Escribe una PREreview

The rapid expansion of AI has intensified concerns about its environmental sustainability. Yet, current assessments predominantly focus on operational carbon emissions using secondary data or estimated values, overlooking environmental impacts in other life cycle stages. This study presents the first comprehensive multi-criteria life cycle assessment (LCA) of AI training, examining 16 environmental impact categories based on detailed primary data collection of the Nvidia A100 SXM 40GB GPU. The LCA results for training BLOOM reveal that the use phase dominates 11 of 16 impact categories including climate change (96\%), while manufacturing dominates the remaining 5 impact categories including human toxicity, cancer (99\%) and mineral and metal depletion (85\%). For training GPT-4, the use phase dominates 10 of 16 impact categories, contributing about 96\% to both the climate change and resource use, fossils category. The manufacturing stage dominates 6 of 16 impact categories including human toxicity, cancer (94\%) and eutrophication, freshwater (81\%). Assessing the cradle-to-gate environmental impact distribution across the GPU components reveals that the GPU chip is the largest contributor across 10 of 16 of impact categories and shows particularly pronounced contributions to climate change (81\%) and resource use, fossils (80\%). While primary data collection results in modest changes in carbon estimates compared to database-derived estimates, substantial variations emerge in other categories. Most notably, minerals and metals depletion increases by 33\%, demonstrating the critical importance of primary data for non-carbon accounting. This multi-criteria analysis expands the Sustainable AI discourse beyond operational carbon emissions, challenging current sustainability narratives and highlighting the need for policy frameworks addressing the full spectrum of AI's environmental impact.