The growing adoption of artificial intelligence (AI) has been accelerating across all parts of society, boosting productivity and addressing pressing global challenges such as climate change. Nonetheless, the technological advancement of AI relies on computationally intensive calculations and thus has led to a surge in resource usage and energy consumption. Even putting aside the environmental toll of server manufacturing and supply chains, AI systems can create a huge environmental cost to communities and regions where they are deployed, including air/thermal pollution due to fossil fuel-based electricity generation and further stressed water resources due to AI’s staggering water footprint.12,25 To make AI more environmentally friendly and ensure that its overall impacts on climate change are positive, recent studies have pursued multifaceted approaches, including efficient training and inference,5 energy-efficient GPU and accelerator designs,19 carbon forecasting,14 carbon-aware task scheduling,1,21 green cloud infrastructures,2 sustainable AI policies,10,18 and more. Additionally, datacenter operators have also increasingly adopted carbon-free energy (such as solar and wind power) and climate-conscious cooling systems, lowering carbon footprint and direct water consumption.8
Although these initiatives are encouraging, unfortunately, a worrisome outcome—environmental inequity—has emerged.3 That is, minimizing the total environmental cost of a globally deployed AI system across multiple regions does not necessarily mean each region is treated equitably. In fact, the environmental cost of AI is often disproportionately higher in certain regions than in others. Even worse, AI’s environmental inequity can be amplified by existing environmental equity agnostic resource allocation, load balancing, and scheduling algorithms and compounded by enduring socioeconomic disparities between regions. For example, geographical load balancing (GLB) algorithms that aggressively exploit regional differences to seek lower electricity prices and/or more renewables7,17 may schedule more workloads to water-inefficient datacenters (located in, for example, water-stressed Arizona), resulting in a disproportionately high water footprint and adding further pressures to local water supplies.9
Addressing the emerging environmental inequity is becoming an integral part of responsible AI.3 It has increasingly received public attention and urgent calls for mitigation efforts. For example, the AI Now Institute compares the uneven regional distribution of AI’s environmental costs to “historical practices of settler colonialism and racial capitalism” in its 2023 Landscape report.11 the United Nations Educational, Scientific and Cultural Organization (UNESCO) recommends against the usage of AI if it creates “disproportionate negative impacts on the environment”;23 California recognizes the need for “ensuring environmental costs are equitably distributed” in its State Report;4 and environmental justice is ranked by Meta as the most critical factor among all environmental-related topics.15
In this article, we advocate environmental equity as a priority for the management of future globally deployed AI systems. Concretely, we explore the potential of harnessing AI workloads’ scheduling flexibility and using equity-aware GLB as a lever to fairly redistribute the environmental cost across regions, ensuring that no single region disproportionately bears the environmental burden. Then, we present key algorithmic challenges to enable AI’s environmental equity without significantly degrading the other performance metrics, such as energy cost and inference accuracy. Finally, we discuss future directions to unleash the full potential of system management for environmentally equitable AI, including coordinated scheduling of AI training and inference, joint optimization of IT and non-IT resources, holistic control of system knobs, and building theoretical foundations.
Our proposal of environmental equity advances the boundaries of existing research on sustainable AI and mitigates the otherwise uneven distribution of AI’s environmental costs across different regions. Equity and fairness are also crucial considerations for AI. The existing research in this space has predominantly tackled prediction unfairness against certain individuals and/or groups.20,26 Thus, environmental equity adds a unique dimension of fairness and significantly complements the existing literature, collaboratively building equitable and responsible AI.
Opportunities and Challenges for Equity-Aware GLB
In this section, we present the potential opportunities of leveraging equity-aware GLB to fairly redistribute the environmental cost across different regions, followed by algorithmic challenges.
Opportunities. The limited power-grid capacity has necessitated increasing flexibility from datacenters to support demand response and maintain grid stability. A notable example is the recent industry initiative to maximize load flexibility for grid-integrated datacenters.6 Specifically, AI workloads exhibit three primary types of flexibility: spatial, where AI training and inference tasks can be distributed across multiple datacenters with minimal impact on latency; temporal, in which AI training tasks can be executed intermittently, provided they meet a given deadline; and performance, where a single inference request can be processed by different AI models, each offering distinct trade-offs between accuracy and resource consumption.
These flexibilities can be exploited to promote environmental equity while satisfying other performance objectives. To achieve this, we can leverage a variety of approaches, such as AI computing resource allocation, load balancing, and job scheduling, which we collectively refer to as system knobs.
In practice, the datacenter fleet of large companies, such as Google and Microsoft, often includes a few tens of self-managed hyperscale datacenters and many more leased third-party colocation datacenter spaces spreading throughout the world.8 By renting virtual machines on public clouds, even a small business can flexibly choose its deployment region and place its computing workloads accordingly. As such, GLB is an important and common knob that can spatially balance computing workloads’ energy demand as well as environmental footprint across different locations.
As a concrete example, we consider moving AI inference workloads around from one datacenter to another and exploit equity-aware GLB to mitigate AI’s environmental inequity. To achieve equitable distribution of AI’s environmental cost, we consider the notion of minimax fairness. Mathematically, denoting as the amount of AI workloads processed in datacenter at time and as the resulting regional environmental cost (for example, due to water consumption12 and air/thermal/waste pollution from non-renewable energy24), we consider an equity-aware objective: , where the first term is the traditional GLB cost (for example, total carbon/water footprint and energy cost) specified based on the prior literature,9 the second term “” serves as the equity regularizer by reducing the highest regional environmental cost, and is the weight.
A snapshot of results. We run a simulation based on the BLOOM model (a large language model) inference trace deployed in 10 different datacenters throughout the world and show a snapshot of our results in Table 1. The details of the simulation are available in Li et al.13 We consider both full GLB (i.e., each request can be flexibly routed to any datacenter) and partial GLB (that is, each request can only be routed to a subset of datacenters depending on its originating location). Compared to common baseline algorithms that simply minimize the total energy cost (GLB-Cost), carbon emission (GLB-Carbon), or workload-to-datacenter distance (GLB-Dist), our algorithm (called eGLB-Off) can effectively mitigate the environmental inequity by reducing the ratio of the maximum to the average regional environmental footprint. Importantly, while there is an inevitable conflict between minimizing the total cost/environmental footprint and addressing the environmental inequity, eGLB-Off can still keep the total cost reasonably low. Additionally, we study a simple online algorithm (called eGLB) based on dual mirror descent to show the potential of mitigating environmental inequity in an online setting. While there is a gap between eGLB and eGLB-Off due to online informational constraints, eGLB outperforms the equity-unaware baseline algorithms in terms of the environmental footprint’s peak-to-average ratio, demonstrating the potential of online GLB to mitigate AI’s environmental inequity.
| GLB | Metric | Algorithm | ||||
|---|---|---|---|---|---|---|
GLB-Cost | GLB-Carbon | GLB-Dist | eGLB-Off | eGLB | ||
| Full | Cost (US$) | 29170 | 45535 | 47038 | 33669 | 33752 |
| PAR (Water) | 1.71 | 1.85 | 1.44 | 1.27 | 1.37 | |
| PAR (Carbon) | 1.68 | 1.70 | 1.41 | 1.13 | 1.22 | |
| Partial | Cost (US$) | 29659 | 45535 | 47038 | 34186 | 34162 |
| PAR (Water) | 1.72 | 1.84 | 1.44 | 1.30 | 1.38 | |
| PAR (Carbon) | 1.69 | 1.71 | 1.41 | 1.12 | 1.22 | |
Challenges. While equity-aware GLB can potentially mitigate AI’s environmental inequity, the equity regularizer “” fundamentally separates our problem from the existing sustainable GLB approaches and creates substantial algorithmic challenges. Specifically, the equity cost “” is unknown until the end of time slots, but complete future information (for example, future workload arrivals and water efficiency) may not be perfectly known in advance. Moreover, even though prediction is often available in practice, it may not be accurate, and its untrusted nature means we cannot simply take the prediction as if it were the ground truth.
Additionally, the traditional design of online competitive algorithms often focuses on guaranteeing the worst-case performance robustness. But, the resulting average performance can be far from optimal due to the conservativeness needed to address potentially worst instances. By contrast, machine learning (ML)-based optimizers (for example, reinforcement learning) can improve the average performance of online decision making by exploiting rich historical data and statistical information, but they typically sacrifice the strong performance robustness needed by real AI systems, especially when there is a distributional shift, the ML model capacity is limited, and/or inputs are adversarial. Thus, to achieve the best of both worlds while pursuing online equitable-aware GLB, we have to carefully balance the usage of traditional competitive algorithms and ML-based optimizers by designing new learning-augmented online algorithms.
Future Directions
We discuss a few future directions to leverage system knobs for environmentally equitable AI.
Coordinated scheduling of AI training and inference. While AI inference offers spatial flexibility, AI model training has great temporal scheduling flexibility, as we can choose when to train the AI models in a stop-and-go manner. We can also choose where to perform AI model training and even possibly change the locations in the middle of the training process. Thus, a potential direction is to explore coordinated scheduling of AI training and inference tasks to fairly distribute AI’s overall environmental costs across different regions.
Joint optimization of IT and non-IT resources. Datacenters have increasingly begun to install onsite carbon-free energy, such as solar power, to partially power the workloads and lower the environmental footprint.21 However, renewables are often intermittent, and the available energy storage capacity is finite. Thus, how to optimize AI demand response given intermittent renewables is challenging, yet worth investigating for addressing AI’s environmental inequity.
Holistic control of system knobs. In addition to GLB, a rich set of system knobs are available and offer flexible trade-offs, such as dynamic model selection for inference, turning servers on/off, and resource allocation to different AI tasks. For example, different AI models can exhibit different energy-accuracy trade-offs for the same task. Holistic control of these system knobs holds enormous potential to curb AI’s resource usage and mitigate environmental inequity, but also presents additional challenges due to the significantly enlarged decision space.
Theoretical foundations. Optimizing a variety of system knobs for environmentally equitable AI has its roots in fair decision making, a classical area that bridges computer systems and algorithms and enjoys a long history with rich theoretical results16,22 and prominent production deployments. However, this classic literature primarily focuses on algorithms that ensure that different job or flow types receive a fair share of system resources, including CPU and memory. Tackling the challenges raised by environmental inequity in modern planet-scale AI systems requires a revisit to the algorithmic foundations and the development of new theoretical tools, which can systematically capture the conflicts between traditional measures of performance, such as accuracy and latency, with measures of emerging importance, such as environmental equity. Thus, it is crucial to build new theoretical foundations to support the design of environmentally equitable AI.
Conclusion
In light of AI’s wildly different environmental costs across different regions, we advocate environmental equity as a priority for the management of future AI systems. We present the potential opportunities and algorithmic challenges of tapping into AI workloads’ scheduling flexibility and leveraging equity-aware GLB to mitigate AI’s environmental inequity. Finally, we discuss a few future directions to unleash the full potential of system knobs for environmentally equitable AI, including coordinated scheduling of AI training and inference, joint optimization of IT and non-IT resources, holistic control of system knobs, and building theoretical foundations. Our proposal of environmental equity pushes forward the boundaries of existing system management for sustainable AI and also adds a unique dimension to AI fairness, collaboratively building equitable and responsible AI.
Acknowledgment
The work of Mohammad Hajiesmaili is supported by NSF CNS-2105494, CNS-2325956, CPS-2136199, and CAREER-2045641. The work of Ramesh K. Sitaraman is supported by NSF CNS-2105494 and CNS-2325956. The work of Adam Wierman is supported by NSF CCF-2326609, CNS-2146814, CPS-2136197, CNS-2106403, and NGSDI-2105648.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment