Abstract Here we quantify the material footprint of artificial intelligence training by linking computational workloads to physical hardware needs. Graphics processing unit demand is expressed as equivalent hardware lifetime consumed and, thus, the material demand involved. As computational demands rise, assessing the environmental impact of artificial intelligence requires moving beyond energy and water consumption to include the material demands of specialized hardware. We analyzed the elemental composition of a widely used graphics processing unit identifying 32 elements. The results show that artificial intelligence hardware consists of approximately 90% heavy metals and only trace amounts of precious metals. Integrating these measurements with computational throughput across varying hardware lifespans, training a large language model requires between 1760 and 8800 graphics processing units. Our findings highlight that incremental model performance gains come at disproportionately high material costs, underscoring the need to incorporate material resource considerations into discussions of artificial intelligence scalability and sustainability.
Falk et al. (2026) studied this question.