AI 훈련의 물질 발자국: A100 GPU 구성과 모델 규모에 따른 자원 요구량 분석

From FLOPs to Footprints: The Resource Cost of Artificial Intelligence Sophia Falk1*, Nicholas Kluge Corrˆea2, Sasha Luccioni3, Lisa Biber-Freudenberger4, and Aimee van Wynsberghe1 1Sustainable AI Lab, Institute for Science and Ethics, Bonn University, Germany *corresponding author: falk@iwe.uni-bonn.de 2Center for Science and Thought, Bonn University, Germany 3Hugging Face 4Center for Development Research, Bonn University, Germany ABSTRACT As computational demands continue to rise, assessing the environmental footprint of artificial intelligence (AI) requires moving beyond energy and water consumption to include the material demands of specialized hardware. This study quantifies the material footprint of AI training by linking computational workloads to physical hardware needs. The elemental composition of the Nvidia A100 SXM 40 GB graphics processing unit (GPU) was analyzed using inductively coupled plasma optical emission spectroscopy (ICP-OES), which identified 32 elements. The results show that AI hardware consists of approximately 90% heavy metals and only trace amounts of precious metals. The elements copper, iron, tin, silicon, and nickel dominate the GPU composition by mass. In a multi-step methodology, we integrate these measurements with computational throughput per GPU across varying lifespans, accounting for the computational requirements of training specific AI models at different training efficiency regimes. Scenario-based analyses reveal that, depending on Model FLOPs Utilization (MFU) and hardware lifespan, training GPT-4 requires between 1,174 and 8,800 A100 GPUs, corresponding to the extraction and eventual disposal of up to 7 tons of toxic elements. Combined software and hardware optimization strategies can substantially reduce material demands: increasing MFU from 20% to 60% lowers GPU requirements by ≈67%, while extending lifespan from one to three years yields comparable savings; implementing both measures together reduces GPU needs by up to 93%. Our findings highlight that incremental performance gains, such as those observed between GPT-3.5 and GPT-4, come at disproportionately high material costs. The study underscores the necessity of incorporating material resource considerations into discussions of AI scalability, emphasizing that future progress in AI must align with principles of resource efficiency and environmental responsibility. Keywords: Sustainable AI, Resource Cost, AI computation, AI model training, FLOPs 1 Introduction Newspaper headlines such as ‘The world is running out of resources for IT’ [1] and ‘Global shortage in computer chips reaches crisis point’ [2] reflect a growing concern regarding the material constraints of the Fourth Industrial Revolution. At the center of this crisis is the semiconductor shortage, rooted in disruptions within the silicon supply chain. The previously significant semiconductor shortage occurred in 2021 during the COVID pandemic [2, 3], and now a surge in demand for hardware accelerators for deep learning workloads could lead to the next global chip shortage [3]. It is important to note that the infrastructure powering emerging technologies, including artificial intelligence (AI)1, is built on far more than just silicon. Data centers, which form the physical backbone of emerging technologies, rely on a wide array of increasingly scarce and valuable materials, including rare earth elements, tantalum, and cobalt [4]. While there is an increasing awareness of the environmental implications and supply chain constraints tied to the extraction of critical resources for ‘green’ technologies and the energy sector, e.g. lithium for electric vehicles [5] or cobalt for renewable energy storage systems [6, 7], there is a lack of equivalent scrutiny when it comes to the material demands of AI. As AI becomes a defining force of the ongoing Fourth Industrial Revolution, it is essential to identify and understand the ‘AI materials’ that constitute the infrastructure and materiality of AI. This becomes more urgent as global investments in digital infrastructure continue to accelerate. The rapid expansion of AI capabilities has driven unprecedented demand for high-performance computing, prompting substantial capital flows into 1By ‘AI’, here we refer specifically to large-scale deep learning systems, such as those used for generative AI applications like chatbots. 1 arXiv:2512.04142v1 [cs.CY] 3 Dec 2025 data center construction and modernization. In 2024 alone, Alphabet, Microsoft, Meta, and Amazon collectively invested approximately $246 USD billion in AI and data center development [8]. Projections indicate that this trajectory will continue: according to McKinsey, AI-ready data center capacity is expected to grow at an average annual rate of 33% between 2023 and 2030, with AI workloads accounting for nearly 70% of total data center demand by the end of the decade [9]. Despite this rapid expansion and increasing electricity demand, data centers accou

…(Full text truncated)…

This content is AI-processed based on ArXiv data.