masanet et al. 2020b). Additionally, stricter water use regulations for semiconductor fabs and e-waste recycling policies can help mitigate AI’s resource consumption. Some jurisdictions have already implemented extended producer responsibility laws, which hold manufacturers accountable for the end-of-life disposal of their products. Expanding these policies to AI hardware could incentivize more sustainable design practices.

At the industry level, collaborative efforts are essential for scaling sustainable AI practices. Leading AI companies and research institutions should establish shared sustainability benchmarks that track energy efficiency, carbon footprint, and resource usage. Furthermore, standardized green AI certifications could guide consumers and enterprises toward more sustainable technology choices (Strubell, Ganesh, and McCallum 2019a). Cloud providers can also commit to 24/7 carbon-free energy (CFE) goals, ensuring that AI workloads are powered by renewable sources in real-time rather than relying on carbon offsets that fail to drive meaningful emissions reductions.

Strubell, Emma, Ananya Ganesh, and Andrew McCallum. 2019a. “Energy and Policy Considerations for Deep Learning in NLP.” In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645–50. Association for Computational Linguistics. https://doi.org/10.18653/v1/p19-1355.

Achieving systemic change in AI sustainability requires a multi-stakeholder approach. Governments, industry leaders, and researchers must work together to set sustainability standards, invest in greener infrastructure, and transition toward a circular AI economy. By embedding sustainability into the entire AI development pipeline, the industry can move beyond incremental optimizations and build a truly sustainable foundation for future innovation.

17.6.4 Case Study: Google’s Framework

To mitigate emissions from rapidly expanding AI workloads, Google engineers identified four key optimization areas—designated as the “4 Ms”—where systematic improvements collectively reduce the carbon footprint of machine learning:

• Model: The selection of efficient AI architectures reduces computation requirements by 5-10X without compromising model quality. Google has extensively researched sparse models and neural architecture search methodologies, resulting in efficient architectures such as the Evolved Transformer and Primer.

• Machine: The implementation of AI-specific hardware offers 2-5X improvements in performance per watt compared to general-purpose systems. Google’s TPUs demonstrate 5-13X greater carbon efficiency relative to non-optimized GPUs.

• Mechanization: The utilization of optimized cloud computing infrastructure with high utilization rates yields 1.4-2X energy reductions compared to conventional on-premise data centers. Google’s facilities consistently exceed industry standards for PUE.

• Map: The strategic positioning of data centers in regions with low-carbon electricity supplies reduces gross emissions by 5-10X. Google maintains real-time monitoring of renewable energy usage across its global infrastructure.

The combined effect of these practices produces multiplicative efficiency gains. For instance, implementing the optimized Transformer model on TPUs in strategically located data centers reduced energy consumption by a factor of 83 and CO₂ emissions by a factor of 747.

Despite substantial growth in AI deployment across Google’s product ecosystem, systematic efficiency improvements have effectively constrained energy consumption growth. A significant indicator of this progress is the observation that AI workloads have maintained a consistent 10% to 15% proportion of Google’s total energy consumption from 2019 through 2021. As AI functionality expanded across Google’s services, corresponding increases in compute cycles were offset by advancements in algorithms, specialized hardware, infrastructure design, and geographical optimization.

Empirical case studies demonstrate how engineering principles focused on sustainable AI development enable simultaneous improvements in both performance and environmental impact. For example, comparative analysis between GPT-3 (considered state-of-the-art in mid-2020) and Google’s GLaM model reveals improved accuracy metrics alongside reduced training computation requirements and lower-carbon energy sources—resulting in a 14-fold reduction in CO₂ emissions within an 18-month development cycle.

Furthermore, Google’s analysis indicates that previous published estimates overestimated machine learning’s energy requirements by factors ranging from 100 to 100,000X due to methodological limitations and absence of empirical measurements. Through transparent reporting of optimization metrics, Google provides a factual basis for efficiency initiatives while correcting disproportionate projections regarding machine learning’s environmental impact.

While substantial progress has been achieved in constraining the carbon footprint of AI operations, Google acknowledges that continued efficiency advancements are essential for responsible innovation as AI applications proliferate. Their ongoing optimization framework encompasses:

1. Life-Cycle Analysis: Demonstrating that computational investments such as neural architecture search, while initially resource-intensive, generate significant downstream efficiencies that outweigh initial costs. Despite higher energy expenditure during the discovery phase compared to manual engineering approaches, NAS ultimately reduces cumulative emissions by generating optimized architectures applicable across numerous deployments.

2. Resource Allocation Prioritization: Concentrating sustainability initiatives on data center and server-side optimization where energy consumption is most concentrated. While Google continues to enhance inference efficiency on edge devices, primary focus remains on training infrastructure and renewable energy procurement to maximize environmental return on investment.

3. Economies of Scale: Leveraging the efficiency advantages inherent in well-designed cloud infrastructure through workload consolidation. As computation transitions from distributed on-premise environments to centralized providers with robust sustainability frameworks, aggregate emissions reductions accelerate.

4. Renewable Energy Integration: Prioritizing renewable energy procurement—with Google achieving 100% matching of energy consumption with renewable sources since 2017—to further reduce the environmental impact of computational workloads.

These integrated approaches indicate that AI efficiency improvements are accelerating rather than plateauing. Google’s multifaceted strategy combining systematic measurement, carbon-aware development methodologies, transparency in reporting, and renewable energy transition establishes a replicable framework for sustainable AI scaling. These empirical results provide a foundation for broader industry adoption of comprehensive sustainability practices.

17.7 Embedded AI and E-Waste

The deployment of AI is rapidly expanding beyond centralized data centers into edge and embedded devices, enabling real-time decision-making without requiring constant cloud connectivity. This shift has led to major efficiency gains, reducing latency, bandwidth consumption, and network congestion while enabling new applications in smart consumer devices, industrial automation, healthcare, and autonomous systems. However, the rise of embedded AI brings new environmental challenges, particularly regarding electronic waste, disposable smart devices, and planned obsolescence.

Unlike high-performance AI accelerators in data centers, which are designed for long-term use and high computational throughput, embedded AI hardware is often small, low-cost, and disposable. Many AI-powered IoT sensors, wearables, and smart appliances are built with short lifespans and limited upgradeability, making them difficult—if not impossible—to repair or recycle (C. P. Baldé 2017). As a result, these devices contribute to a rapidly growing electronic waste crisis, one that remains largely overlooked in discussions on AI sustainability.

The scale of this issue is staggering. The number of Internet of Things (IoT) devices is projected to exceed 30 billion by 2030, with AI-powered chips embedded into everything from household appliances and medical implants to industrial monitoring systems and agricultural sensors (Statista 2022). Many of these devices will become obsolete within just a few years, leading to an unprecedented surge in e-waste. Without sustainable design practices and improved lifecycle management, the expansion of AI at the edge risks exacerbating global electronic waste accumulation and straining recycling infrastructure.

Statista. 2022. “Number of Internet of Things (IoT) Connected Devices Worldwide from 2019 to 2030.” https://www.statista.com/statistics/802690/worldwide-connected-devices-by-access-technology/.

Figure 17.13: Number of Internet of Things (IoT) connected devices worldwide from 2019 to 2023. Source: Statista.

While AI-powered data centers have been scrutinized for their carbon footprint and energy demands, far less attention has been paid to the environmental cost of embedding AI into billions of short-lived devices. Addressing this challenge requires rethinking how AI hardware is designed, manufactured, and disposed of, ensuring that edge AI systems contribute to technological progress without leaving behind an unsustainable legacy of waste.

17.7.1 E-Waste Crisis

Electronic waste, or e-waste, is one of the fastest-growing environmental challenges of the digital age. Defined as discarded electronic devices containing batteries, circuit boards, and semiconductor components, e-waste presents severe risks to both human health and the environment. Toxic materials such as lead, mercury, cadmium, and brominated flame retardants, commonly found in AI-enabled hardware, can contaminate soil and groundwater when improperly disposed of. Despite the potential for recycling and material recovery, most e-waste remains improperly handled, leading to hazardous waste accumulation and significant environmental degradation.

The scale of the problem is staggering. Today, global e-waste production exceeds 50 million metric tons annually, with projections indicating that this figure will surpass 75 million tons by 2030 as consumer electronics and AI-powered IoT devices continue to proliferate. According to the United Nations, e-waste generation could reach 120 million tons per year by 2050 if current consumption patterns persist (Un and Forum 2019). The combination of short product lifespans, rising global demand, and limited recycling infrastructure has accelerated this crisis.

AI-driven consumer devices, such as smart speakers, fitness trackers, and home automation systems, are among the most significant contributors to e-waste. Unlike modular and serviceable computing systems, many of these devices are designed to be disposable, meaning that when a battery fails or a component malfunctions, the entire product is discarded rather than repaired. This built-in disposability exacerbates the unsustainable cycle of consumption and waste, leading to higher material extraction rates and increased pressure on waste management systems.

Developing nations are disproportionately affected by e-waste dumping, as they often lack the infrastructure to process obsolete electronics safely. In 2019, only 13% to 23% of e-waste in lower-income countries was formally collected for recycling, with the remainder either incinerated, illegally dumped, or manually dismantled in unsafe conditions (Un and Forum 2019). Many discarded AI-powered devices end up in informal recycling operations, where low-paid workers are exposed to hazardous materials without proper protective equipment. Open-air burning of plastic components and crude metal extraction methods release toxic fumes and heavy metals into the surrounding environment, posing severe health risks.

Un, and World Economic Forum. 2019. A New Circular Vision for Electronics, Time for a Global Reboot. PACE - Platform for Accelerating the Circular Economy. https://www3.weforum.org/docs/WEF\_A\_New\_Circular\_Vision\_for\_Electronics.pdf.

The global recycling rate for e-waste remains alarmingly low, with only 20% of all discarded electronics processed through environmentally sound recycling channels. The remaining 80% is either landfilled, incinerated, or dumped illegally, leading to long-term environmental contamination and resource depletion. Without stronger policies, better product design, and expanded e-waste management systems, the rapid growth of AI-powered devices will significantly worsen this crisis.

AI-driven electronics should not become another major contributor to the global e-waste problem. Tackling this challenge requires a multi-pronged approach, including more sustainable design practices, stronger regulatory oversight, and greater investment in global e-waste recycling infrastructure. Without intervention, AI’s environmental impact will extend far beyond its energy consumption, leaving behind a legacy of toxic waste and resource depletion.

17.7.2 Disposable Electronics

The rapid proliferation of low-cost AI-powered microcontrollers, smart sensors, and connected devices has transformed various industries, from consumer electronics and healthcare to industrial automation and agriculture. While these embedded AI systems enable greater efficiency and automation, their short lifespans and non-recyclable designs pose a significant sustainability challenge. Many of these devices are treated as disposable electronics, designed with limited durability, non-replaceable batteries, and little to no repairability, making them destined for the waste stream within just a few years of use.

One of the primary drivers of AI-powered device disposability is the falling cost of microelectronics. The miniaturization of computing hardware has enabled manufacturers to embed tiny AI processors and wireless connectivity modules into everyday products, often for under $1 per chip. As a result, AI functionality is increasingly being integrated into single-use and short-lived products, including smart packaging, connected medical devices, wearables, and home appliances. While these innovations improve convenience and real-time data collection, they lack proper end-of-life management strategies, leading to a surge in hard-to-recycle electronic waste (V. Forti 2020).

V. Forti, R. Kuehr, C. P. Baldé. 2020. The Global e-Waste Monitor 2020: Quantities, Flows, and Circular Economy Potential. United Nations University, International Telecommunication Union,; International Solid Waste Association. https://ewastemonitor.info.

Non-Replaceable Batteries Cost

Many disposable AI devices incorporate sealed, non-replaceable lithium-ion batteries, making them inherently unsustainable. Smart earbuds, wireless sensors, and even some fitness trackers lose functionality entirely once their batteries degrade, forcing consumers to discard the entire device. Unlike modular electronics with user-serviceable components, most AI-powered wearables and IoT devices are glued or soldered shut, preventing battery replacement or repair.

This issue extends beyond consumer gadgets. Industrial AI sensors and remote monitoring devices, often deployed in agriculture, infrastructure, and environmental monitoring, frequently rely on non-replaceable batteries with a limited lifespan. Once depleted, these sensors—many of which are installed in remote or difficult-to-access locations—become e-waste, requiring costly and environmentally disruptive disposal or replacement (Ciez and Whitacre 2019).

Ciez, Rebecca E., and J. F. Whitacre. 2019. “Examining Different Recycling Processes for Lithium-Ion Batteries.” Nature Sustainability 2 (2): 148–56. https://doi.org/10.1038/s41893-019-0222-5.

The environmental impact of battery waste is particularly concerning. Lithium mining, essential for battery production, is an energy-intensive process that consumes vast amounts of water and generates harmful byproducts. Additionally, the improper disposal of lithium batteries poses fire and explosion risks, particularly in landfills and waste processing facilities. As the demand for AI-powered devices grows, addressing the battery sustainability crisis will be critical to mitigating AI’s long-term environmental footprint.

Recycling Challenges

Unlike traditional computing hardware—such as desktop computers and enterprise servers, which can be disassembled and refurbished—most AI-enabled consumer electronics are not designed for recycling. Many of these devices contain mixed-material enclosures, embedded circuits, and permanently attached components, making them difficult to dismantle and recover materials from (Patel et al. 2016).

Patel, Paresh D., Absar Lakdawala, Sajan Chourasia, and Rajesh N. Patel. 2016. “Bio Fuels for Compression Ignition Engine: A Review on Engine Performance, Emission and Life Cycle Analysis.” Renewable and Sustainable Energy Reviews 65 (November): 24–43. https://doi.org/10.1016/j.rser.2016.06.010.

Additionally, AI-powered IoT devices are often too small to be efficiently recycled using conventional e-waste processing methods. Large-scale electronics, such as laptops and smartphones, have well-established recycling programs that allow for material recovery. In contrast, tiny AI-powered sensors, earbuds, and embedded chips are often too costly and labor-intensive to separate into reusable components. As a result, they frequently end up in landfills or incinerators, contributing to pollution and resource depletion.

The environmental impact of battery waste is particularly concerning. Lithium mining, essential for battery production, is an energy-intensive process that consumes vast amounts of water and generates harmful byproducts (Bouri 2015). Additionally, the improper disposal of lithium batteries poses fire and explosion risks, particularly in landfills and waste processing facilities. As the demand for AI-powered devices grows, addressing the battery sustainability crisis will be critical to mitigating AI’s long-term environmental footprint (Zhan, Oldenburg, and Pan 2018).

Bouri, Elie. 2015. “A Broadened Causality in Variance Approach to Assess the Risk Dynamics Between Crude Oil Prices and the Jordanian Stock Market.” Energy Policy 85 (October): 271–79. https://doi.org/10.1016/j.enpol.2015.06.001.

Zhan, Ruiting, Zachary Oldenburg, and Lei Pan. 2018. “Recovery of Active Cathode Materials from Lithium-Ion Batteries Using Froth Flotation.” Sustainable Materials and Technologies 17 (September): e00062. https://doi.org/10.1016/j.susmat.2018.e00062.

Need for Sustainable Design

Addressing the sustainability challenges of disposable AI electronics requires a fundamental shift in design philosophy. Instead of prioritizing cost-cutting and short-term functionality, manufacturers must embed sustainability principles into the development of AI-powered devices. This includes:

• Designing for longevity: AI-powered devices should be built with replaceable components, modular designs, and upgradable software to extend their usability.
• Enabling battery replacement: Consumer and industrial AI devices should incorporate easily swappable batteries rather than sealed enclosures that prevent repair.
• Standardizing repairability: AI hardware should adopt universal standards for repair, ensuring that components can be serviced rather than discarded.
• Developing biodegradable or recyclable materials: Research into eco-friendly circuit boards, biodegradable polymers, and sustainable packaging can help mitigate waste.

Incentives and regulations can also encourage manufacturers to prioritize sustainable AI design. Governments and regulatory bodies can implement right-to-repair laws, extended producer responsibility policies, and e-waste take-back programs to ensure that AI-powered devices are disposed of responsibly. Additionally, consumer awareness campaigns can educate users on responsible e-waste disposal and encourage sustainable purchasing decisions.

The future of AI-powered electronics must be circular rather than linear, ensuring that devices are designed with sustainability in mind and do not contribute disproportionately to the global e-waste crisis. By rethinking design, improving recyclability, and promoting responsible disposal, the industry can mitigate the negative environmental impacts of AI at the edge while still enabling technological progress.

17.7.3 AI Hardware Obsolescence

The concept of planned obsolescence refers to the intentional design of products with artificially limited lifespans, forcing consumers to upgrade or replace them sooner than necessary. While this practice has long been associated with consumer electronics and household appliances, it is increasingly prevalent in AI-powered hardware, from smartphones and wearables to industrial AI sensors and cloud infrastructure. This accelerated replacement cycle not only drives higher consumption and production but also contributes significantly to the growing e-waste crisis (Slade 2007).

Slade, Giles. 2007. Made to Break: Technology and Obsolescence in America. Harvard University Press. https://doi.org/10.4159/9780674043756.

Luna, William Fernando Martı́nez. 2018a. “CONSUMER PROTECTION AGAINST PLANNED OBSOLESCENCE. AN INTERNATIONAL PRIVATE LAW ANALYSIS.” In Planned Obsolescence and the Rulle of Law, 12:229–80. 3. Universidad del Externado de Colombia. https://doi.org/10.2307/j.ctv1ddcwvh.9.

One of the most visible examples of planned obsolescence in AI hardware is the software-driven degradation of device performance. Many manufacturers introduce software updates that, while ostensibly meant to enhance security and functionality, often degrade the performance of older devices. For example, Apple has faced scrutiny for deliberately slowing down older iPhone models via iOS updates (Luna 2018a). While the company claimed that these updates were meant to prevent battery-related shutdowns, critics argued that they pushed consumers toward unnecessary upgrades rather than encouraging repair or battery replacement.

This pattern extends to AI-powered consumer electronics, where firmware updates can render older models incompatible with newer features, effectively forcing users to replace their devices. Many smart home systems, connected appliances, and AI assistants suffer from forced obsolescence due to discontinued cloud support or software services, rendering hardware unusable even when physically intact (Luna 2018b)⁹.

———. 2018b. “CONSUMER PROTECTION AGAINST PLANNED OBSOLESCENCE. AN INTERNATIONAL PRIVATE LAW ANALYSIS.” In Planned Obsolescence and the Rulle of Law, 15:229–80. 2. Universidad del Externado de Colombia. https://doi.org/10.2307/j.ctv1ddcwvh.9.

⁹ Apple was fined €25 million by French regulators in 2020 for intentionally slowing down older iPhone models without informing users, a practice that has sparked global debate on software-induced obsolescence.

Lock-In and Proprietary Components

Another form of planned obsolescence arises from hardware lock-in, where manufacturers deliberately prevent users from repairing or upgrading their devices. Many AI-powered devices feature proprietary components, making it impossible to swap out batteries, upgrade memory, or replace failing parts. Instead of designing for modularity and longevity, manufacturers prioritize sealed enclosures and soldered components, ensuring that even minor failures lead to complete device replacement (Johnson 2018).

Russell, Mark. 2022. “Tech Industry Trends in Hardware Lock-in and Their Sustainability Implications.” Sustainable Computing Journal 10 (1): 34–50.

For example, many AI wearables and smart devices integrate non-replaceable batteries, meaning that when the battery degrades (often in just two to three years), the entire device becomes e-waste. Similarly, smartphones, laptops, and AI-enabled tablets increasingly use soldered RAM and storage, preventing users from upgrading hardware and extending its lifespan (Russell 2022).

Planned obsolescence also affects industrial AI hardware, including AI-powered cameras, factory sensors, and robotics. Many industrial automation systems rely on vendor-locked software ecosystems, where manufacturers discontinue support for older models to push customers toward newer, more expensive replacements. This creates a cycle of forced upgrades, where companies must frequently replace otherwise functional AI hardware simply to maintain software compatibility (Sharma 2020)¹⁰.

Sharma, Amit. 2020. “Industrial AI and Vendor Lock-in: The Hidden Costs of Proprietary Ecosystems.” AI and Industry Review 8 (3): 55–70.

¹⁰ Many industrial AI systems rely on proprietary software ecosystems, where manufacturers discontinue updates and support for older hardware, forcing companies to purchase new equipment to maintain compatibility.

Environmental Cost

Planned obsolescence is not just a financial burden on consumers—it has severe environmental consequences. By shortening product lifespans and discouraging repairability, manufacturers increase the demand for new electronic components, leading to higher resource extraction, energy consumption, and carbon emissions.

The impact of this cycle is particularly concerning given the high environmental cost of semiconductor manufacturing. Producing AI chips, GPUs, and other advanced computing components requires vast amounts of water, rare earth minerals, and energy¹¹. For example, a single 5nm semiconductor fabrication plant consumes millions of gallons of ultrapure water daily and relies on energy-intensive processes that generate significant CO₂ emissions (Mills and Le Hunte 1997; Harris 2023). When AI-powered devices are discarded prematurely, the environmental cost of manufacturing is effectively wasted, amplifying AI’s overall sustainability challenges.

¹¹ Semiconductor fabrication is one of the most resource-intensive manufacturing processes, consuming vast amounts of water and energy while generating hazardous chemical waste.

Mills, Andrew, and Stephen Le Hunte. 1997. “An Overview of Semiconductor Photocatalysis.” Journal of Photochemistry and Photobiology A: Chemistry 108 (1): 1–35. https://doi.org/10.1016/s1010-6030(97)00118-4.

Harris, Michael. 2023. “The Environmental Cost of Next-Generation AI Chips: Energy, Water, and Carbon Impacts.” Journal of Green Computing 17 (1): 22–38.

Puckett, Jim. 2016. E-Waste and the Global Environment: The Hidden Cost of Discarded Electronics. MIT Press.

Additionally, many discarded AI devices contain hazardous materials, including lead, mercury, and brominated flame retardants, which can leach into the environment if not properly recycled (Puckett 2016). The acceleration of AI-powered consumer electronics and industrial hardware turnover will only worsen the global e-waste crisis, further straining waste management and recycling systems.

Extending Hardware Lifespan

Addressing planned obsolescence requires a shift in design philosophy, moving toward repairable, upgradable, and longer-lasting AI hardware. Some potential solutions include:

• Right-to-Repair Legislation: Many governments are considering right-to-repair laws, which would require manufacturers to provide repair manuals, replacement parts, and diagnostic tools for AI-powered devices. This would enable consumers and businesses to extend hardware lifespans rather than replacing entire systems (Johnson 2018).
• Modular AI Hardware: Designing AI-powered devices with modular components—such as replaceable batteries, upgradeable memory, and standardized ports—can significantly reduce electronic waste while improving cost-effectiveness for consumers (Inc. 2022).
• Extended Software Support: Companies should commit to longer software support cycles, ensuring that older AI-powered devices remain functional rather than being rendered obsolete due to artificial compatibility constraints (Brown 2021).
• Consumer Awareness & Circular Economy: Encouraging trade-in and recycling programs, along with consumer education on sustainable AI purchasing, can help shift demand toward repairable and long-lasting devices (Cheshire 2021).

Johnson, Rebecca. 2018. “The Right to Repair Movement and Its Implications for AI Hardware Longevity.” Technology and Society Review 20 (4): 87–102.

Inc., Framework Computer. 2022. “Modular Laptops: A New Approach to Sustainable Computing.”

Brown, Samantha. 2021. “Long-Term Software Support: A Key Factor in Sustainable AI Hardware.” Computer Ethics and Sustainability 14 (2): 112–30.

Cheshire, David. 2021. “Circular Economy and Sustainable AI: Designing Out Waste in the Tech Industry.” In The Handbook to Building a Circular Economy, 48–61. RIBA Publishing. https://doi.org/10.4324/9781003212775-8.

Several tech companies are already experimenting with more sustainable AI hardware. For example, Framework, a startup focused on modular laptops, offers fully repairable, upgradeable systems that prioritize long-term usability over disposable design. Similar efforts in the smartphone and AI-driven IoT sectors could help reduce the environmental footprint of planned obsolescence.

The widespread adoption of AI-powered devices presents a critical opportunity to rethink the lifecycle of electronics. If left unchecked, planned obsolescence will continue to drive wasteful consumption patterns, accelerate e-waste accumulation, and exacerbate the resource extraction crisis. However, with policy interventions, industry innovation, and consumer advocacy, AI hardware can be designed for durability, repairability, and sustainability.

The future of AI should not be disposable. Instead, companies, researchers, and policymakers must prioritize long-term sustainability, ensuring that AI’s environmental footprint is minimized while its benefits are maximized. Addressing planned obsolescence in AI hardware is a key step toward making AI truly sustainable—not just in terms of energy efficiency but in its entire lifecycle, from design to disposal.

17.8 Policy and Regulation

The increasing energy consumption and carbon emissions of AI systems have raised concerns among policymakers, industry leaders, and environmental advocates. As AI adoption accelerates, regulatory frameworks are becoming essential to ensure that AI development and deployment align with global sustainability goals. Without policy intervention, the rapid scaling of AI infrastructure risks exacerbating climate change, resource depletion, and electronic waste generation (Vinuesa et al. 2020).

Vinuesa, Ricardo, Hossein Azizpour, Iolanda Leite, Madeline Balaam, Virginia Dignum, Sami Domisch, Anna Felländer, Simone Daniela Langhans, Max Tegmark, and Francesco Fuso Nerini. 2020. “The Role of Artificial Intelligence in Achieving the Sustainable Development Goals.” Nature Communications 11 (1): 233. https://doi.org/10.1038/s41467-019-14108-y.

Policymakers face a delicate balancing act—on one hand, AI innovation drives economic growth and scientific progress, but on the other, its unchecked expansion could have significant environmental consequences. To address this, policy mechanisms such as measurement and reporting mandates, emission restrictions, financial incentives, and self-regulatory initiatives are being explored worldwide. Government agencies, international organizations, and private sector coalitions are working to establish standardized methodologies for assessing AI’s carbon footprint, encourage efficiency improvements, and promote green AI infrastructure investments.

However, policy fragmentation across regions poses challenges. The European Union, for instance, is leading regulatory efforts through initiatives like the AI Act¹² and sustainability disclosure requirements under the Corporate Sustainability Reporting Directive (CSRD)¹³, while U.S. policymakers have largely relied on voluntary reporting and market-based incentives. China and other nations are taking their own approaches, creating potential barriers to a unified global AI sustainability strategy

¹² The European Commission’s AI Act regulates AI systems’ development, deployment, and use, including provisions for sustainability reporting and energy efficiency requirements.

¹³ The Corporate Sustainability Reporting Directive (CSRD) mandates that large companies disclose environmental and social information, including AI-related emissions and energy consumption.

This section explores the various policy tools available for mitigating AI’s environmental impact, analyzing the role of governments, regulatory bodies, and industry-led efforts. By examining both mandatory and voluntary approaches, we assess how regulations can drive AI sustainability without impeding technological progress.

17.8.1 Measurement and Reporting

A critical first step toward mitigating AI’s environmental impact is accurate measurement and transparent reporting of energy consumption and carbon emissions. Without standardized tracking mechanisms, it is difficult to assess AI’s true sustainability impact or identify areas for improvement. Government regulations and industry initiatives are beginning to mandate energy audits, emissions disclosures, and standardized efficiency metrics for AI workloads. These policies aim to increase transparency, inform better decision-making, and hold organizations accountable for their environmental footprint.

The lack of universally accepted metrics for assessing AI’s environmental impact has been a significant challenge. Current sustainability evaluations often rely on ad hoc reporting by companies, with inconsistent methodologies for measuring energy consumption and emissions. To address this, policymakers and industry leaders are advocating for formalized sustainability benchmarks that assess AI’s carbon footprint at multiple levels. Computational complexity and model efficiency are key factors, as they determine how much computation is required for a given AI task. Data center efficiency, often measured through power usage effectiveness, plays a crucial role in evaluating how much of a data center’s power consumption directly supports computation rather than being lost to cooling and infrastructure overhead. The carbon intensity of energy supply is another critical consideration, as AI operations running on grids powered primarily by fossil fuels have a far greater environmental impact than those powered by renewable energy sources.

Several industry efforts are working toward standardizing sustainability reporting for AI. The MLCommons benchmarking consortium has begun incorporating energy efficiency as a factor in AI model assessments, recognizing the need for standardized comparisons of model energy consumption. Meanwhile, regulatory bodies are pushing for mandatory disclosures. In Europe, the proposed AI Act includes provisions for requiring organizations using AI at scale to report energy consumption and carbon emissions associated with their models. The European Commission has signaled that sustainability reporting requirements for AI may soon be aligned with broader environmental disclosure regulations under the CSRD.

One of the biggest challenges in implementing AI sustainability reporting is balancing transparency with the potential burden on organizations. While greater transparency is essential for accountability, requiring detailed reporting for every AI workload could create excessive overhead, particularly for smaller firms and research institutions. To address this, policymakers are exploring scalable approaches that integrate sustainability considerations into existing industry standards without imposing rigid compliance costs. Developing lightweight reporting mechanisms that leverage existing monitoring tools within data centers and cloud platforms can help ease this burden while still improving visibility into AI’s environmental footprint.

To be most constructive, measurement and reporting policies should focus on enabling continuous refinement rather than imposing simplistic restrictions or rigid caps. Given AI’s rapid evolution, regulations that incorporate flexibility while embedding sustainability into evaluation metrics will be most effective in driving meaningful reductions in energy consumption and emissions. Rather than stifling innovation, well-designed policies can encourage AI developers to prioritize efficiency from the outset, fostering a culture of responsible AI design that aligns with long-term sustainability goals.

17.8.2 Restriction Mechanisms

Beyond measurement and reporting mandates, direct policy interventions can restrict AI’s environmental impact through regulatory limits on energy consumption, emissions, or model scaling. While AI’s rapid growth has spurred innovation, it has also introduced new sustainability challenges that may require governments to impose guardrails to curb excessive environmental costs. Restrictive mechanisms, such as computational caps, conditional access to public resources, financial incentives, and even outright bans on inefficient AI practices, are all potential tools for reducing AI’s carbon footprint. However, their effectiveness depends on careful policy design that balances sustainability with continued technological advancement.

One potential restriction mechanism involves setting limits on the computational power available for training large AI models. The European Commission’s proposed AI Act has explored this concept by introducing economy-wide constraints on AI training workloads. This approach mirrors emissions trading systems (ETS)¹⁴ in environmental policy, where organizations must either operate within predefined energy budgets or procure additional capacity through regulated exchanges. While such limits could help prevent unnecessary computational waste, they also raise concerns about limiting innovation, particularly for researchers and smaller companies that may struggle to access high-performance computing resources (Schwartz et al. 2020).

¹⁴ Emissions trading systems (ETS) are market-based mechanisms that cap the total amount of greenhouse gas emissions allowed within a jurisdiction, with organizations required to purchase or trade emissions allowances to meet their compliance obligations.

Schwartz, Roy, Jesse Dodge, Noah A. Smith, and Oren Etzioni. 2020. “Green AI.” Communications of the ACM 63 (12): 54–63. https://doi.org/10.1145/3381831.

Another policy tool involves conditioning access to public datasets and government-funded computing infrastructure based on model efficiency. AI researchers and developers increasingly rely on large-scale public datasets and subsidized cloud resources to train models. Some have proposed that governments could restrict these resources to AI projects that meet strict energy efficiency criteria. For instance, the MLCommons benchmarking consortium could integrate sustainability metrics into its standardized performance leaderboards, incentivizing organizations to optimize for efficiency alongside accuracy. However, while conditioned access could promote sustainable AI practices, it also risks creating disparities by limiting access to computational resources for those unable to meet predefined efficiency thresholds.

Financial incentives and disincentives represent another regulatory mechanism for driving sustainable AI. Carbon taxes on AI-related compute consumption could discourage excessive model scaling while generating funds for efficiency-focused research. Similar to existing environmental regulations, organizations could be required to pay fees based on the emissions associated with their AI workloads, encouraging them to optimize for lower energy consumption. Conversely, tax credits could reward companies developing efficient AI techniques, fostering investment in greener computing technologies. While financial mechanisms can effectively guide market behavior, they must be carefully calibrated to avoid disproportionately burdening smaller AI developers or discouraging productive use cases.

In extreme cases, outright bans on particularly wasteful AI applications may be considered. If measurement data consistently pinpoints certain AI practices as disproportionately harmful with no feasible path to remediation, governments may choose to prohibit these activities altogether. However, defining harmful AI use cases is challenging due to AI’s dual-use nature, where the same technology can have both beneficial and detrimental applications. Policymakers must approach bans cautiously, ensuring that restrictions target clearly unsustainable practices without stifling broader AI innovation.

Ultimately, restriction mechanisms must strike a careful balance between environmental responsibility and economic growth. Well-designed policies should encourage AI efficiency while preserving the flexibility needed for continued technological progress. By integrating restrictions with incentives and reporting mandates, policymakers can create a comprehensive framework for guiding AI toward a more sustainable future.

17.8.3 Government Incentives

In addition to regulatory restrictions, governments can play a proactive role in advancing sustainable AI development through incentives that encourage energy-efficient practices. Financial support, tax benefits, grants, and strategic investments in Green AI research can drive the adoption of environmentally friendly AI technologies. Unlike punitive restrictions, incentives provide positive reinforcement, making sustainability a competitive advantage rather than a regulatory burden.

One common approach to promoting sustainability is through tax incentives. Governments already offer tax credits for adopting renewable energy sources, such as the U.S. Residential Clean Energy Credit and commercial energy efficiency deductions. Similar programs could be extended to AI companies that optimize their models and infrastructure for lower energy consumption. AI developers who integrate efficiency-enhancing techniques, such as model pruning, quantization, or adaptive scheduling, could qualify for tax reductions, creating a financial incentive for Green AI development.

Beyond tax incentives, direct government funding for sustainable AI research is an emerging strategy. Spain has already committed 300 million euros toward AI projects that explicitly focus on sustainability. Such funding can accelerate breakthroughs in energy-efficient AI by supporting research into novel low-power algorithms, specialized AI hardware, and eco-friendly data center designs. Public-private partnerships can further enhance these efforts, allowing AI companies to collaborate with research institutions and government agencies to pioneer sustainable solutions.

Governments can also incentivize sustainability by integrating Green AI criteria into public procurement policies. Many AI companies provide cloud computing, software services, and AI-driven analytics to government agencies. By mandating that vendors meet sustainability benchmarks—such as operating on carbon-neutral data centers or using energy-efficient AI models—governments can use their purchasing power to set industry-wide standards. Similar policies have already been applied to green building initiatives, where governments require contractors to meet environmental certifications. Applying the same approach to AI could accelerate the adoption of sustainable practices.

Another innovative policy tool is the introduction of carbon credits specifically tailored for AI workloads. Under this system, AI companies could offset emissions by investing in renewable energy projects or carbon capture technologies. AI firms exceeding predefined emissions thresholds would be required to purchase carbon credits, creating a market-based mechanism that naturally incentivizes efficiency. This concept aligns with broader cap-and-trade programs that have successfully reduced emissions in industries like manufacturing and energy production. However, as seen with the challenges surrounding unbundled Energy Attribute Certificates (EACs), carbon credit programs must be carefully structured to ensure genuine emissions reductions rather than allowing companies to simply “buy their way out” of sustainability commitments.

While government incentives offer powerful mechanisms for promoting Green AI, their design and implementation require careful consideration. Incentives should be structured to drive meaningful change without creating loopholes that allow organizations to claim benefits without genuine improvements in sustainability. Additionally, policies must remain flexible enough to accommodate rapid advancements in AI technology. By strategically combining tax incentives, funding programs, procurement policies, and carbon credit systems, governments can create an ecosystem where sustainability is not just a regulatory requirement but an economic advantage.

17.8.4 Self-Regulation

While government policies play a crucial role in shaping sustainable AI practices, the AI industry itself has the power to drive significant environmental improvements through self-regulation. Many leading AI companies and research organizations have already adopted voluntary commitments to reduce their carbon footprints, improve energy efficiency, and promote sustainable development. These efforts can complement regulatory policies and, in some cases, even set higher standards than those mandated by governments.

One of the most visible self-regulation strategies is the commitment by major AI companies to operate on renewable energy. Companies like Google, Microsoft, Amazon, and Meta have pledged to procure enough clean energy to match 100% of their electricity consumption. Google has gone further by aiming for 24/7 Carbon-Free Energy by ensuring that its data centers run exclusively on renewables every hour of every day. These commitments not only reduce operational emissions but also create market demand for renewable energy, accelerating the transition to a greener grid. However, as seen with the use of unbundled EACs, transparency and accountability in renewable energy claims remain critical to ensuring genuine decarbonization rather than superficial offsets.

Another form of self-regulation is the internal adoption of carbon pricing models. Some companies implement shadow pricing, where they assign an internal cost to carbon emissions in financial decision-making. By incorporating these costs into budgeting and investment strategies, AI companies can prioritize energy-efficient infrastructure and low-emission AI models. This approach mirrors broader corporate sustainability efforts in industries like aviation and manufacturing, where internal carbon pricing has proven to be an effective tool for driving emissions reductions.

Beyond energy consumption, AI developers can implement voluntary efficiency checklists that guide sustainable design choices. Organizations like the AI Sustainability Coalition have proposed frameworks that outline best practices for model development, hardware selection, and operational energy management. These checklists can serve as practical tools for AI engineers to integrate sustainability into their workflows. Companies that publicly commit to following these guidelines set an example for the broader industry, demonstrating that sustainability is not just an afterthought but a core design principle.

Independent sustainability audits further enhance accountability by providing third-party evaluations of AI companies’ environmental impact. Firms specializing in technology sustainability, such as Carbon Trust and Green Software Foundation, offer audits that assess energy consumption, carbon emissions, and adherence to green computing best practices. AI companies that voluntarily undergo these audits and publish their findings help build trust with consumers, investors, and regulators. Transparency in environmental reporting allows stakeholders to verify whether companies are meeting their sustainability commitments.

Self-regulation in AI sustainability also extends to open-source collaborations. Initiatives like CodeCarbon and ML \(\textrm{CO}_2\) Impact provide tools that allow developers to estimate and track the carbon footprint of their AI models. By integrating these tools into mainstream AI development platforms like TensorFlow and PyTorch, the industry can normalize sustainability tracking as a standard practice. Encouraging developers to measure and optimize their energy consumption fosters a culture of accountability and continuous improvement.

While self-regulation is an important step toward sustainability, it cannot replace government oversight. Voluntary commitments are only as strong as the incentives driving them, and without external accountability, some companies may prioritize profit over sustainability. However, when combined with regulatory frameworks, self-regulation can accelerate progress by allowing industry leaders to set higher standards than those mandated by law. By embedding sustainability into corporate strategy, AI companies can demonstrate that technological advancement and environmental responsibility are not mutually exclusive.

17.8.5 Global Impact

While AI sustainability efforts are gaining traction, they remain fragmented across national policies, industry initiatives, and regional energy infrastructures. AI’s environmental footprint is inherently global, spanning supply chains, cloud data centers, and international markets. A lack of coordination between governments and corporations risks inefficiencies, contradictory regulations, and loopholes that allow companies to shift environmental burdens rather than genuinely reduce them. Establishing global frameworks for AI sustainability is therefore crucial for aligning policies, ensuring accountability, and fostering meaningful progress in mitigating AI’s environmental impact.

One of the primary challenges in global AI sustainability efforts is regulatory divergence. Countries and regions are taking vastly different approaches to AI governance. The European Union’s AI Act, for example, introduces comprehensive risk-based regulations that include provisions for energy efficiency and environmental impact assessments for AI systems. By contrast, the United States has largely adopted a market-driven approach, emphasizing corporate self-regulation and voluntary sustainability commitments rather than enforceable mandates. Meanwhile, China has prioritized AI dominance through heavy government investment, with sustainability playing a secondary role to technological leadership. This regulatory patchwork creates inconsistencies in how AI-related emissions, resource consumption, and energy efficiency are tracked and managed.

One proposed solution to this fragmentation is the standardization of sustainability reporting metrics for AI systems. Organizations such as the OECD, IEEE, and United Nations have pushed for unified environmental impact reporting standards similar to financial disclosure frameworks. This would allow companies to track and compare their carbon footprints, energy usage, and resource consumption using common methodologies. The adoption of LCA standards for AI—as seen in broader environmental accounting—would enable more accurate assessments of AI’s total environmental impact, from hardware manufacturing to deployment and decommissioning.

Beyond reporting, energy grid decarbonization remains a critical global consideration. The sustainability of AI is heavily influenced by the carbon intensity of electricity in different regions. For example, training a large AI model in a coal-powered region like Poland results in significantly higher carbon emissions than training the same model in hydroelectric-powered Norway. However, market-based energy accounting practices—such as purchasing unbundled EACs—have allowed some companies to claim carbon neutrality despite operating in high-emission grids. This has led to concerns that sustainability claims may not always reflect actual emissions reductions but instead rely on financial instruments that shift carbon responsibility rather than eliminating it. As a response, Google has championed 24/7 Carbon-Free Energy (CFE), which aims to match local energy consumption with renewable sources in real-time rather than relying on distant offsets. If widely adopted, this model could become a global benchmark for AI sustainability accounting.

Another key area of global concern is AI hardware supply chains and electronic waste management. The production of AI accelerators, GPUs, and data center hardware depends on a complex network of raw material extraction, semiconductor fabrication, and electronic assembly spanning multiple continents. The environmental impact of this supply chain—ranging from rare-earth mineral mining in Africa to chip manufacturing in Taiwan and final assembly in China—often falls outside the jurisdiction of AI companies themselves. This underscores the need for international agreements on sustainable semiconductor production, responsible mining practices, and e-waste recycling policies.

The Basel Convention¹⁵, which regulates hazardous waste exports, could provide a model for addressing AI-related e-waste challenges at a global scale. The convention restricts the transfer of toxic electronic waste from developed nations to developing countries, where unsafe recycling practices can harm workers and pollute local ecosystems. Expanding such agreements to cover AI-specific hardware components, such as GPUs and inference chips, could ensure that end-of-life disposal is handled responsibly rather than outsourced to regions with weaker environmental protections.

¹⁵ Basel Convention: An international treaty regulating the transboundary movement of hazardous waste to prevent its disposal in countries with weaker environmental protections.

International collaboration in AI sustainability is not just about mitigating harm but also leveraging AI as a tool for environmental progress. AI models are already being deployed for climate forecasting, renewable energy optimization, and precision agriculture, demonstrating their potential to contribute to global sustainability goals. Governments, research institutions, and industry leaders must align on best practices for scaling AI solutions that support climate action, ensuring that AI is not merely a sustainability challenge but also a powerful tool for global environmental resilience.

Ultimately, sustainable AI requires a coordinated global approach that integrates regulatory alignment, standardized sustainability reporting, energy decarbonization, supply chain accountability, and responsible e-waste management. Without such collaboration, regional disparities in AI governance could hinder meaningful progress, allowing inefficiencies and externalized environmental costs to persist. As AI continues to evolve, establishing global frameworks that balance technological advancement with environmental responsibility will be critical in shaping an AI-driven future that is not only intelligent but also sustainable.

17.9 Public Engagement

As artificial intelligence (AI) becomes increasingly intertwined with efforts to address environmental challenges, public perception plays a pivotal role in shaping its adoption, regulation, and long-term societal impact. While AI is often viewed as a powerful tool for advancing sustainability—through applications such as smart energy management, climate modeling, and conservation efforts—it also faces scrutiny over its environmental footprint, ethical concerns, and transparency.

Public discourse surrounding AI and sustainability is often polarized. On one side, AI is heralded as a transformative force capable of accelerating climate action, reducing carbon emissions, and optimizing resource use. On the other, concerns persist about the high energy consumption of AI models, the potential for unintended environmental consequences, and the opaque nature of AI-driven decision-making. These contrasting viewpoints influence policy development, funding priorities, and societal acceptance of AI-driven sustainability initiatives.

Bridging the gap between AI researchers, policymakers, and the public is essential for ensuring that AI’s contributions to sustainability are both scientifically grounded and socially responsible. This requires clear communication about AI’s capabilities and limitations, greater transparency in AI decision-making processes, and mechanisms for inclusive public participation. Without informed public engagement, misunderstandings and skepticism could hinder the adoption of AI solutions that have the potential to drive meaningful environmental progress.

17.9.1 AI Awareness

Public understanding of AI and its role in sustainability remains limited, often shaped by media narratives that highlight either its transformative potential or its risks. Surveys such as the Pew Research Center poll found that while a majority of people have heard of AI, their understanding of its specific applications—particularly in sustainability—remains shallow. Many associate AI with automation, recommendation systems, or chatbots but may not be aware of its broader implications in climate science, energy optimization, and environmental monitoring.

A key factor influencing public perception is the framing of AI’s sustainability contributions. Optimistic portrayals emphasize AI’s ability to enhance renewable energy integration, improve climate modeling accuracy, and enable smart infrastructure for reduced emissions. Organizations such as Climate Change AI actively promote AI’s potential in environmental applications, fostering a positive narrative. Conversely, concerns about AI’s energy-intensive training processes, ethical considerations, and potential biases contribute to skepticism. Studies analyzing public discourse on AI sustainability reveal an even split between optimism and caution, with some fearing that AI’s environmental costs may outweigh its benefits.

In many cases, public attitudes toward AI-driven sustainability efforts are shaped by trust in institutions. AI systems deployed by reputable environmental organizations or in collaboration with scientific communities tend to receive more favorable reception. However, corporate-led AI sustainability initiatives often face skepticism, particularly if they are perceived as greenwashing—a practice where companies exaggerate their commitment to environmental responsibility without substantial action.

To foster informed public engagement, increasing AI literacy is crucial. This involves education on AI’s actual energy consumption, potential for optimization, and real-world applications in sustainability. Universities, research institutions, and industry leaders can play a pivotal role in making AI’s sustainability impact more accessible to the general public through open reports, interactive tools, and clear communication strategies.

17.9.2 Messaging and Discourse

How AI is communicated to the public significantly influences perceptions of its role in sustainability. The messaging around AI-driven environmental efforts must balance technical accuracy, realistic expectations, and transparency to ensure constructive discourse.

Optimistic narratives emphasize AI’s potential as a powerful tool for sustainability. Initiatives such as Climate Change AI and AI-driven conservation projects highlight applications in wildlife protection, climate modeling, energy efficiency, and pollution monitoring. These examples are often framed as AI augmenting human capabilities, enabling more precise and scalable solutions to environmental challenges. Such positive framing encourages public support and investment in AI-driven sustainability research.

However, skepticism remains, particularly regarding AI’s own environmental footprint. Critical perspectives highlight the massive energy demands of AI model training, particularly for large-scale neural networks. The Asilomar AI Principles and other cautionary frameworks stress the need for transparency, ethical guardrails, and energy-conscious AI development. The rise of generative AI models has further amplified concerns about data center energy consumption, supply chain sustainability, and the long-term viability of compute-intensive AI workloads.

A key challenge in AI sustainability messaging is avoiding extremes. Public discourse often falls into two polarized views: one where AI is seen as an indispensable tool for solving climate change, and another where AI is portrayed as an unchecked technology accelerating ecological harm. Neither view fully captures the nuanced reality. AI, like any technology, is a tool whose environmental impact depends on how it is developed, deployed, and governed.

To build public trust and engagement, AI sustainability messaging should prioritize three key aspects. First, it must acknowledge clear trade-offs by presenting both the benefits and limitations of AI for sustainability, including energy consumption, data biases, and real-world deployment challenges. Second, messaging should rely on evidence-based claims, communicating AI’s impact through data-driven assessments, lifecycle analyses, and transparent carbon accounting rather than speculative promises. Third, the framing should remain human-centered, emphasizing collaborative AI systems that work alongside scientists, policymakers and communities rather than fully automated, opaque decision-making systems. Through this balanced, transparent approach, AI can maintain credibility while driving meaningful environmental progress.

Effective public engagement relies on bridging the knowledge gap between AI practitioners and non-experts, ensuring that AI’s role in sustainability is grounded in reality, openly discussed, and continuously evaluated.

17.9.3 Transparency and Trust

As AI systems become more integrated into sustainability efforts, transparency and trust are crucial for ensuring public confidence in their deployment. The complexity of AI models, particularly those used in environmental monitoring, resource optimization, and emissions tracking, often makes it difficult for stakeholders to understand how decisions are being made. Without clear explanations of how AI systems operate, concerns about bias, accountability, and unintended consequences can undermine public trust.

A key aspect of transparency involves ensuring that AI models used in sustainability applications are explainable and interpretable. The National Institute of Standards and Technology (NIST) Principles for Explainable AI provide a framework for designing systems that offer meaningful and understandable explanations of their outputs. These principles emphasize that AI-generated decisions should be contextually relevant, accurately reflect the model’s logic, and clearly communicate the limitations of the system (Phillips et al. 2020). In sustainability applications, where AI influences environmental policy, conservation strategies, and energy management, interpretability is essential for public accountability.

Phillips, P. Jonathon, Carina A. Hahn, Peter C. Fontana, David A. Broniatowski, and Mark A. Przybocki. 2020. “Four Principles of Explainable Artificial Intelligence.” Gaithersburg, Maryland. National Institute of Standards; Technology (NIST). https://doi.org/10.6028/nist.ir.8312-draft.

Transparency is also necessary in AI sustainability claims. Many technology companies promote AI-driven sustainability initiatives, yet without standardized reporting, it is difficult to verify the actual impact. The Montréal Carbon Pledge offers a valuable framework for accountability in this space:

“As institutional investors, we must act in the best long-term interests of our beneficiaries. In this fiduciary role, long-term investment risks are associated with greenhouse gas emissions, climate change, and carbon regulation. Measuring our carbon footprint is integral to understanding better, quantifying, and managing the carbon and climate change-related impacts, risks, and opportunities in our investments. Therefore, as a first step, we commit to measuring and disclosing the carbon footprint of our investments annually to use this information to develop an engagement strategy and identify and set carbon footprint reduction targets.” — Montréal Carbon Pledge

This commitment to measuring and disclosing carbon footprints serves as a model for how AI sustainability claims could be validated. A similar commitment for AI, where companies disclose the environmental footprint of training and deploying models, would provide the public with a clearer picture of AI’s sustainability contributions. Without such measures, companies risk accusations of “greenwashing,” where claims of sustainability benefits are exaggerated or misleading.

Beyond corporate accountability, transparency in AI governance ensures that AI systems deployed for sustainability are subject to ethical oversight. The integration of AI into environmental decision-making raises questions about who has control over these technologies and how they align with societal values. Efforts such as the OECD AI Policy Observatory highlight the need for regulatory frameworks that require AI developers to disclose energy consumption, data sources, and model biases when deploying AI in critical sustainability applications. Public accessibility to this information would enable greater scrutiny and foster trust in AI-driven solutions.

Building trust in AI for sustainability requires not only clear explanations of how models function but also proactive efforts to include stakeholders in decision-making processes. Transparency mechanisms such as open-access datasets, public AI audits, and participatory model development can enhance accountability. By ensuring that AI applications in sustainability remain understandable, verifiable, and ethically governed, trust can be established, enabling broader public support for AI-driven environmental solutions.

17.9.4 Engagement and Awareness

Public engagement plays a crucial role in shaping the adoption and effectiveness of AI-driven sustainability efforts. While AI has the potential to drive significant environmental benefits, its success depends on how well the public understands and supports its applications. Widespread misconceptions, limited awareness of AI’s role in sustainability, and concerns about ethical and environmental risks can hinder meaningful engagement. Addressing these issues requires deliberate efforts to educate, involve, and empower diverse communities in discussions about AI’s impact on environmental sustainability.

Surveys indicate that while AI is widely recognized, the specific ways it intersects with sustainability remain unclear to the general public. A study conducted by the Pew Research Center found that while 87% of respondents had some awareness of AI, only a small fraction could explain how it affects energy consumption, emissions, or conservation efforts. This gap in understanding can lead to skepticism, with some viewing AI as a potential contributor to environmental harm due to its high computational demands rather than as a tool for addressing climate challenges. To build public confidence in AI sustainability initiatives, clear communication is essential.

Efforts to improve AI literacy in sustainability contexts can take multiple forms. Educational campaigns highlighting AI’s role in optimizing renewable energy grids, reducing food waste, or monitoring biodiversity can help demystify the technology. Programs such as Climate Change AI and Partnership on AI actively work to bridge this gap by providing accessible research, case studies, and policy recommendations that illustrate AI’s benefits in addressing climate change. Similarly, media representation plays a significant role in shaping perceptions, and responsible reporting on AI’s environmental potential—alongside its challenges—can provide a more balanced narrative.

Beyond education, engagement requires active participation from various stakeholders, including local communities, environmental groups, and policymakers. Many AI-driven sustainability projects focus on data collection and automation but lack mechanisms for involving affected communities in decision-making. For example, AI models used in water conservation or wildfire prediction may rely on data that overlooks the lived experiences of local populations. Creating channels for participatory AI design—where communities contribute insights, validate model outputs, and influence policy—can lead to more inclusive and context-aware sustainability solutions.

Transparency and public input are particularly important when AI decisions affect resource allocation, environmental justice, or regulatory actions. AI-driven carbon credit markets, for instance, require mechanisms to ensure that communities in developing regions benefit from sustainability initiatives rather than facing unintended harms such as land displacement or exploitation. Public consultations, open-data platforms, and independent AI ethics committees can help integrate societal values into AI-driven sustainability policies.

Ultimately, fostering public engagement and awareness in AI sustainability requires a multi-faceted approach that combines education, communication, and participatory governance. By ensuring that AI systems are accessible, understandable, and responsive to community needs, public trust and support for AI-driven sustainability solutions can be strengthened. This engagement is essential to aligning AI innovation with societal priorities and ensuring that environmental AI systems serve the broader public good.

17.9.5 Equitable AI Access

Ensuring equitable access to AI-driven sustainability solutions is essential for fostering global environmental progress. While AI has demonstrated its ability to optimize energy grids, monitor deforestation, and improve climate modeling, access to these technologies remains unevenly distributed. Developing nations, marginalized communities, and small-scale environmental organizations often lack the infrastructure, funding, and expertise necessary to leverage AI effectively. Addressing these disparities is crucial to ensuring that the benefits of AI sustainability solutions reach all populations rather than exacerbating existing environmental and socio-economic inequalities.

One of the primary barriers to equitable AI access is the digital divide. Many AI sustainability applications rely on advanced computing infrastructure, cloud resources, and high-quality datasets, which are predominantly concentrated in high-income regions. A recent OECD report on national AI compute capacity highlighted that many countries lack a strategic roadmap for developing AI infrastructure, leading to a growing gap between AI-rich and AI-poor regions (Oecd 2023). Without targeted investment in AI infrastructure, lower-income countries remain excluded from AI-driven sustainability advancements. Expanding access to computing resources, supporting open-source AI frameworks, and providing cloud-based AI solutions for environmental monitoring could help bridge this gap.

Oecd. 2023. “A Blueprint for Building National Compute Capacity for Artificial Intelligence.” 350. Organisation for Economic Co-Operation; Development (OECD). https://doi.org/10.1787/876367e3-en.

In addition to infrastructure limitations, a lack of high-quality, region-specific data poses a significant challenge. AI models trained on datasets from industrialized nations may not generalize well to other geographic and socio-economic contexts. For example, an AI model optimized for water conservation in North America may be ineffective in regions facing different climate patterns, agricultural practices, or regulatory structures. Efforts to localize AI sustainability applications—by collecting diverse datasets, partnering with local organizations, and integrating indigenous knowledge—can enhance the relevance and impact of AI solutions in underrepresented regions.

Access to AI tools also requires technical literacy and capacity-building initiatives. Many small environmental organizations and community-driven sustainability projects do not have the in-house expertise needed to develop or deploy AI solutions effectively. Capacity-building efforts, such as AI training programs, knowledge-sharing networks, and collaborations between academic institutions and environmental groups, can empower local stakeholders to adopt AI-driven sustainability practices. Organizations like Climate Change AI and the Partnership on AI have taken steps to provide resources and guidance on using AI for environmental applications, but more widespread efforts are needed to democratize access.

Funding mechanisms also play a critical role in determining who benefits from AI-driven sustainability. While large corporations and well-funded research institutions can afford to invest in AI-powered environmental solutions, smaller organizations often lack the necessary financial resources. Government grants, philanthropic funding, and international AI-for-good initiatives could help ensure that grassroots sustainability efforts can leverage AI technologies. For instance, Spain has allocated 300 million euros specifically for AI and sustainability projects, setting a precedent for public investment in environmentally responsible AI innovation. Expanding such funding models globally could foster more inclusive AI adoption.

Beyond technical and financial barriers, policy interventions are necessary to ensure that AI sustainability efforts are equitably distributed. Without regulatory frameworks that prioritize inclusion, AI-driven environmental solutions may disproportionately benefit regions with existing technological advantages while neglecting areas with the most pressing sustainability challenges. Governments and international bodies should establish policies that encourage equitable AI adoption, such as requiring AI sustainability projects to consider social impact assessments or mandating transparent reporting on AI-driven environmental initiatives.

Ensuring equitable access to AI for sustainability is not merely a technical challenge but a fundamental issue of environmental justice. As AI continues to shape global sustainability efforts, proactive measures must be taken to prevent technology from reinforcing existing inequalities. By investing in AI infrastructure, localizing AI applications, supporting capacity-building efforts, and implementing inclusive policies, AI can become a tool that empowers all communities in the fight against climate change and environmental degradation.

17.10 Future Challenges

As AI continues to evolve, its role in environmental sustainability is set to expand. Advances in AI have the potential to accelerate progress in renewable energy, climate modeling, biodiversity conservation, and resource efficiency. However, realizing this potential requires addressing significant challenges related to energy efficiency, infrastructure sustainability, data availability, and governance. The future of AI and sustainability hinges on balancing innovation with responsible environmental stewardship, ensuring that AI-driven progress does not come at the cost of increased environmental degradation.

17.10.1 Future Directions

A major priority in AI sustainability is the development of more energy-efficient models and algorithms. Optimizing deep learning models to minimize computational cost is a key research direction, with techniques such as model pruning, quantization, and low-precision numerics demonstrating significant potential for reducing energy consumption without compromising performance. These strategies aim to improve the efficiency of AI workloads while leveraging specialized hardware accelerators to maximize computational throughput with minimal energy expenditure. The continued development of non-von Neumann computing paradigms, such as neuromorphic computing and in-memory computing, presents another avenue for energy-efficient AI architectures, as explored in the Hardware Acceleration chapter.

Another crucial direction involves the integration of renewable energy into AI infrastructure. Given that data centers are among the largest contributors to AI’s carbon footprint, shifting towards clean energy sources like solar, wind, and hydroelectric power is imperative. The feasibility of this transition depends on advancements in sustainable energy storage technologies, such as those being developed by companies like Ambri, an MIT spinoff working on liquid metal battery solutions. These innovations could enable data centers to operate on renewable energy with greater reliability, reducing dependency on fossil fuel-based grid power. However, achieving this transition at scale requires collaborative efforts between AI companies, energy providers, and policymakers to develop grid-aware AI scheduling and carbon-aware workload management strategies, ensuring that compute-intensive AI tasks are performed when renewable energy availability is at its peak.

Beyond energy efficiency, AI sustainability will also benefit from intelligent resource allocation and waste reduction strategies. Improving the utilization of computing resources, reducing redundant model training cycles, and implementing efficient data sampling techniques can substantially decrease energy consumption. A key challenge in AI model development is the trade-off between experimentation and efficiency—techniques such as neural architecture search and hyperparameter optimization can improve model performance but often require vast computational resources. Research into efficient experimentation methodologies could help strike a balance, allowing for model improvements while mitigating the environmental impact of excessive training runs.

17.10.2 Challenges

Despite these promising directions, significant obstacles must be addressed to make AI truly sustainable. One of the most pressing challenges is the lack of standardized measurement and reporting frameworks for evaluating AI’s environmental footprint. Unlike traditional industries, where LCA methodologies are well-established, AI systems require more comprehensive and adaptable approaches that account for the full environmental impact of both hardware (compute infrastructure) and software (model training and inference cycles). While efforts such as MLCommons have begun integrating energy efficiency into benchmarking practices, a broader, globally recognized standard is necessary to ensure consistency in reporting AI-related emissions.

Another critical challenge is optimizing AI infrastructure for longevity and sustainability. AI accelerators and data center hardware must be designed with maximized utilization, extended operational lifespans, and minimal environmental impact in mind. Unlike conventional hardware refresh cycles, which often prioritize performance gains over sustainability, future AI infrastructure must prioritize reusability, modular design, and circular economy principles to minimize electronic waste and reduce reliance on rare earth materials.

From a software perspective, minimizing redundant computation is essential to reducing energy-intensive workloads. The practice of training larger models on increasingly vast datasets, while beneficial for accuracy, comes with diminishing returns in sustainability. A data-centric approach to AI model development, as highlighted in recent work (Wu et al. 2022), suggests that the predictive value of data decays over time, making it crucial to identify and filter the most relevant data subsets. Smarter data sampling strategies can optimize training processes, ensuring that only the most informative data is used to refine models, reducing the energy footprint without sacrificing model quality.

Wu, Carole-Jean, Ramya Raghavendra, Udit Gupta, Bilge Acun, Newsha Ardalani, Kiwan Maeng, Gloria Chang, et al. 2022. “Sustainable Ai: Environmental Implications, Challenges and Opportunities.” Proceedings of Machine Learning and Systems 4: 795–813.

A further challenge lies in data accessibility and transparency. Many AI sustainability efforts rely on corporate and governmental disclosures of energy usage, carbon emissions, and environmental impact data. However, data gaps and inconsistencies hinder efforts to accurately assess AI’s footprint. Greater transparency from AI companies regarding their sustainability initiatives, coupled with open-access datasets for environmental impact research, would enable more rigorous analysis and inform best practices for sustainable AI development.

Finally, the rapid pace of AI innovation poses challenges for regulation and governance. Policymakers must develop agile, forward-looking policies that promote sustainability while preserving the flexibility needed for AI research and innovation. Regulatory frameworks should encourage efficient AI practices, such as promoting carbon-aware computing, incentivizing energy-efficient AI model development, and ensuring that AI-driven environmental applications align with broader sustainability goals. Achieving this requires close collaboration between AI researchers, environmental scientists, energy sector stakeholders, and policymakers to develop a regulatory landscape that fosters responsible AI growth while minimizing ecological harm.

17.10.3 Towards Sustainable AI

The future of AI in sustainability is both promising and fraught with challenges. To harness AI’s full potential while mitigating its environmental impact, the field must embrace energy-efficient model development, renewable energy integration, hardware and software optimizations, and transparent environmental reporting. Addressing these challenges will require multidisciplinary collaboration across technical, industrial, and policy domains, ensuring that AI’s trajectory aligns with global sustainability efforts.

By embedding sustainability principles into AI system design, optimizing compute infrastructure, and establishing clear accountability mechanisms, AI can serve as a catalyst for environmental progress rather than a contributor to ecological degradation. The coming years will be pivotal in shaping AI’s role in sustainability, determining whether it amplifies existing challenges or emerges as a key tool in the fight against climate change and resource depletion.

17.11 Conclusion

The integration of AI into environmental sustainability presents both immense opportunities and formidable challenges. As AI systems continue to scale in complexity and influence, their environmental footprint must be addressed through energy-efficient design, responsible infrastructure deployment, transparent accountability measures, and policy-driven interventions. While AI offers powerful capabilities for climate modeling, emissions reduction, resource optimization, and biodiversity conservation, its reliance on compute-intensive hardware, large-scale data processing, and energy-hungry model training necessitates a careful balance between progress and sustainability.

This chapter has explored the full lifecycle impact of AI systems, from their carbon footprint and energy consumption to hardware manufacturing, e-waste concerns, and the role of embedded AI in the growing “Internet of Trash.” We have examined strategies for mitigating AI’s environmental impact, including advances in green AI infrastructure, energy-aware model optimization, and lifecycle-aware AI development. Additionally, we have highlighted the importance of policy and regulatory frameworks in shaping a sustainable AI ecosystem, emphasizing the need for measurement and reporting mandates, incentive structures, and governance mechanisms that align AI innovation with long-term environmental goals.

Public perception and engagement remain central to the discourse on AI and sustainability. Transparent AI practices, explainable models, and ethical governance frameworks will be key to fostering trust and ensuring that AI solutions are inclusive, equitable, and accountable. The responsible deployment of AI in sustainability efforts must incorporate stakeholder input, interdisciplinary collaboration, and a commitment to minimizing unintended consequences.

Looking ahead, the path toward sustainable AI requires continuous advancements in hardware efficiency, carbon-aware computing, renewable energy integration, and equitable access to AI resources. Overcoming challenges such as data gaps, inconsistent environmental reporting, and planned obsolescence in AI hardware will require collective efforts from AI researchers, environmental scientists, policymakers, and industry leaders. By embedding sustainability at the core of AI development, we can ensure that AI not only accelerates technological progress but also contributes meaningfully to a more sustainable and resilient future.

AI has the potential to be a force for good in the fight against climate change and resource depletion, but its long-term impact depends on the choices we make today. Through innovation, regulation, and collective responsibility, AI can evolve as a technology that enhances environmental sustainability rather than exacerbating ecological strain. The decisions made by AI practitioners, policymakers, and society at large will shape whether AI serves as a tool for sustainable progress or an unchecked driver of environmental harm. The imperative now is to act deliberately, designing AI systems that align with global sustainability goals and contribute to a future where technological advancement and ecological well-being coexist harmoniously.