Abstract
While "Green AI" initiatives have successfully drawn attention to the massive energy and carbon footprints of large-scale models, they frequently overlook the geographical distribution of these impacts. This paper introduces the concept of environmental equity in AI, arguing that minimizing global footprints is insufficient if it results in localized environmental crises. We propose using Geographical Load Balancing (GLB) to redistribute AI workloads, ensuring no single region bears a disproportionate burden of resource consumption. By employing a minimax fairness objective, we demonstrate that significant gains in regional equity—measured by the Peak-to-Average Ratio (PAR) of water and carbon impacts—can be achieved with a manageable increase in operational costs.
Introduction
The rapid proliferation of Large Language Models (LLMs) has sparked a critical dialogue regarding the sustainability of AI. However, current resource management strategies are often "equity-agnostic," prioritizing global cost minimization or carbon reduction without considering the localized sociotechnical consequences. For instance, a system might minimize its total carbon footprint by shifting workloads to a data center in a water-stressed region, exacerbating local environmental injustices. This paper argues that environmental equity must be integrated into AI fairness research as a primary objective, alongside traditional metrics like prediction accuracy and algorithmic bias.
Methodology
The core of our approach is the application of Geographical Load Balancing (GLB) as a mechanism for achieving equity. We frame the workload distribution problem using a minimax fairness objective. Rather than simply minimizing the sum of all costs, our objective function incorporates a regularizer that penalizes the maximum regional environmental impact:
$$\text{Objective} = \sum \text{Total Cost} + \lambda \cdot \max_{\text{region}} (\text{Regional Environmental Impact})$$
We explore three dimensions of workload flexibility:
- Spatial Flexibility: Dynamically routing inference requests between geographically dispersed data centers.
- Temporal Flexibility: Scheduling non-urgent tasks (e.g., model training or fine-tuning) during periods of high renewable energy availability or low environmental stress.
- Performance Flexibility: Selecting between model variants that offer different trade-offs between inference accuracy and resource consumption.
We evaluate two algorithmic implementations: eGLB-Off, an offline optimization providing a theoretical upper bound on performance, and eGLB, a practical online algorithm based on dual mirror descent designed for real-time decision-making under uncertainty.
Experimental Setup
Our simulations utilize inference traces from the BLOOM model, a 176-billion parameter open-access LLM. We model a hypothetical deployment across ten global data centers, incorporating real-world data on regional water consumption factors and carbon intensity. To measure inequity, we utilize the Peak-to-Average Ratio (PAR). A PAR of 1 indicates perfect equity across regions, while higher values signify that specific regions are bearing a disproportionate share of the environmental burden.
Key Results
Our findings highlight a fundamental trade-off between global cost efficiency and regional environmental equity. As shown in Table 1, traditional cost-minimizing strategies (GLB-Cost) result in high PAR values for both water (1.71) and carbon (1.68).
The proposed eGLB-Off algorithm significantly improves these metrics, reducing the PAR for water to 1.27 and carbon to 1.13. This improvement comes with a modest "equity tax"—a roughly 15% increase in total operational costs (from $29,170 to $33,669). Notably, the real-time eGLB algorithm maintains high performance, achieving PAR values of 1.37 (water) and 1.22 (carbon) while remaining computationally feasible for live deployments. These results demonstrate that substantial progress toward environmental justice is possible without prohibitive financial overhead.
Discussion and Future Work
This research expands the scope of AI fairness to include the physical communities hosting data center infrastructure. While our results are promising, several challenges remain. Online algorithms must be further refined to handle the inherent unpredictability of workload arrivals and renewable energy availability, perhaps through learning-augmented approaches. Additionally, future work should explore the synergy between spatial and temporal flexibility, as well as the potential for holistic control systems that dynamically adjust model architectures based on real-time local environmental stress. Establishing robust theoretical foundations for the conflict between performance and equity will be essential for the next generation of sustainable AI systems.
Summary updated