Architectural Imperative: Distributed Computer Systems Infrastructure for a Sustainable AI World

Abstract

Rapid advances in artificial intelligence (AI), driven primarily by the scaling of large language models (LLMs), have exposed critical ecological and infrastructural limitations inherent in prevailing centralized computing paradigms. Electricity consumption in global data centers reached approximately 415 terawatt-hours (TWh) in 2024 and is projected to surpass 945 TWh by 2030, a trajectory that existing hardware efficiency improvements alone are incapable of reversing given the compounding effect of Jevons' Paradox. This brief contends that distributed computer systems infrastructure represents the essential architectural foundation for any feasible trajectory towards sustainable AI development. Through systematic examination of how core distributed systems principles—encompassing consensus protocols, fault tolerance mechanisms, state replication strategies, and distributed storage architectures—intersect with the operational demands of large-scale AI workloads, a concrete engineering pathway toward reduced environmental impact is identified and elaborated. The roles of container orchestration, 3D hybrid parallelism, carbon-aware workload scheduling, and high-performance interconnect topologies in maximizing resource utilization and minimizing idle energy consumption are examined in depth. Federated learning and edge inference are further explored as mechanisms for displacing computation toward the network edge, where training data originates and where sunk device energy can be productively leveraged. This brief is structured as a detailed blueprint for expansion into a full-length 40-page journal article, providing the technical scaffolding required to demonstrate, through distributed systems engineering rather than policy aspiration, that meaningful decoupling of AI capability growth from ecological degradation is architecturally achievable.