The Quest for the Greenest LLM: A Deep Dive into AI’s Environmental Impact

My name is Jean Philippe Barrett, and with over 25 years in IT and a background in renewable energy, I have been passionately involved in IT sustainability. Through my company, Better Bytes, I’ve been exploring how the tech industry, especially AI, can align with environmental goals. Today, I want to share insights from the first comprehensive assessment of the environmental impact of a leading LLM supplier. This study offers unprecedented transparency and data-driven guidance on how AI can be developed and deployed more sustainably.

Why Digital Sustainability Matters Now More Than Ever

Despite the green reputation of places like New Zealand, where I live, we face significant challenges in energy supply and sustainability. Our beautiful country struggles to provide enough electricity for both industry and households, leading us to import and burn coal. This situation mirrors global trends where data centers consume a growing share of green energy capacity, intensifying the urgency to transition away from fossil fuels.

AI plays a complex role here. On one hand, it is a driver behind some big tech companies delaying their carbon neutrality goals. On the other hand, hyperscalers in regions like New Zealand are investing in renewable infrastructure, though renewable capacity takes time to build—just like data centers themselves.

The big question remains: are data centers’ energy and carbon challenges temporary or a long-term issue? Unfortunately, the truth is we don’t have reliable data, especially about AI’s environmental footprint. Many AI companies keep their environmental data secret or use questionable methodologies, making it difficult for businesses to factor sustainability into their AI strategies effectively.

The Risk of AI Undermining Corporate Sustainability

There’s a real risk that AI’s environmental impact could undermine companies’ own sustainability commitments, particularly regarding scope 3 emissions, category 1—which covers purchased goods and services. This makes transparent, comprehensive data on AI’s environmental footprint not just desirable, but urgent.

Introducing the First Comprehensive Environmental Assessment of an LLM Supplier

Today’s discussion centers around the first-ever in-depth environmental impact assessment of a leading LLM supplier. My team in Europe had the privilege of peer reviewing this study, which is groundbreaking for three reasons:

1. Full Life Cycle Coverage: Unlike many previous analyses, this study evaluates the entire life cycle of AI delivery—from training to inference to hardware disposal.
2. Multiple Environmental Factors: It considers not only carbon emissions but also water consumption and resource depletion.
3. Firsthand Data: The study was commissioned by the LLM supplier itself, granting auditors unrestricted access to real operational data rather than estimates or proxies.

While I can’t yet reveal the name of the AI firm—though it’s not OpenAI or GPT-4, but a popular competitor—the report will soon be publicly available on LinkedIn and the Green IO website.

Understanding the Study: Two Models, Two Scales

The study examined two LLMs:

• A smaller model with 3 billion parameters, comparable in performance to LLaMA 3B.
• A larger model with approximately 120 billion parameters, comparable to GPT-4.

Using the MMLU benchmark, the smaller model scored 61, while the larger model scored 85. The functional unit for comparison was the total life cycle of each model—cradle to grave—including even the carbon footprint of the computers used by professionals during fine-tuning.

Carbon Emissions: The Bigger the Model, The Larger the Footprint

The larger model has so far caused the release of around 21,000 tons of CO₂ equivalent. While that might sound massive, it’s still three to four times less than the estimated emissions of GPT-4. To put this in perspective, Netflix alone emitted around 1.5 billion tons per year as of 2021.

At an individual level, however, this footprint can feel frustrating—especially if AI models are used for trivial purposes, potentially wasting the efforts and emissions of thousands of people. Your personal perspective on this number may vary, but one thing is clear: the scale of energy consumption is significant.

Electricity: The Dominant Emission Source and Its Geographic Impact

Electricity use accounts for 86% to 90% of the emissions in this study. This high percentage is largely due to the training location: Nevada, whose electricity grid remains highly carbon-intensive.

The primary recommendation from the audit is to relocate operations to regions with cleaner electricity grids, such as Northern Europe, where emissions could be reduced by a factor of ten. This is a simple yet powerful lever to dramatically reduce AI’s carbon footprint.

Inference Impact: Small but Not Negligible

The study also evaluated the environmental impact of inference—the process of running AI queries, such as chatbot conversations. It accounted for the full life cycle of an inference, including the user’s device (e.g., smartphone).

For a typical small conversation of 800 output tokens:

• The larger model emits approximately 1.14 grams of CO₂.
• The smaller model emits about 0.16 grams.

This is roughly equivalent to sending a simple email, which seems negligible. However, when training emissions are amortized over the number of inferences so far, the numbers change significantly:

• 37 grams of CO₂ per inference for the larger model.
• 70 grams for the smaller model, due to less usage and therefore less amortization.

In other words, inference emissions look small until you consider the heavy upfront training cost. Over time, as models are used more frequently, inference emissions become a smaller part of the total footprint.

Reducing Inference Emissions

How can we reduce the environmental impact of inference? Here are some practical strategies:

• Prompt Engineering: Crafting efficient prompts can reduce unnecessary computations. The greenest prompt, however, is the one you don’t send.
• UX Streamlining: Avoid redundant queries by optimizing user experience and interface design.
• Local Execution: If open-source models are available, running them locally with renewable energy sources can significantly cut emissions.

We are currently in a phase of AI hype where wasteful computing power is common. A more deliberate approach to AI usage can yield substantial environmental benefits.

Water Consumption: An Often Overlooked Factor

Water is a critical but frequently neglected resource in AI sustainability discussions. The larger model consumed approximately 314,000 cubic meters of freshwater, with 60% used for cooling the heat generated by AI computing. This volume is equivalent to about 3,000 Olympic swimming pools.

When distributed over inferences, each conversation uses roughly half a liter of freshwater. While this may not seem large per interaction, the cumulative effect is significant given AI’s rapid growth.

Water Scarcity in Context

Water scarcity is more severe than many realize. To illustrate:

• The largest bubble in the commonly circulated water availability graphic includes all water on Earth—saltwater from oceans and freshwater locked in polar ice caps.
• The second bubble, often mistaken for freshwater availability, includes swamps and other inaccessible water sources.
• The truly available freshwater for human consumption comes from rivers and lakes, represented by a tiny bubble.

Given this reality, the study’s recommendation to relocate AI operations to cool, humid regions like Northern Europe also helps reduce water consumption due to more efficient cooling needs.

Resource Depletion: The Hidden Cost of AI Hardware

Perhaps the most alarming insight is the impact of AI on raw materials. The larger model required 750 kilograms of antimony equivalent (SbEq), a unit measuring resource depletion similar to CO₂ equivalent for greenhouse gases. This doesn’t mean AI uses only antimony; instead, it relies on elements spanning almost the entire periodic table.

Hardware manufacturing accounts for 61% of this resource depletion. Extracting these materials involves moving massive amounts of earth to harvest shrinking quantities of metals and minerals.

For a typical 400-token conversation, the model’s full life cycle consumes about 1.3 milligrams of antimony equivalent—roughly the mass of a two-cent coin. My personal daily AI usage of ten conversations would fit into a small bag of this material.

Hardware Demand and Material Scarcity

The larger model training consumed 25,000 NVIDIA H100 GPUs, compared to 60,000 for GPT-4. These GPUs typically remain in use for 5 to 6 years before being replaced. However, the fast pace of AI development might accelerate hardware refresh rates, increasing demand for scarce materials.

Critical minerals like tin, lead, indium, and many rare earth elements face economic exhaustion within 10 to 20 years. This scarcity threatens not only AI but also essential sectors like healthcare and renewable energy infrastructure.

McKinsey analysis warns that supply cannot keep pace with demand, particularly for rare earths, leading to soaring costs and potential geopolitical conflicts over these vital resources.

Key Takeaways for Sustainable AI Development

Based on this comprehensive assessment, here are practical recommendations for organizations and developers aiming to reduce AI’s environmental footprint:

1. Choose Smaller or Specialized Models: The difference in environmental impact between small and large models is often two orders of magnitude. For tasks that do not require deep reasoning or complex creativity—such as headline generation, sentiment analysis, or spam detection—a smaller model is more sustainable.
2. Leverage Renewable Energy Regions: Train and operate models in areas with abundant renewable energy and cool, humid climates to reduce both carbon and water footprints.
3. Run Open-Source Models Locally: If possible, run AI models on local infrastructure powered by clean energy where you have full control over the energy mix.
4. Apply Software Eco-Design Principles: Optimize software to reduce the number of inferences needed, avoiding unnecessary computation.

Balancing Efficiency with Demand: The Jevons Paradox

Even with these strategies, reducing the environmental footprint is not guaranteed. The Jevons Paradox teaches us that efficiency improvements often lead to increased consumption. As AI becomes more efficient, its adoption and usage may accelerate, potentially increasing overall digital consumption.

For example, the rapid adoption of AI tools like DeepSeek demonstrates how efficiency can stimulate new behaviors and economic shifts, leading to more digital products and services.

Choosing the Future We Want to Accelerate

Ultimately, the question is not just about frugal AI usage, but about the kind of future we want to build. Will AI be used to optimize yesterday’s problems, such as improving fossil fuel extraction, or will it help create tomorrow’s solutions for a sustainable planet?

Our collective goal should be to harness AI for net positive outcomes, ensuring that its power is directed towards environmental and social good. The responsibility to shape this future rests in our hands.

Conclusion

The environmental impact of AI, especially large language models, is becoming impossible to ignore. Through the first comprehensive, data-driven assessment of an LLM supplier’s footprint, we now have clearer insights into the carbon emissions, water consumption, and resource depletion associated with AI development and deployment.

By choosing smaller models, relocating operations to renewable energy-rich regions, optimizing inference, and applying eco-design principles, we can significantly reduce AI’s environmental footprint. However, we must remain vigilant against rebound effects like the Jevons Paradox and ensure that AI accelerates solutions for a sustainable future.

As AI continues to evolve at breakneck speed, so too must our commitment to digital sustainability. The power is in our hands to make AI a force for good.