From FLOPs to Footprints: The Resource Cost of Artificial Intelligence

• The introduction frames scrutiny of AI materials as increasingly relevant amid concern about scarce resources and expanding digital infrastructure.It notes that existing environmental scrutiny has focused on other technologies and energy-sector materials.Source: §1 · pp. 1–2
• The paper focuses on large-scale deep learning systems, including generative AI applications such as chatbots.Source: §1 · p. 1
• Efficiency techniques can lower computational requirements and broaden deployment, while wider adoption can reinforce demand for infrastructure and materials.The passage relates this pattern to Jevons’ Paradox, in which efficiency gains may accompany greater consumption.Source: §1 · p. 2
• AI’s environmental footprint includes materials in specialized hardware, which energy and water assessments alone do not capture.Source: §1 · p. 2
• The study links model training FLOPs to A100 hardware capacity over its lifespan to estimate material consumption.It combines elemental-composition measurements with a framework mapping training requirements to GPU operational capacity.Source: §1 · p. 2

2 Method

• The framework combines prior A100 elemental measurements with GPU lifetime-compute estimates and model FLOP requirements to assess training material demand.It compares model compute requirements with GPU capacity under operating-lifetime and training-efficiency assumptions.Source: §2 · p. 3

2.1 Elemental Analysis Data

• The A100 was disassembled into five component groups and analyzed by ICP-OES, with preparation and mineralization procedures designed to detect a broad range of elements.The paper reuses elemental data reported in Falk et al. (2025) [27] while adding analytical detail and interpretation.Source: §2.1 · p. 3

2.2 GPU Throughput and AI model Demands

• GPU requirements are estimated by comparing lifetime throughput with model compute demand, assuming sequential GPU use until each device reaches end of life.The estimates vary operational lifespan and the share of delivered compute effectively used for training.Source: §2.2 · pp. 3–4

2.2.1 Nvidia A100: Computational Throughput

• At continuous operation, an A100’s theoretical annual throughput is 9.8×10^21 FLOPs, and Table 1 scopes lifetime throughput across three lifespan scenarios.The estimate uses the A100’s 312 teraFLOPs-per-second peak BF16 throughput.Source: §2.2.1 · p. 4, Table 1 · p. 4

2.2.2 AI Model Training: Computational Requirements

• Dense-transformer training compute is estimated with 6 FLOPs per parameter per token, while GPT-4 is modeled across plausible active-parameter MoE scenarios.The GPT-4 architecture is undisclosed, so the paper accounts for uncertainty in its assumed active-parameter configurations.Source: §2.2.2 · pp. 4–5

2.2.3 Efficiency Adjustments for Real-World Training Scenarios

• GPU counts are adjusted for training efficiency across 20%–50% MFU and 1–3-year lifespans, while estimates exclude extra delays and training interruptions.Real-world MFU varies with infrastructure and software, and the paper treats its scenarios as conservative because excluded interruptions would increase resource use.Source: §2.2.3 · p. 5

2.3 Material Footprint of AI Model Training

• The material estimate scales each GPU’s measured mass for each element by the adjusted GPU count across training scenarios.This calculation is repeated for all 32 detected elements and varies with MFU and hardware-lifespan assumptions.Source: §2.3 · p. 6

3.1 GPU requirement per AI model

• For GPT-4, increasing hardware lifespan from one to three years reduces GPU demand from 5,029 to 1,676 at 35% MFU; across the stated MFU and lifespan bounds, requirements range from 1,174 to 8,800 GPUs.GPU requirements also vary substantially across models: GPT-4 is highest, while Pythia has the lowest requirements, at 11 to 2 GPUs across the scenarios.Source: §3.1 · p. 7, Table 4 · p. 7, Figure 4 · p. 10

3.2 Elemental Composition of the Nvidia A100 GPU

• ICP-OES detected 32 elements in the A100; copper, iron, tin, silicon, and nickel dominate by mass, and component profiles differ considerably.The heatsink is 98.1% copper; the PCB is 46.5% copper and 28% iron, while the GPU chip is primarily chromium, silicon, and tin. The authors argue that model growth will require complete GPU units rather than mainly larger individual chips.Source: §3.2 · pp. 7–9, Figure 2 · p. 8, Figure 3 · p. 8

3.3 The Resource Cost of AI training

• At 35% MFU and a one-year lifespan, GPT-4 training requires 5,029 A100 GPUs and entails estimated extraction and eventual disposal of 7,003 kg of toxic materials.Across nine models, the 20% MFU, one-year scenario requires 13,315 GPUs and 19,940 kg of extracted resources, including 18,545 kg toxic materials; at 50% MFU and three years, these fall to 2,740 kg and 2,550 kg toxic materials.Source: §3.3 · pp. 9–10, Figure 4 · p. 10

3.4 Performance Vs Resource Consumption

• GPT-4 used 31.5 times more GPU resources than GPT-3.5 (2,515 vs 80 GPUs), with gains of +61.1% on MATH and +39.3% on HumanEval but only modest gains on other benchmarks.The authors describe these returns as diminishing relative to computational investment.Source: §3.4 · p. 11, Figure 6 · p. 12
• With nearly identical training budgets, GPT-3.5 outperforms LLaMa 2 across the evaluated benchmarks despite LLaMa 2’s higher MFU of 53%, versus an estimated 20 to 35% for GPT-3.5.The passage notes that data quality, architecture, and other optimization strategies may contribute to different outcomes.Source: §3.4 · p. 12, Figure 6 · p. 12
• GPT-4 required about 81 times Falcon’s training compute and achieved more than 7 times Falcon’s MATH performance, while its HellaSwag improvement was 14%.The passage presents capability scaling as differing across benchmarks.Source: §3.4 · p. 12, Figure 6 · p. 12

3.5 Resource Savings via Training Efficiency and Lifespan Improvements

• Raising MFU from 20% to 60% or extending lifespan from one to three years each reduces GPU needs by about 67%; combining 60% MFU with five years cuts them about 93%.For GPT-4, the combined scenario requires 587 GPUs, compared with 8,800 at 20% MFU and a one-year lifespan; MFU-related savings diminish as hardware lifespan lengthens.Source: §3.5 · p. 13, Figure 7 · p. 13

4 Discussion

• Scaling shows diminishing returns for most capabilities, while mathematical reasoning may require disproportionate compute; architectural and training innovations may improve performance more effectively than raw scaling.Source: §4 · p. 13
• Sequential training scenarios standardize comparisons but do not mirror parallelized workflows, heterogeneous GPU wear, or all effects of distributed-training choices on MFU.Lower MFU can reflect deliberate choices, such as distributing moderate batches across more GPUs to shorten training time.Source: §4 · p. 13
• FLOP estimates omit implementation details that can change real-world compute, yet remain useful for comparing model complexity and as conservative resource baselines.The omitted factors include optimized kernels, re-materialization, and sparse architectures.Source: §4 · pp. 13–14
• GPT-4 compute and material estimates depend on an informed active-parameter estimate because its architecture is undisclosed; alternative configurations would change the results.An interactive tool lets users enter alternative parameters and calculate corresponding results.Source: §4 · p. 14
• The material assessment excludes carbon lost during polymer pyrolysis, oxygen in metal oxides, refining and manufacturing losses, and non-GPU data-center infrastructure, making resource estimates conservative.Unmeasured elements account for 1.78% of the GPU die, 1.73% of the heatsink, 7.55% of the PoP, and 4.05% of the PCB.Source: §4 · p. 14
• Reducing AI training’s material impacts calls for software- and hardware-level optimization and coordinated action, including utilization disclosures and greater transparency from developers and data-center operators.Source: §4 · p. 14

5 Conclusion

• The assessment links training compute to material demand and finds AI’s environmental impacts include mineral extraction and disposal of toxic materials embedded in hardware, beyond operational energy use.Source: §5 · p. 14
• Performance gains from continued computational scaling diminish across many capability domains, so architectural and algorithmic innovation may be more sustainable than brute-force model scaling.Source: §5 · p. 14
• Reducing impacts requires coordinated action by developers, hardware manufacturers, and data-center operators, including utilization reporting and real-world performance data in environmental assessments.Source: §5 · p. 14