The Environmental Impact of Large Language Models: Balancing Efficiency and Performance

Reconciling the contrasting narratives on the environmental impact of large language models - Scientific Reports

In addition to the expensive option of GPT-4-32k, OpenAI offers multiple alternative LLMs for ChatGPT with lower prices. While the cheapest option GPT-3.5-Turbo only costs $2.00 per 1M output tokens, we considered the recently released GPT-4o offered by OpenAI at $15.00 per 1M output tokens as of May 202437. Therefore, the human-to-LLM cost ratio is 1200 in the U.S. and 130 in India. Comparing the environmental impact of LLMs to an amortized portion of a human's total footprint may seem unconventional, but we believe it is necessary for accurately assessing environmental costs in labor and production. When a company employs someone, they are not just paying for task-specific energy expenditure. They are effectively renting a portion of that person's life -- typically 40 hours per week. During this time, all of the individual's environmental impacts, from commuting to basic life functions, are part of the cost of their labor. This view aligns with how industrial civilizations think about economic compensation. Companies do not pay employees only for the calories they burn typing; they pay them a salary that supports their entire life (during work hours at least, and possibly even a living wage). Environmental accounting should follow the same principle. This approach is reasonable when comparing human labor to LLMs. While LLMs have quantifiable energy and resource costs, human labor involves interconnected environmental impacts beyond immediate tasks. Considering total environmental impact during work hours provides a more accurate representation of human work costs compared to LLM alternatives. Our approach aims to reframe the discussion of environmental impacts in labor and production. Instead of allowing corporations to externalize costs by focusing only on task-specific impacts, we argue for a view that acknowledges the full scope of resources dedicated to work activities. This invites deeper consideration of environmental trade-offs in various production modes, including potential substitution of human labor with LLMs. At first glance, this study's findings suggest that replacing human labor with AI could lead to substantial environmental benefits, as the direct environmental footprint of LLMs is significantly lower than that of humans for the same output. The comparative LCA results highlight the substantial environmental and economic advantages of Llama-3-70B over human labor in content creation. Across all four metrics -- energy consumption, water consumption, carbon emissions, and costs -- Llama-3-70B outperforms human labor by orders of magnitude, with human-to-LLM ratios ranging from 40 to 150. When compared to a lightweight AI model (Gemma-2B-it), the ratios range from 1200 to 4400. For the case of India, the human-to-AI ratios are between 3.4 and 16 for a typical LLM and between 130 and 1100 for a lightweight LLM. These findings emphasize the potential of LLMs to reduce the environmental impact of knowledge work and creative tasks, while simultaneously reducing costs. Despite our conservative comparison (e.g., using lower energy and cost values for human labor when applicable), however, we should interpret this study's findings with cautious optimism. As model sizes continue growing (e.g., recently released Llama-3.1-405B), the energy consumption of LLMs as well as the environmental footprint will likely increase substantially. As a consequence, LLMs may be more energy-consuming than human labor, especially for Indian residents. Thus, we emphasize the need for ongoing research efforts to ensure the energy efficiency and sustainability of LLMs in the long term. The economic effects of LLM adoption extend beyond the immediate environmental benefits shown in our analysis. While LLMs can reduce the environmental impact of content creation compared to human labor, a straightforward replacement is improbable. LLM integration into various industries will likely be influenced by factors such as the rebound effect and profit-seeking behavior. Moreover, the current pricing provided by OpenAI may be heavily subsidized in order to drive continued user growth. As a result, the economic cost of LLMs could rise in the future. LLM adoption will likely have immediate, mid-term, and long-term consequences for the economy and society. The initial impact will likely be increased productivity in content creation. Writers, illustrators, and other creators can use LLMs to work more efficiently and potentially produce better output by exploring more possibilities in their creative process. This could increase content volume and variety, possibly benefiting consumers through lower prices and more choices. However, it might also decrease demand for traditional content creation jobs. As these changes occur, the job market will likely shift. While some traditional content creation roles may diminish, new opportunities will likely appear. These could include content creators skilled in using LLMs, who might earn higher wages, as well as supervisory roles like editors, curators, and LLM system managers. There will also be a need for technicians to maintain and improve LLM systems. This shift will require changes in education and job training to develop new skills. The lower costs and easier entry into content creation could also encourage new business models and increase competition in content-focused industries. The long-term effects of widespread LLM adoption could be significant and may take years to become apparent. Industries that rely heavily on content creation may need to change how they operate. There is a risk of growing inequality between those who can use AI technologies effectively and those who cannot. How we view creativity, originality, and the nature of work may change as AI-generated content becomes common. We might see a split in the content market: high-volume, low-cost AI content alongside more expensive human-created content. LLMs are already affecting copyright laws, which could change how we value and protect intellectual property. To illustrate potential economic impacts, we can consider a hypothetical scenario where LLMs are adopted for a significant portion of content creation tasks in the U.S. over the next decade. Assuming current minimum wage rates and LLM costs, this could result in substantial direct cost savings in labor costs for content creation. We might see significant job market shifts, with potential displacement of many content creation jobs, partially offset by the creation of new roles in LLM management and specialized content creation. Notable productivity gains could emerge, with a potential multiple-fold increase in content output per dollar spent. Additionally, we might observe market expansion in content-related industries due to reduced costs and increased accessibility. These potential outcomes highlight the need for careful planning and proactive policies as LLM use increases. While the immediate benefits in resource efficiency and cost savings are clear, the broader effects on society, the economy, and the environment are varied and interconnected. The actual path of LLM integration will likely involve both human and LLM-driven work, rather than LLMs simply replacing humans. How LLM capabilities, human skills, market needs, and consumer preferences interact will shape the future of content creation and distribution in ways we can not fully predict. As this transition occurs, more research is needed to measure these impacts accurately and develop ways to reduce potential negative consequences while maximizing the benefits of this technology. The main challenge is how to use the environmental and productivity advantages of LLMs while ensuring fair economic outcomes and maintaining the value of human creativity and expertise. Furthermore, while LLMs like Llama-3-70B have demonstrated impressive language generation capabilities, they are also known to produce biased, inconsistent, or factually incorrect outputs. (Whether LLMs are more biased, etc., than typical humans, though, is an open question.) The widespread use of LLMs for content creation may exacerbate the spread of misinformation, perpetuate societal biases, or lead to the erosion of trust in online content, or conversely, reduce those effects by replacing even more flawed humans. Addressing these challenges will require the development of robust quality control mechanisms, fact-checking processes, and ethical guidelines for the use of LLMs in content creation. The potentially substantial environmental benefits of LLMs over human labor for content creation tasks present a complex ethical landscape that merits careful consideration. On one hand, the dramatic reduction in energy consumption, water usage, and carbon emissions offered by LLMs aligns with urgent global sustainability goals and could contribute significantly to mitigating climate change. This environmental advantage creates a strong ethical argument for their widespread adoption. On the other hand, the potential societal impacts of LLMs-including job displacement, the spread of misinformation, and the perpetuation of biases-raise equally important ethical concerns. The tension between these competing ethical considerations highlights the need for a nuanced approach to LLM implementation. It may be necessary to develop frameworks that balance the environmental benefits of LLMs with strategies to mitigate their potential negative societal impacts. This could include investing in retraining programs for displaced workers, implementing strict content verification processes, and continuously refining LLMs to reduce biases. Ultimately, the ethical deployment of LLMs will require ongoing attention to ensure that the pursuit of environmental sustainability through LLMs does not come at the cost of social equity and information integrity. It is also equally important to strengthen research efforts to ensure the long-term sustainability of LLMs, especially as model sizes and energy consumption of LLMs continue to increase. This study has several limitations that we address here. These limitations also point toward opportunities for future research. First, the environmental and economic impacts of Llama-3-70B/Gemma-2B-it and human labor may vary depending on the nature of the content creation task. Our analysis focused on a relatively simple task of writing a 500-word page of content, and the results may not be generalizable to more complex or domain-specific tasks. Second, our study relied on publicly available data and assumptions about the energy consumption, water consumption, carbon emissions, and economic costs of Llama-3-70B/Gemma-2B-it and human labor. While we have made efforts to use the most reliable and up-to-date data sources available, there may be uncertainties and variations in the actual impacts based on the specific hardware and infrastructure used, the geographic location, and the individual behavior of human workers. For example, where the human worker lives makes a large difference in their impact per unit of work produced. Third, our LCA approach does not account for the potential long-term environmental and economic impacts of LLM adoption, such as the effects on job displacement, skills development, and innovation. Future research should seek to address these concerns. Finally, our study compared Llama-3-70B/Gemma-2B-it to human labor for content creation, but there are other LLMs and AI-based content creation tools available, each with their own environmental and economic impacts. In future research, we would like to include a broader range of LLMs and content creation approaches, as well as explore the potential for combining human and AI capabilities for optimal performance and sustainability. As the U.S. National Academies wrote in their 2012 report, Computing Research for Sustainability, "sustainability is not, at its root, a technical problem, nor will merely technical solutions be sufficient. Instead, deep economic, political, and cultural adjustments will ultimately be required, along with a major, long-term commitment in each sphere to deploy the requisite technical solutions at scale. Nevertheless, technological advances and enablers have a clear role in supporting such change." Our findings demonstrate that LLMs can significantly reduce the environmental footprint of content creation in comparison to human labor, highlighting the potential of this technology to contribute to sustainability efforts in the realm of work. However, the actual impact of LLMs on sustainability will depend on a range of cultural, social, and economic factors that shape their development and deployment, which could lead to either a net reduction or increase in environmental impact. We present the analyses described below as a step toward broader understanding of the role of LLMs in the future of sustainable work.

Small Models, Big Impact: The Sustainable Future of AI Language Models

It is well established by now that generative AI systems have demonstrated remarkable capabilities, from generating nuanced human-like text and engaging in complex problem-solving to assisting with code development and creative tasks. Their impact spans across industries, revolutionizing content creation, customer service, and data analysis. The power of these models stems from their sheer size and complexity. LLMs are built on neural networks with an enormous number of parameters - the adjustable elements that allow the model to learn and make predictions. The scale of these models has grown exponentially in recent years, driven by the principle that larger models, trained on more data, generally perform better across a wide range of tasks. This approach, often referred to as "scaling laws" in machine learning, has led to the development of increasingly massive models. For context, OpenAI's GPT-3, introduced in 2020, boasts possibly more than 175 billion parameters. More recent models have pushed these boundaries even further, with Google's PaLM featuring 540 billion parameters and DeepMind's Gopher reaching 280 billion. Some models, like Wu Dao 2.0 developed by the Beijing Academy of Artificial Intelligence, have even surpassed the trillion-parameter milestone. However, as these models grow increasingly powerful and complex, a significant concern has surfaced: the energy consumption for model training and the subsequent environmental impact. The immense computational resources required to train and operate these massive models translate into considerable energy consumption and, consequently, a larger carbon footprint. Training a model with hundreds of billions of parameters requires vast amounts of electricity, often equivalent to the annual energy consumption of hundreds of households. This growing environmental cost has sparked a crucial debate within the AI community and beyond. Researchers, industry experts, and environmentalists are now grappling with a pressing question: Is it possible to develop more compact, energy-efficient models that can rival the performance of their larger counterparts? To understand the significance of this question, we must first grasp the scale of energy consumption associated with today's leading LLMs. Take GPT-3, for instance, one of the most well-known language models. According to a study [1], its training process alone consumed an estimated 1,287 MWh of electricity - equivalent to the annual energy use of 120 average American households. While more recent models like Meta's Llama 3.2 have shown improvements in efficiency. According to Meta's model card, both its 1 billion and 3 billion parameter variants consumed just over 581 MWh combined to train - about half the energy required for GPT-3 [2]. To put it another way, training Llama 3 used roughly as much energy as a large airliner consumes during a 7-hour flight. The estimated total greenhouse-gas emissions resulting from Llama 3.2's training was 240 tons CO2eq, however almost 100% of the electricity was based on renewable energy, making it largely carbon neutral. As AI continues to advance, some projections are alarming. The study projected that by 2027 the AI sector's annual energy consumption could rival that of entire countries like the Netherlands [1]. While much attention is paid to the energy required for training these massive models, the ongoing energy costs of inference - using the model to generate outputs - are also significant. Research done at Northeastern University reports that the 65-billion parameter version of the LLaMA model consumes about 4 Joules per output token [2]. This means that generating a passage of about 250 words requires approximately 0.00037 kilowatt-hours of energy. Note: Only model output tokens were considered for this statistical illustration. In response to these challenges, researchers are exploring ways to create smaller, more efficient models that can deliver comparable performance to their larger counterparts. This shift towards "small models with big impact" is driven by both environmental concerns and practical considerations. One promising approach comes from researchers at UC Santa Cruz, who developed a method to run a billion-parameter-scale language model on just 13 watts of power - about the same energy required to power a light bulb [3]. This represents a more than 50-fold improvement in efficiency compared to typical hardware. In this study, researchers focused on eliminating computationally expensive operations like matrix multiplication during model training while maintaining performance. Some are exploring efficient parameterizations and meta-learning techniques to reduce the need for ever-larger models. Another relatively smaller model named Orca was trained on a student-teacher mechanism, which involved imitating the reasoning processes of the strong GPT-4 model as a teacher [4]. Orca has 13 billion parameters and was trained with synthetic data generated by gpt-3.5-turbo and GPT-4 using 20 NVIDIA A100 chips for a total of 200 hours. While it's significantly smaller than any GPT model, it still outperformed gpt-3.5-turbo (ChatGPT) model when evaluated with GPT-4, while only consuming approximately 1,600 kWh of energy during its training process. Yet another example is the Phi-1.5 model, which has merely 1.3 billion parameters and is found to outperform models 5x its size on natural language tasks [5]. The Phi-1.5 model, trained on an A100 chip, consumed approximately 600 kWh of energy. While Phi-1.5's performance compared to other larger models is controversial and somewhat disputed, it's undeniable that the model trained on a domain-specific dataset can be a significant value proposition and tremendously cost-effective. The pursuit of more efficient models isn't just about reducing size - it's about maintaining or even improving performance. Various techniques to create application-specific smaller models that can outperform their larger counterparts are a reality. Some of the key approaches include: As we look to the future of AI language models, it's clear that finding the right balance between performance and sustainability will be crucial. While large models have demonstrated impressive capabilities, the environmental and economic costs associated with training and running them are becoming increasingly difficult to ignore. The development of smaller, more efficient models offers a promising path forward. By leveraging advanced techniques in model compression, knowledge distillation, and hardware optimization, AI practitioners aim to create models that can deliver comparable results to their larger counterparts while consuming significantly less energy. This shift towards more sustainable AI doesn't just benefit the environment - it also has the potential to democratize access to powerful language models. Smaller, more efficient models could be run on less powerful hardware, making advanced AI capabilities accessible to a wider range of organizations and individuals. As we stand at the crossroads of AI innovation and environmental responsibility, the pursuit of smaller, more efficient language models represents a crucial step towards a more sustainable future. By focusing on creating "small models with big impact," the AI community can continue to push the boundaries of what's possible while also addressing the pressing need for more environmentally friendly technologies. The challenge ahead is clear: to develop models that can match or exceed the performance of today's largest language models while dramatically reducing their energy footprint. As research continues to make strides in this direction, we move closer to a future where powerful AI tools can be deployed widely and responsibly, without compromising on performance or placing undue strain on our planet's resources. In the end, the true measure of AI's success may not just be its ability to process language or generate human-like text, but its capacity to do so in a way that's sustainable for both our digital and natural environments. The ongoing research into smaller, more efficient models promises to play a crucial role in shaping this sustainable AI future. Anjanava Biswas is an award-winning Senior AI Specialist Solutions Architect at Amazon Web Services (AWS), a public speaker, and author with more than 16 years of experience in enterprise architecture, cloud systems, and transformation strategy. He is dedicated to artificial intelligence, machine learning, and generative AI research, development, and innovation projects for the past seven years, working closely with organizations from the healthcare, financial services, technology startup, and public sector industries. Biswas holds a Bachelor of Technology degree in Information Technology and Computer Science and is a TOGAF certified enterprise architect; he also holds 7 AWS Certifications. Biswas is a Senior IEEE member, and a Fellow at IET (UK), BCS (UK), and IETE (India). Connect with Anjanava Biswas at anjan.biswas@IEEE.org or on LinkedIn.