What is the environmental impact of artificial intelligence?
Contents
What is the environmental impact of artificial intelligence?
L’intelligence artificielle a un effet réel et croissant sur l’environnement, concentré dans les centres de données qui l’entraînent et la font fonctionner. Ces infrastructures représentaient environ 1,5 % de l’électricité mondiale en 2024, soit 415 TWh, et cette part pourrait atteindre près de 3 % en 2030, portée en grande partie par l’IA. S’y ajoutent l’eau mobilisée pour le refroidissement et les ressources nécessaires aux serveurs. C’est précisément pour cela que l’IA devient un sujet central du numérique responsable : l’enjeu n’est pas de renoncer à cette technologie, mais de l’utiliser avec sobriété, en mesurant et en maîtrisant ce qu’elle coûte au climat.
Key Takeaways
- Data centers consumed approximately 415 TWh in 2024 (1.5% of global electricity) and could reach 945 to 950 TWh by 2030.
- The share of AI in this demand could rise from 5–15% today to 35–50% by 2030.
- A query on a GPT-4-type model consumes about 10 Wh, which is roughly 30 times more than a typical online search.
- The cost of AI is multifaceted: electricity, cooling water, resources, and electronic waste.
- Digital frugality involves using AI in a thoughtful way: choosing the right model for the right need, using low-carbon hosting, and measuring actual consumption.
What are the main regulations, standards, and methodologies for GHG accounting?
Where does AI's environmental impact come from?
L’essentiel de l’empreinte de l’IA se situe dans les centres de données. Selon l’Agence internationale de l’énergie (AIE), leur demande électrique augmente d’environ 15 % par an entre 2024 et 2030, soit plus de quatre fois la croissance des autres secteurs. Ces mêmes infrastructures supportent aussi des usages plus banals, et l’impact des mails sur l’environnement en est un bon exemple.
How much energy does a data center consume?
These facilities accounted for approximately 415 TWh in 2024, equivalent to the annual electricity consumption of a country like France. The IEA projects that this figure will nearly double by 2030, reaching around 945 to 950 TWh, and then approximately 1,200 TWh in 2035. The share specifically attributable to AI could triple over this period, rising from 5–15% of the total to 35–50% by 2030. This growth, driven largely by the cloud and online services, explains why the issue has become a priority for both governments and businesses.
How much does an AI query “cost”?
A query on a large language model such as GPT-4 consumes about 10 Wh, compared to about 0.3 Wh for a typical online search—a ratio of about 30 to 1. The initial training of a model is even more energy-intensive: training GPT-3 reportedly required about 1,287 MWh and emitted nearly 552 metric tons of greenhouse gases—the equivalent of the annual emissions from about 100 cars.
A key point for decision-makers: Over the lifetime of a consumer model, daily use (inference) often accounts for a much larger share than the initial training. Increasing the number of unnecessary queries is therefore not climate-neutral.
What about water, natural resources, and electronic waste?
The cost isn’t limited to electricity.Cloud infrastructure uses water for cooling, and the hardware —servers, graphics cards—relies on metals and resources whose extraction itself emits greenhouse gases. At the end of their life cycle, these devices contribute to the growing stream of electronic waste, one of the most difficult types of waste to recycle.
Estimates regarding water usage vary depending on the model and location, but they highlight one key point: AI’s ecological footprint depends heavily on the electricity mix of the country where the servers are located and on cooling conditions. A single query can therefore have a very different environmental impact depending on whether it relies on low-carbon electricity or fossil fuels.
Why Is AI Part of Responsible Digital Technology?
Responsible digital technology aims to reduce the environmental and social impact of digital tools while maintaining their usefulness. AI has become one of the top priorities in this area for three reasons:
- Demand for it is growing faster than for any other digital service;
- its cost is often invisible to the end user, which encourages unnecessary use;
- Conversely, if it is well-targeted, it can generate environmental benefits that outweigh its costs.
A responsible approach, therefore, does not involve banning AI, but rather striking a balance: deploying it where it creates real value, and avoiding it where the benefits do not justify the cost. This is exactly the spirit of sustainable development as applied to technology: balancing utility, moderation, and performance to achieve a more sustainable digital future.
How Can We Reduce the Environmental Footprint of AI?
The principles of simplicity apply directly to AI. They relate primarily to eco-design and the frugal use of resources:
- Choose the right model for the right need: a small, specialized model is often sufficient in situations where a large, general-purpose model would be overkill;
- minimize unnecessary and redundant requests;
- give priority to accommodations in facilities powered by low-carbon electricity;
- reuse and share models rather than retrain them unnecessarily;
- Measure what we actually consume so we can manage it.
📗 Did you know?
There is growing discussion offrugal AI: an approach that seeks to achieve a useful result with the minimum amount of computational resources, energy, and data. For many business needs, a frugal, well-tuned model delivers the same results as a massive model—at a fraction of the cost.
The table below illustrates the energy-saving measures and their effects:
| A Catalyst for Responsible Digital Technology | Expected effect |
|---|---|
| Choose a model that's the right size (not the largest one) | Significant decrease in energy consumption per request |
| Reduce Unnecessary Requests | Direct Decrease in Inference |
| Low-Carbon Accommodations | Decrease in CO₂ intensity per kWh |
| Reuse of Existing Templates | Fewer Expensive Training Sessions |
| Usage Measurement | Possible control and arbitration |
Use Case: A service company with 1,000 employees
A service company that rolls out an AI assistant to its 1,000 employees may see the number of requests skyrocket out of control. A prudent approachinvolves identifying priority cases, selecting models scaled to meet specific needs, and measuring energy consumption—rather than allowing requests with no added value to multiply. This management process then becomes a genuine management issue, just like controlling the electricity bill.
AI can also support the ecological transition
The net impact of AI can be positive when it optimizes processes that generate high emissions. By optimizing delivery routes and the load factor of its fleet, a transportation company simultaneously reduces its emissions and fuel costs. In the construction industry, AI controls heating and air conditioning based on actual occupancy; in the energy sector, it helps forecast renewable energy production and balance the grids.
One regulatory limitation remains: carbon data generated by AI is not auditable. For a GHG inventory, a product carbon footprint (PCF), or enforceable transportation reporting, the calculation must follow a standardized method—ISO 14067 or ISO 14083—based on traceable and reproducible data. AI provides quick estimates and guides decision-making, but it does not replace a method that can be verified by a third party.
FAQ
Is artificial intelligence bad for the environment?
Not by nature. It has a real and increasing energy cost, but it can also generate emissions reductions that exceed that cost when it is well-targeted. The net result depends on how it is used and on the electricity mix.
What is the impact of a query on a generative AI?
About 10 Wh for a query on a GPT-4-type model—roughly 30 times that of a typical online search. The exact carbon footprint depends on the country and the time of day the server is running.
What is responsible digital technology as it applies to AI?
This refers to the set of practices aimed at reducing the AI’s environmental footprint while maintaining its usefulness: choosing the right model, limiting unnecessary queries, prioritizing low-carbon hosting, and measuring consumption.
Should we give up on AI for environmental reasons?
No. The key is to use it in a targeted and frugal manner, reserving the most resource-intensive treatments for cases where the value created truly justifies the cost.
Will AI increase global electricity consumption?
Yes, mechanically speaking. The IEA estimates that demand from data centers will nearly double by 2030, driven largely by AI. That is why energy efficiency is becoming a key management priority in its own right.
At Decarbo’Solution®, we view AI as a catalyst for our business, never as a substitute for methodological rigor. Our carbon footprint calculations remain standardized, traceable, and verifiable—which is what makes them valuable to clients, investors, and commercial@decarbosolution.com | www.decarbosolution.com