ChatGPT and Its Carbon Footprint

Environmental Impact & Sustainable AI Alternatives

Brought to you by:

Ms. Muskan Sanduja - 2023/08/114
Mr. Saurabh Mittal - 2023/08/076
Ms. Ankur - 2023/08/141
Ms. Muskan Tiwari - 2023/08/146
Ms. Harshita - 2023/08/086
AI’S ENVIRONMENTAL IMPACT
AI has rapidly transformed industries, but its increasing ✅ AI’s Growth & Energy Demand
adoption comes with significant environmental AI adoption is expanding across industries,
consequences. Large AI models like ChatGPT require increasing energy consumption.
massive computational power, consuming electricity and AI models rely on high-performance
contributing to carbon emissions. Understanding and computing (HPC), requiring vast data
addressing AI’s carbon footprint is crucial for sustainable centers.
technological development.
✅ Why Carbon Footprint Matters?
AI-driven cloud computing emits millions of
metric tons of CO₂ annually.
Sustainability concerns arise due to fossil
fuel-dependent energy sources.

✅ Research Focus
Investigating ChatGPT’s energy consumption
and environmental impact.
Exploring sustainable AI alternatives for
lower carbon emissions.
LIFECYCLE OF AI ENERGY CONSUMPTION
1. Training Phase
AI models require extensive computational power during training.
Massive datasets and complex neural networks increase electricity consumption.
Large models like GPT-4 require weeks of processing on thousands of GPUs.
🔴 Impact: Can generate hundreds of metric tons of CO₂ per training cycle.
2. Inference Phase
Once trained, AI models like ChatGPT process billions of queries daily.
Each interaction requires real-time computing, consuming energy continuously.
🔴 Impact: AI usage contributes daily emissions, scaling with user demand.
3. Data Storage & Cloud Dependence
AI relies on global data centers to store and process information.
These data centers operate 24/7, consuming vast amounts of power and cooling resources.
🔴 Impact: Data centers account for ~1% of global electricity use and growing.
LIFECYCLE OF AI ENERGY CONSUMPTION

GPU- Graphics Processing Units Cloud Servers

 ENERGY CONSUMPTION OF CHATGPT

✅ How ChatGPT Processes Billions of Queries Daily

ChatGPT operates on massive cloud infrastructure, running
on thousands of GPUs.
Every user interaction triggers neural network computations,
requiring electricity.
The more queries processed, the higher the carbon footprint.
✅ Comparison: AI Energy Use vs. Traditional Computing
AI models consume exponentially more energy than regular
software.
Example: A single Google Search consumes 0.3 Wh, while an
AI-generated response can use 10x more energy.
Data centers supporting AI require cooling systems, further
increasing power use.
ENERGY CONSUMPTION OF CHATGPT
 COMPARING CHATGPT WITH OTHER AI MODELS

Estimated CO₂ Energy

AI Model Emissions (Metric Consumption Equivalent Real-World Impact
Tons) (MWh)

GPT-2 ~10 ~50 Powering a home for 1 year

GPT-3 ~552 ~1,287 5 cars’ lifetime emissions

GPT-4 ~1,000+ ~2,500+ 1,000+ transatlantic flights

BERT ~284 ~650 60,000+ phone charges

T5 ~200 ~480 200,000 km driven by a car

Case Study: AI-Generated vs. Traditional Content

AI-generated images (DALL·E, Midjourney, etc.) require high GPU power, leading to higher carbon emissions.
Studio Ghibli’s traditional hand-drawn animation consumes far less energy but takes longer to produce.
Text-based AI (like ChatGPT) consumes less energy than AI image or video generation models.
STUDIO GHIBLI’S TRADITIONAL ANIMATION
Hand-Drawn Animation & Minimal Carbon Footprint
Studio Ghibli is renowned for its frame-by-frame hand-drawn animation,
requiring skilled artists instead of power-hungry AI models.
Unlike AI-generated digital art, traditional animation doesn’t rely on high-
energy GPU clusters.
Example: "Spirited Away" (2001) had over 112,000 hand-drawn frames,
created with minimal environmental impact compared to AI-generated
visuals.

Craftsmanship & Sustainability in Traditional Art

Ghibli’s artistic approach focuses on human creativity rather than data-
intensive AI models.
Traditional methods use paper, ink, and watercolors, which have lower energy
consumption than AI requiring massive cloud-based GPU computing.
Studio Ghibli’s philosophy promotes slow, sustainable art creation, ensuring
eco-friendly production practices.
 AI-GENERATED ART AND ITS CARBON IMPACT

AI-Generated Art (DALL·E, Traditional Digital Rendering Traditional Digital

Factor Factor AI-Generated Art
MidJourney, etc.) (Photoshop, Blender, etc.) Art

Energy High (due to GPU-intensive Moderate (depends on project Energy per Image ~0.02 - 0.05 kWh ~0.2 - 1.0 kWh
Consumption model inference) complexity)
CO₂ Emissions per
~10-50g CO₂ ~100-500g CO₂
CO₂ Emissions Estimated 2-5x higher than Lower, but increases with high-resolution Image
per Image manual design rendering
Water Usage (for Minimal (local ~2-5 liters (data
cooling) machine) centers)
Computational Requires cloud-based AI models, Runs on local machines or smaller cloud-
Requirements data centers based servers
Hardware Power ~2,000-3,000W (AI
~100-300W (PC)
Usage server)
Large-scale GPUs (e.g., NVIDIA
Hardware Usage Consumer GPUs (RTX 4090, M1 Ultra)
A100, H100)

Water
Consumption High (for data center cooling) Lower, mostly air-cooled workstations
(Cooling)

Sustainability Some providers use renewable Depends on user’s hardware and energy
Measures energy source
 SUSTAINABLE AI AND FUTURE DIRECTIONS

Energy-Efficient AI Models
Smaller, task-specific AI models
Model pruning & quantization to reduce computation
Low-power AI chips (e.g., ARM-based, neuromorphic
computing)

Renewable Energy for AI

AI data centers powered by solar, wind, and hydro energy
Carbon-neutral cloud computing initiatives (e.g., Google,
Microsoft)

Policy Recommendations
Regulations on AI energy use and efficiency
Carbon footprint transparency in AI development
Incentives for green AI research
 ALTERNATIVE AI SOLUTIONS FOR SUSTAINABILITY

Green AI Innovations
AI models designed for minimal energy consumption
Open-source lightweight AI frameworks (e.g., TinyML)

Decentralized AI Computing (Edge AI)

AI processing on local devices (reduces cloud dependency)
Uses lower-power hardware (smartphones, IoT devices)
Reduces data transfer energy costs

Optimization Techniques for Lower Energy Usage

Model compression (pruning, quantization)
Efficient training techniques (transfer learning, low-rank adaptation)
Hardware acceleration (AI-specific chips like TPUs, neuromorphic processors)

AI’s Role in Sustainability

Smart energy grids (AI-driven power distribution)
AI for climate modeling & carbon tracking
Sustainable supply chain optimization
CONCLUSION
AI is a powerful yet resource-intensive technology, requiring significant energy for training and inference. However,
advancements in energy-efficient models, decentralized computing, and renewable-powered data centers can help mitigate its
environmental impact. Responsible AI development, combined with sustainable policies and industry innovation, is essential to
balance progress with ecological responsibility. By prioritizing green AI practices, we can harness AI’s potential while minimizing
its carbon footprint.

REFERENCE
Bolón-Canedo, V., Morán-Fernández, L., Cancela, B., & Alonso-Betanzos, A. (2024). A review of green artificial intelligence:
Towards a more sustainable future. Neurocomputing, 570, 126104. https://doi.org/10.1016/j.neucom.2024.126104
Verdecchia, R., Sallou, J., & Cruz, L. (2023). A systematic review of Green AI. arXiv preprint arXiv:2301.11047.
https://arxiv.org/abs/2301.11047
van Wynsberghe, A. (2021). Sustainable AI: AI for sustainability and the sustainability of AI. AI and Ethics, 1(3), 213–218.
https://doi.org/10.1007/s43681-021-00043-6
Yokoyama, A. M., Ferro, M., de Paula, F. B., Vieira, V. G., & Schulze, B. (2023). Investigating hardware and software aspects in the
energy consumption of machine learning: A Green AI‐centric analysis. Concurrency and Computation: Practice and Experience,
35(19), e7825. https://doi.org/10.1002/cpe.7825
Hao, K., Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. Proceedings
of the 57th Annual Meeting of the Association for Computational Linguistics, 3645–3650. https://doi.org/10.18653/v1/P19-1355
Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L. M., Rothchild, D., & Dean, J. (2021). Carbon emissions and large neural
network training. arXiv preprint arXiv:2104.10350. https://arxiv.org/abs/2104.10350