Large-scale computational research requires vast compute clusters and extensive data storage, producing measurable environmental effects across energy systems, materials supply chains, and local ecosystems. Evidence from computational science and climate-oriented reviews demonstrates that the environmental footprint is multifaceted and grows as models scale.
Energy use and carbon emissions
Energy intensity of training and operating large models creates significant electricity demand. Emma Strubell University of Massachusetts Amherst documented that deep learning model development and repeated experimental cycles can multiply energy use beyond initial training, increasing the overall carbon budget for research programs. This demand interacts with local power mixes: research institutions supplied by low-carbon grids can reduce net emissions, while those dependent on fossil-fueled electricity magnify climate impacts. Nuanced choices about scheduling, model reuse, and efficiency directly alter cumulative emissions.
Resource extraction, water use, and local consequences
Material impacts arise from the need for specialized hardware such as GPUs and accelerators. Supply chains for metals and rare earths involve mining, refining, and transport, with documented social and environmental harms in producing regions. David Rolnick McGill University and colleagues highlight that scaling compute without attention to lifecycle impacts shifts burdens to territories supplying hardware. Data centers also demand water and land for cooling and infrastructure; in some regions this leads to competition with local agricultural needs and alters land use patterns. Communities hosting data centers or mines often face distinct trade-offs between economic opportunities and environmental costs.
Causes, mitigation, and longer-term consequences
The principal causes are model scaling, inefficient model architecture search, and incentive structures that reward state-of-the-art results over energy-aware design. Consequences include increased greenhouse gas emissions, accelerated resource depletion, and localized environmental degradation that can exacerbate social inequities. Mitigation strategies supported in the literature emphasize algorithmic efficiency, transparent reporting of energy and emissions, reuse of pretrained models, and siting decisions that prioritize renewable energy and responsible supply chains. Policy, procurement, and cultural norms within research communities influence how quickly these practices diffuse.
Recognizing the environmental dimension of computational research reframes technical progress as coupled to energy systems and territorial justice. Promoting efficiency and accountability at institutional and community levels can reduce harms while preserving the scientific and societal benefits of large-scale computation.