AI’s Environmental Impact

What Every Developer and User Should Know

When I first started using AI tools, I was amazed by their power. But I never stopped to ask: What’s the real cost of all this intelligence?

Turns out, the answer is both fascinating and urgent. Let’s take a journey through the hidden world of AI’s environmental footprint, and discover what we can do about it.

---

The Big Picture: Why AI’s Footprint Matters

AI is everywhere, from chatbots to image generators to code assistants. But every query, every model, every “aha!” moment comes with a real-world cost: energy, water, and carbon emissions. And as AI scales, so does its impact.

Did you know? Training a single large AI model can use as much water as a small town’s daily needs, and emit as much CO₂ as 100+ cars over their lifetimes ^{1, 2}.

---

Under the Hood: Why AI Needs So Much Energy and Water

Let’s peek inside the black box:

• Model Size: Today’s language models have billions (sometimes trillions) of parameters. Think of each as a tiny dial the model can adjust. Training them means running through massive datasets, again and again, to find the best setting for every dial ^{1, 2, 3}.
• Training: At the heart of deep learning are matrix multiplications, huge grids of numbers being multiplied together. For transformers (the architecture behind most modern AI), this means not just multiplying, but also calculating “attention” scores: for every word or token, the model computes how much it should pay attention to every other word in the sequence. This attention mechanism is powerful, but it means the number of calculations grows rapidly with input length ².
• Inference: When you use an AI model, it generates text one token at a time. For each token, your input and all previous tokens are passed through dozens or even hundreds of neural network layers. Each layer involves matrix multiplications, nonlinear activations, and normalization steps. For a model with billions of parameters, these operations are repeated billions of times for every user query ^{1, 2, 3}.
• Memory Access: Every calculation involves reading the model’s parameters (weights) from memory. For very large models, just moving this data around inside the server is a significant energy cost, sometimes as much as the calculations themselves.
• Hardware Utilization: Training and inference are performed on specialized hardware (GPUs, TPUs) designed for parallel processing. These chips are extremely powerful but also energy-hungry. To serve millions of users, data centers keep these chips running 24/7, often with many running below full capacity to ensure low-latency responses ³.
• Cooling: All this computation generates heat. Data centers use advanced cooling systems, often water based, to prevent overheating. Evaporative cooling is efficient but consumes water that cannot be reused, and the hotter the climate, the more water is needed ^{1, 4}.

Analogy: Imagine a stadium full of people (parameters), each doing math with every other person (attention), all at once, for every word you type. Now imagine keeping the stadium air-conditioned, day and night, so the math never stops. That’s the scale of modern AI.

Pro tip: The bigger the model, the bigger the footprint. But smart engineering, like using smaller models, quantization, batching, and efficient hardware, can cut energy and water use by up to 90% ^{1, 3}.

---

The Real Numbers: How AI Compares

• A single AI query can use 10 to 50x more energy than a Google search ^{1, 2}.
• Training a large model can use as much energy as thousands of households, and as much water as a small town ^{1, 3}.
• Inference (serving users) now accounts for 60 to 90% of a model’s total emissions, because it happens millions of times ^{5, 6}.

---

Foundation Models: The Heavyweights

• GPT-4 (OpenAI): Training estimated at 12,000 to 15,000 tCO₂ ⁵.
• Llama 2/4 (Meta): Llama-2: 539 tCO₂; Llama 4 Scout/Maverick: 1,999 tCO₂ ^{5, 7}.
• Gemini (Google): No public data, but Google’s total emissions and water use are rising fast ^{3, 8}.
• Claude (Anthropic): No public data; claims eco-efficiency ⁵.

Takeaway: The biggest models are also the biggest emitters. But smaller, distilled, or quantized models can do the same job with a fraction of the impact ^{1, 3}.

---

Why Does Inference Dominate AI’s Environmental Impact?

• Training is a one-time cost, inference is ongoing: Training happens once, but inference (serving users) happens millions or billions of times ^{1, 2, 5, 6}.
• Scale: Popular models serve millions daily. The cumulative effect quickly surpasses training ^{5, 6}.
• Hardware: Inference requires powerful GPUs running 24/7, even when not fully used ^{3, 5}.
• Model size: Each inference runs the full model, billions of parameters, every time ^{1, 2, 3}.

Bottom line: As AI gets more popular, serving users becomes the main source of emissions. That’s why efficient inference matters so much.

---

The Providers: Who’s Doing What?

Google/Alphabet

• Emissions: Up 13% in 2023; total GHG emissions up 1,500% since 2010 ^{3, 8}.
• Water: 11 billion gallons withdrawn in 2024 ⁸.
• Energy: 100% renewable matched, but absolute use rising ³.
• Transparency: No model-specific data for Gemini/Gemma ⁵.

Microsoft (Azure/OpenAI)

• Emissions: Rising sharply; missed interim sustainability targets ⁹.
• Water: Increasing with AI expansion ⁹.
• Transparency: No model-specific data for GPT-4; estimated 12,000 to 15,000 tCO₂ ⁵.

Meta

• Emissions: Scope 3 and water use up ⁷.
• Transparency: Most open, publishes Llama-2/4 emissions ^{5, 7}.

Anthropic

• Emissions/Water: No public data; claims eco-efficiency ⁵.

Others (DeepSeek, Mistral, Cohere)

• Transparency: Generally low; Mistral-7B estimated at ~10 tCO₂ ⁵.

Pro tip: Choose providers with transparent, independently audited sustainability reports ^{3, 7, 5}.

---

Architectures & Approaches: What’s Greener?

• Standard LLM API: High per-query cost, especially with large models ^{1, 3}.
• RAG (Retrieval-Augmented Generation): Can reduce repeated LLM calls, but embedding/indexing is resource-intensive ^{1, 3}.
• Multi-Agent Systems: Can multiply resource use if poorly designed; specialized agents can be more efficient ⁷.
• Fine-Tuning: Full retraining is costly; parameter-efficient fine-tuning (LoRA, QLoRA) is much greener ^{1, 3}.
• Local Inference: Zero cloud water use, much lower energy per query, great for repetitive or privacy-sensitive tasks ^{1, 3}.

Best practice: Use the smallest, most efficient model that meets your needs. Batch queries, cache results, and pick green cloud regions ^{1, 3, 6, 10}.

---

What You Can Do: Action Steps for Developers & Users

For Developers

1. Benchmark and select efficient models ^{1, 3}.
2. Use green cloud regions ^{4, 3, 6, 10}.
3. Batch and cache queries ³.
4. Monitor your emissions (try CodeCarbon) ³.
5. Advocate for transparency by asking your provider for real data ^{3, 7, 5}.

For Users

• Prefer efficient model options when available.
• Batch your queries and use local tools for simple tasks ^{1, 3}.
• Support providers who publish real environmental data.

---

Quick Reference Table

“High” and “Very High” are relative to current industry standards. See references for data.

---

Final Thoughts: Building a Greener AI Future

AI is powerful, but with great power comes great responsibility. As developers and users, we have real choices, and real influence. By demanding transparency, choosing efficient tools, and advocating for responsible AI, we can help steer the industry toward sustainability.

What’s your experience with sustainable AI? Have you considered the environmental impact of your projects? I’d love to hear your thoughts and ideas for building a greener digital world.

---

References

1. 1

   Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren, Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv, 2023, revised 2024.

2. 2

   Emma Strubell, Ananya Ganesh and Andrew McCallum, Energy and Policy Considerations for Deep Learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019, pages 3645 to 3650.

3. 3

   Google, 2024 Environmental Report.

4. 4

   Arnault Pachot and Céline Patissier, Towards Sustainable Artificial Intelligence: An Overview of Environmental Protection Uses and Issues. arXiv, 2022.

5. 5

   DitchCarbon, The Real Carbon Cost of an AI Token, 2025. Model by model estimates of training emissions.

6. 6

   Carbone4, Is Generative AI Generating Climate Change?, 2025.

7. 7

   Meta, 2024 Sustainability Report.

8. 8

   Kairos Fellowship, No Climate Results Found, 2024. On the rise in Google's emissions and water use.

9. 9

   Microsoft, 2024 Environmental Sustainability Report.

10. 10

    Universal AI University, AI's Environmental Footprint: The Energy and Water Consumption of Leading AI Models, 2025.