Why is AI’s environmental impact becoming a major issue?
AI’s environmental footprint remains difficult to isolate. However, estimates from the International Energy Agency (IEA) suggest that global electricity consumption from data centers could reach between 700 and 1,700 TWh by 2035, depending on the scenario.
By comparison, France currently consumes 451 TWh per year (All of France's electricity data in real time). This potential increase is driven by the rapid adoption of AI and by intensive use cases, especially real-time applications or video generation, which are the most energy demanding.
The issue is becoming critical as computing capacity concentrates among a small number of providers and in certain data centers, creating localized energy pressure points.
How do we define frugal AI, and why is it becoming essential?
We often talk about sustainable AI or responsible AI, but the concept of frugal AI, as defined in the AFNOR framework I collaborated on on, aims to reduce the environmental footprint of artificial intelligence to the strict minimum while delivering a positive outcome for the environment. This applies both to how systems are designed, which should be efficient by default, and to how they are used.
When we look at environmental impact, the question of rebound effects is central. It is about ensuring that making AI available does not in fact create additional impacts. The origin of the word itself reflects the idea of not taking more than the Earth can regenerate, a principle that helps guide this approach.
What questions should we ask to make better AI choices?
The first question is simple. Can we do without AI? Many use cases can rely on lighter solutions, including applications that are just as effective. There is no need to use a large model like GPT-5 or Claude Sonnet 4.5 for tasks such as translation. It is often oversized.
Then it becomes a question of trade-offs. You need to balance efficiency and performance with social and environmental considerations. When AI does make sense, it is important to choose the smallest model possible. In many cases, it is sufficient and far less energy intensive. Frugality is based on informed decisions and thoughtful trade-offs. It is about choosing the right tool for the right use.
How is Orange putting more responsible AI into practice?
Our Group relies on the open-source EcoLogits tool to measure the carbon impact of AI queries. Not all generative AI models have the same impact. This tool makes them comparable and allows employees to assess the footprint of each use. Our internal AI tool, Dinootoo (Live Intelligence), displays the carbon impact of each query as well as the monthly total. This helps raise awareness and supports better decision-making. The company has also defined eco-design principles to guide data scientists in building more efficient AI systems.
Another key initiative is the AI Carbon Value Navigator, a dedicated management tool that supports decision-making around integrating AI into our products and services. It compares potential financial gains with the carbon emissions generated and helps ensure that each AI solution aligns with Orange’s net zero target by 2040. It also enables reporting from project level up to the Executive Committee. This provides a clear and structured way to assess carbon impact alongside business value.
How are more frugal AI practices promoted across the Group?
The shift toward frugal AI was first made possible through strong leadership engagement. Leaders took ownership of the topic and helped drive it across the organization. Training is also guides the process. Some employees follow dedicated training on frugal AI, while others are trained on Dinootoo, our internal generative AI tool, which includes a focus on environmental impact. This tool allows everyone to track the impact of their AI usage and receive guidance on the most appropriate models. Integrating these topics into communication, tools, and training helps strengthen awareness.
This is complemented by eco-design principles for teams working with data and developing AI models. We also organize major initiatives such as the AI Boost Camp, two days fully dedicated to artificial intelligence, with keynotes, conferences, and training sessions. Frugality is becoming a shared mindset across the organization, from the Executive Committee to operational teams.
What habits support more frugal AI in daily use?
The first is to always use the smallest models when possible. For decision-makers, it is important to integrate carbon into their decision-making. AI can also help reduce the carbon impact of a product, a service, or an organization. This is the case for the Orange network, for example. The different LLMs we use support the development of AI solutions that improve operational efficiency and open up new possibilities in customer experience and smart networks.
However, it is important to keep in mind that generative AI, in some intensive use cases, may not align with environmental transition goals. When designing a product, generative AI should not be considered the only option. Traditional AI approaches, such as expert systems or machine learning, can deliver better results while being more efficient and less costly over time. In practice, frugality can also drive positive business impact.
