How to use AI in a sustainable way
AI may seem like something abstract, happening "in the cloud," but it has very real, tangible consequences.
It consumes vast amounts of energy and water.
Yet, you can’t escape it—AI is all around us and already deeply integrated into our daily lives.
Just as with other aspects of life, you can make conscious choices here too.
What should you pay attention to?
1. Energy Consumption
As a rule: smaller AI models use less energy than large models.
Large models require more training and have more parameters. For example, GPT-4 is estimated to have up to 1.8 trillion parameters, while Gemini Ultra/Pro has up to 100 billion. There are also smaller models optimized for efficiency.
Examples of small models: Gemini Nano, Mistral 7B, Llama 2 7B.
Additionally, there’s a distinction between MoE (Mixture of Experts) models and Dense models:
- MoE models (e.g., Gemini) activate only a specific subset of data per query.
- Dense models (e.g., ChatGPT) scan the entire network for every prompt.
2. Training Efficiency
Most energy is consumed during the training phase of AI models. Therefore, pay attention to the following aspects:
- Training efficiency:
- How much data must be trained before the model is usable?
- Do data centers use efficient hardware and operate in locations powered by renewable energy?
What you can do to make more sustainable choices:
✅ Use smaller models (e.g., Gemini Nano, Mistral 7B, Llama 2 7B).
✅ Run AI locally when possible (e.g., Pixel 8’s Gemini Nano).
✅ Don’t use AI for nonsense, keep conversations short—every token consumes energy.
✅ Demand transparent AI—insight into carbon footprint.
✅ Push for regulation—e.g., the EU AI Act requires energy reporting.
✅ Always choose a lean setup: make sure the model can inspect a specific dataset, and does not have to use the entire network.
✅ Compensate your CO2 emissions: we recommend compensating double the amount of the emissions, based on a worst case scenario of 5 watt-hours per question (=approx 1.35 grams of CO2).