Interesting read;

The adoption of artificial intelligence has been rapidly accelerating across all parts of society, bringing the potential to address shared global challenges such as climate change and drought mitigation. Yet underlying the excitement surrounding AI’s transformative potential are increasingly large and energy-intensive deep neural networks. And the growing demands of these complex models are raising concerns about AI’s environmental impact.

Importantly, beyond their global climate impact, the environmental effects of AI have significant implications at the local and regional levels. While recent initiatives present promising steps for sustainable AI, they often prioritize easily measurable environmental metrics such as the total amount of carbon emissions and water consumption. They do not give enough attention to environmental equity — the imperative that AI’s environmental costs be equitably distributed across different regions and communities.

The escalating and localized environmental costs of AI

Even putting aside the environmental toll of chip manufacturing and supply chains, the training process for a single AI model, such as a large language model, can consume thousands of megawatt hours of electricity and emit hundreds of tons of carbon. This is roughly equivalent to the annual carbon emissions of hundreds of households in America. Furthermore, AI model training can lead to the evaporation of an astonishing amount of fresh water into the atmosphere for data center heat rejection, potentially exacerbating stress on our already limited freshwater resources.

All these environmental impacts are expected to escalate considerably, with the global AI energy demand projected to exponentially increase to at least 10 times the current level and exceed the annual electricity consumption of a small country like Belgium by 2026. In the United States, the rapidly growing AI demand is poised to drive data center energy consumption to about 6% of the nation’s total electricity usage in 2026, adding further pressure on grid infrastructures and highlighting the urgent need for sustainable solutions to support continued AI advancement.

The generation of electricity, particularly through fossil fuel combustion, results in local air pollution, thermal pollution in water bodies, and the production of solid wastes, including even hazardous materials. Elevated carbon emissions in a region come with localized social costs, potentially leading to higher levels of ozone, particulate matter, and premature mortality. Furthermore, the strain on local freshwater resources imposed by the substantial water consumption associated with AI, both directly for onsite server cooling and indirectly for offsite electricity generation, can worsen prolonged droughts in water-stressed regions like Arizona and Chile.

Initiatives to advance AI’s environmental sustainability

In the era of heightened environmental consciousness, a variety of multifaceted initiatives have been gaining traction, aiming to advance AI’s environmental sustainability and ensure its positive net contribution to mitigating climate change.

Advancements in data center power and cooling infrastructures have made substantial strides in reducing the once-hefty energy overheads of AI computing, as evidenced by the decline in power usage effectiveness (PUE) from 2.0 to 1.1 or even lower in cutting-edge data center facilities.

Other key innovations include the design of efficient AI model architectures, optimization algorithms to accelerate AI training and inference, techniques like weight pruning and quantization to reduce model sizes, and the creation of energy-efficient GPUs and accelerators.

At the system level, holistic management of both computing and non-computing resources is essential for building sustainable AI systems. For instance, geographical load balancing, a well-established technique, can dynamically align energy demand with real-time grid operating conditions and carbon intensities across a network of distributed data centers. Its effectiveness in mitigating environmental impact has been demonstrated in real-world systems, such as Google’s carbon-intelligent computing platform.

Furthermore, data center operators have pursued diverse strategies to achieve “net zero” emissions, including the development of large-scale solar farms and procurement of renewable energy credits. Likewise, in acknowledgment of fresh water as a vital social resource, industry leaders have set an ambitious goal of “water positive” by 2030 by replenishing more water than they consume.

The troubling rise of AI’s environmental inequity

Unfortunately, there remains a widening disparity in how different regions and communities are affected by AI’s environmental impacts. In many cases, adverse environmental impacts of AI disproportionately burden communities and regions that are particularly vulnerable to the resulting environmental harms. For instance, in 2022, Google operated its data center in Finland on 97% carbon-free energy; that number drops to 4–18% for its data centers in Asia. This highlights a significant disparity in the local consumption of fossil fuels and the creation of air pollution. Similarly, the water-consumption rate for data center heat rejection can be disproportionately higher in drought-stricken regions such as Arizona due to their hotter climates.

Moreover, existing approaches to deploying and managing AI computing often exacerbate environmental inequity, which is compounded by persistent socioeconomic disparities between regions. For instance, geographical load balancing that prioritizes the total energy costs or carbon footprint may inadvertently increase the water footprint of data centers in water-stressed regions, further straining local freshwater resources. It could also disproportionately add to the grid congestion and raise locational marginal prices for electricity, potentially leading to increased utility rates and unfairly burdening local residents with higher energy costs.

A strategic imperative to address AI’s environmental inequity

The troubling rise of AI’s environmental inequality has impeded progress toward environmentally responsible AI. This issue has garnered increasing public attention, prompting urgent calls for mitigation efforts. For instance, in its 2023 Landscape report, the AI Now Institute draws parallels between the uneven regional distribution of AI’s environmental costs and historical practices of settler colonialism and racial capitalism. Additionally, UNESCO advises against using AI if it leads to disproportionate negative environmental impacts, while the State of California acknowledges the importance of equitably distributing AI’s environmental costs in its recent Benefits and Risks of Generative Artificial Intelligence report.

AI’s emerging environmental inequality across different regions is deeply intertwined with global socioeconomic disparities. Therefore, taking proactive steps to address this issue is not just a moral necessity but a strategic imperative for minimizing potential business risks and narrowing the widening socioeconomic gap.

We advocate environmental equity as a vital component for fostering truly responsible AI. Our goal is to address AI’s environmental impact in the most adversely affected and vulnerable regions as a priority, mitigating the widening disparity between different regions.

Promoting environmentally equitable AI via geographical load balancing

Admittedly, achieving uniform environmental impact for AI data centers across all geographical locations is challenging, as some regions inherently face greater environmental disadvantages than others.

However, the ability to flexibly deploy and manage AI computing across a network of geographically distributed data centers offers substantial opportunities to tackle AI’s environmental inequality by prioritizing disadvantaged regions and equitably distributing the overall negative environmental impact. While redirecting car traffic on a freeway to address the associated air pollution presents logistical hurdles, using geographical load balancing to spatially reroute “AI traffic,” including AI training tasks and certain inference requests, is easy and immediately actionable. This technique enables dynamic redistribution of AI computing across geographically distributed data centers, considering real-time local factors such as the proportion of fossil-based energy sources and water efficiency. Crucially, it applies regardless of whether a company operates its own geographically distributed data center infrastructures or uses cloud-based AI services. For instance, a major technology company could easily move most of its computing tasks to other data centers without impacting its services, while a small business could dynamically and flexibly shift its computing needs across different cloud regions.

By fairly distributing AI computing across different data centers, we can effectively redistribute AI’s regional environmental costs, fostering a more equitable outcome. The key innovation centers on prioritizing regions bearing the most notable adverse environmental impact, such as heightened air pollution and depleting freshwater resources. This is accomplished by dynamically assigning a higher importance weight to otherwise disproportionately burdened regions when optimizing the geographical load balancing decisions in real time.

However, implementing equitable-aware geographical load balancing presents some practical challenges. Even when dynamically balancing AI computing in real time, it’s difficult (if not impossible) to predict AI’s future computing demands to account for the long-term environmental impact. And ensuring a consistent level of AI model inference performance and quality is essential. To tackle these challenges, we can utilize the history data and leverage state-of-the-art reinforcement learning techniques to identify the optimal AI computing management across geographically distributed data centers, prioritizing environmental equity while minimizing the overall environmental impact.

• • •

The uneven distribution of AI’s environmental impacts often remains hidden from public view and can potentially create unforeseen socioeconomic ramifications. We aim to raise awareness within the business community and among the general public regarding AI’s emerging environmental inequity. As we strive to develop environmentally responsible AI, we must not solely concentrate on easily measurable sustainability metrics, such as the total carbon emission and water consumption, while overlooking equity in the process. AI’s environmental impacts must resonate with the priorities and interests of local regions.

Authors:

• Shaolei Ren is an associate professor of electrical and computer engineering at the University of California, Riverside. He has taught and researched computational sustainability for more than a decade. His work on sustainable AI has been featured in many international AI governance and ethics guidelines, contributed to K-12 education materials, led to industry innovations like real-time water footprint reporting tools, and gained worldwide media coverage.

• Adam Wierman is the Carl F. Braun Professor in the Department of Computing and Mathematical Sciences at Caltech. His research strives to make the networked systems that govern our world sustainable and resilient. He is best known for his work spearheading the design of algorithms for sustainable data centers, which has seen significant industry adoption, as well as his work on heavy tails, including his coauthored book, The Fundamentals of Heavy Tails.