Source-linked AI summary
Wasted large language models: A life cycle thinking approach
Erik Johannes Husom, Maria Emine Nylund, Ophelia Prillard
TL;DR
LLMs have a growing carbon footprint, while efficiency gains alone have not reduced consumption. This paper applies life cycle thinking and the EU waste hierarchy to LLMs, highlighting prevention and reuse as approaches to reduce environmental impact.
Problem
LLMs require substantial energy for training and use, and their growing climate impact motivates approaches beyond improving energy efficiency.
Method
The paper applies life cycle thinking and the EU waste hierarchy to LLM development and deployment, examining prevention, reuse, recycling, recovery, and disposal.
Results
Applying the waste hierarchy highlights avoiding new model training and reusing existing models as key ways to reduce LLM-related energy consumption and environmental impact.
Takeaways & Limitations
The framework can help developers and users use existing LLMs more consciously and counter a throw-away mentality around model development.
Takeaways & Limitations
The paper discusses improving use of existing models while assuming they have already been produced and will be used, rather than examining reduced model use.
Abstract
from arXiv · showhide
Large Language Models (LLMs) are machine learning (ML) models that have an increasingly large carbon footprint through their development and use. Efforts to increase the energy efficiency of these models have not translated into reduced consumption due to rebound effects such as Jevons Paradox - that increased efficiency drives increased use. There is therefore a need for additional measures to solve this problem. We suggest that one possible way forward is to use life cycle thinking, and view LLMs as products that can become waste. With this perspective, we investigate the potential of the waste hierarchy from the EU's Waste Framework Directive, which suggests five different measures for how to manage waste: prevention, reuse, recycling, recovery, and disposal. We examine how these measures can inform and motivate new types of thinking and approaches to reducing LLM waste and their environmental impact in general. Applying the waste hierarchy to LLMs highlights that preventing waste is essential for reducing the models' environmental impact, mainly because it reduces the need for training new models. Prevention can be achieved through many existing methods for reusing, "recycling", and "recovering" LLMs. Additionally, disposal can be important both for saving energy and for keeping a considerate attitude to the resources being spent on training LLMs. We also call to attention that prevention of unnecessary use of LLMs carry huge potential for lowering the climate impact of the models.
1 Introduction
LLMs offer productivity and efficiency benefits but require substantial energy to train and use, creating a need for more environmentally sustainable AI. The paper applies life-cycle thinking and the EU waste hierarchy to reduce unnecessary training costs and improve reuse of existing LLM artifacts.
- Motivation: LLMs can improve productivity and efficiency, especially in software development, but their training and use at scale require enormous amounts of energy.The paper identifies this energy demand as a source of considerable carbon footprint.
- Motivation: Energy-efficiency research and rapidly advancing training methods create incentives for increasingly frequent LLM training and retraining.The introduction links improved pre-training, post-training, and mid-training methods to faster development cycles.
- Problem: Growing numbers of LLM developers and rising demand from proprietary services and open-weight models contribute to a large number of models being trained.The introduction cites frontier labs, record-breaking user numbers, and a broad ecosystem of downloadable models.
- Approach: Life-cycle thinking uses the EU waste management hierarchy—prevention, reduction, reuse, recycle, and responsible disposal—to examine how products can deliver benefits with less cost and waste.The passage presents waste management as part of life-cycle thinking and lists the hierarchy's five steps.
- Contribution: The paper applies life-cycle thinking and the EU waste hierarchy to LLM development and deployment, emphasizing reduced unnecessary training costs and improved reuse of existing artifacts.It also discusses weaknesses of LLMs in the context of life-cycle thinking.
2 Background
This section introduces life cycle thinking and the EU Waste Framework Directive’s waste hierarchy as foundations for applying waste-management principles to LLMs. It frames LLMs as software products whose lifecycle and waste can be assessed despite differences from traditional computer code.
- Life cycle thinking: Life cycle thinking assesses products and services across their full lifecycle, including production, use, and what happens after their useful purpose ends.A product may be defined as waste when it is no longer useful or valuable for its original intended purpose.
- EU waste hierarchy: The EU Waste Framework Directive establishes a prioritized hierarchy of prevention, preparing for re-use, recycling, recovery, and disposal to improve environmental outcomes.The directive provides a legal framework for preventing and managing waste in the EU.
- Applying the hierarchy to LLMs: The waste hierarchy is generally aimed at material products, but this paper investigates its application to immaterial products such as software and LLMs.The section presents this application as a way to investigate waste-management principles for machine learning models.
- LLMs as software: LLMs are software because they contain low-level instructions executed by computers, represented as numbers that transform inputs into outputs.Unlike traditional code, LLMs are produced using training data and an ML algorithm that optimizes parameters for a task.
- Software as products: Viewing both computer code and LLMs as human-created and used products provides the basis for defining software waste.Their differences affect how the waste hierarchy may reduce LLMs’ environmental impact.
3 Waste hierarchy for LLMs
The waste hierarchy frames LLMs as products whose environmental impact can be reduced by preventing unnecessary model production, extending reuse, adapting existing models, and disposing of unused models thoughtfully. Although waste management primarily affects development, storage, movement, and inference energy, avoiding unnecessary LLM use is also an important climate measure.
- Defining LLM waste: LLM waste arises when older models are deprecated after better or cheaper models appear, despite software being easy to copy and reuse.Because a copy may still run elsewhere, complete waste is difficult to establish definitively.
- Environmental impact: Reducing waste can lower energy use across developing, storing, moving, and using models, mainly by avoiding new model production and reusing existing models.Hardware, land-use, and cooling impacts remain important but are secondary to these four processes in this waste-management approach.
- Preventing unnecessary use: Preventing unnecessary LLM use is distinct from preventing product waste but remains an important way to reduce environmental impact.The paper highlights a need for HCI that makes users aware of LLM environmental costs and supports life cycle thinking.
- Prevention and reuse: Reuse is a central prevention measure because unlimited copying can amortize the energy spent developing, storing, and moving LLMs.Extending model lifespans through sufficiency, slower turnover, retrieval-augmented generation, and reusable instructions can counter deprecation.
- Recycling: Recycling LLMs means extending or adapting existing models, including fine-tuning and low-rank adaptation, rather than training entirely new models.Unlike ordinary code, LLMs require more complex techniques to recycle.
- Disposal: Disposing of unused LLMs can reduce climate impact because storing models consumes substantial energy, with typical model sizes ranging from a few to several hundreds of GBs.Deletion can nevertheless encourage a throwaway mentality that obscures the energy and time invested in model development.
4 Discussion
The discussion presents waste prevention—especially avoiding unnecessary training—as central to reducing LLM climate impacts, while acknowledging that new models may sometimes be necessary. It frames the waste hierarchy as a consciousness-raising approach rather than an immediate legal or mainstream-practice guideline.
- Waste prevention: Avoiding new model training is identified as a main way to use the waste hierarchy to reduce LLM climate impacts.The proposed overview is not exhaustive but offers a way to improve LLM waste management.
- Waste prevention: New models may be necessary to study training improvements or address bias and hallucinations, although augmentation can reduce some existing weaknesses.The discussion emphasizes judging when new training is worthwhile to avoid unnecessary waste.
- Model lifespans: 62 LLMs tracked from 2018-2025 across 108 000 citing papers showed that LLM lifespans are getting shorter.Understanding model lifespans, usage decline, and changing performance and efficiency requires better data collection.
- Limitations: The paper does not assess LLMs’ net value, because benefits and long-term effects remain unclear amid rebound effects such as cognitive outsourcing and technical debt.It treats LLMs as products and focuses on reducing their waste rather than determining their overall value.
- Limitations: Reuse and recycling may struggle to gain traction while model capabilities increase rapidly and growth-based economic incentives remain dominant.Climate-conscious motivations include avoiding climate damage, preserving resources for critical uses, and reducing nature destruction.
- Framework scope: The framework encourages developers and users to use existing models and become more conscious of development resources, countering a throw-away mentality.It is presented as a starting point, not as a short-term law or mainstream-practice guideline.
5 Conclusion and future work
The paper applies life cycle thinking and the EU Waste Framework Directive’s waste hierarchy to LLMs, reframing their development and deployment as processes that can generate waste. This perspective supports more considerate approaches to reducing LLMs’ environmental and climate impact.
- Conclusion and future work: Life cycle thinking reframes LLM development and deployment as processes involving products that can become waste.This reframing shifts attention toward reducing the technology’s environmental impact.
- Conclusion and future work: The EU Waste Framework Directive’s waste hierarchy provides five measures for analyzing how to reduce LLMs’ environmental impact.The measures are prevention, reuse, recycling, recovery, and disposal.
- Conclusion and future work: A more considerate approach to LLM development and deployment can lead to less climate impact.The paper examines the waste hierarchy as a framework for motivating such approaches.