The environmental impact of generative AI like ChatGPT is a concern, but their actual energy consumption remains difficult to assess. Are the prompts generated by the chatbot really that energy-intensive?
On April 16, on X, a user asked, “I wonder how much money OpenAI has lost in electricity costs because of people saying ‘please’ and ‘thank you’ to their models.” Sam Altman responded with a touch of humor, saying it represented ” tens of millions of dollars well spent—you never know .” This was enough to reignite the discussion on the environmental cost of generative AI. If simple, polite phrases can generate such electricity costs, what about chatbots as a whole? Here are some answers.
Energy Consumption on ChatGPT: What Are We Talking About?
Every time a query is made on ChatGPT, powerful servers are called upon. Hosted in data centers around the world, these computing-intensive machines run continuously and require constant power, which generates greenhouse gases.
This consumption includes not only the calculations performed, but also the cooling required to prevent overheating. These air conditioning systems, which are very energy-intensive, increase the carbon footprint. Added to this is the transfer of data through digital infrastructures, which are also powered by electricity.
The environmental impact of ChatGPT therefore rests on two levels:
- Design-related pollution, concentrated at the time of model training, represents a massive but one-time cost. It is independent of end users.
- Conversely, usage-related pollution, repeated with each interaction, accumulates over time. Users are directly responsible for it. But in a way, it offsets the energy consumption generated by the model’s construction.
ChatGPT: The environmental impact of model training
Over the past four years, a number of studies (of varying degrees of rigor) have attempted to estimate the energy consumed to create ChatGPT. As early as 2021, a joint study conducted by researchers at Google and the University of Berkeley estimated that training GPT-3, with its 175 billion parameters, required approximately 1,287 MWh of electricity, equivalent to the annual consumption of 120 American homes.
And subsequent models only exacerbated this energy consumption. Data scientist Kasper Groes Albin Ludvigsen, for example, estimated in a Medium article that training GPT-4 would have consumed between 51 and 62 million kWh, 40 to 48 times the consumption of GPT-3. This represents between 24,600 and 29,600 tons of CO₂, or the equivalent of approximately 12,300 to 14,800 round trips between Paris and New York.
Consumption of a prompt: how much does a question weigh in ChatGPT?
If the models used by ChatGPT prove to be particularly polluting in production, what about their daily use? Some pessimistic estimates seem to indicate that conversations with users are particularly costly in terms of energy. In March 2024, in Le Figaro, sustainable finance expert and professor at ESSCA Dejan Glavas stated in particular that a query made on ChatGPT would emit “sixty times more carbon” than a Google search. This data alone is enough to dissuade people from saying “thank you” to the chatbot.
However, assessments of the actual impact of a prompt vary, and a number of them include, for example, the environmental cost of training the models, which can distort the result. In a publication dated February 7, 2025, the independent research institute Epoch AI, which analyzes the evolution of artificial intelligence and its impacts on society, questions certain estimates, which it considers overestimated. According to its calculations, a typical query addressed to ChatGPT would consume approximately 0.3 watt-hours, ten times less than the figures regularly put forward. This difference is explained by the improvement of the models and chips used, and by a more realistic reassessment of the number of tokens generated during a typical interaction. The figure thus remains marginal compared to the daily consumption of an average American household, estimated at more than 28,000 watt-hours per day.
For comparison, 0.3 watt-hours is less than the amount of electricity consumed by an LED light bulb or a laptop in a few minutes. And even for a heavy user, ChatGPT’s energy cost is only a small fraction of the overall electricity consumption of a person in a developed country, the report states.
Lightweight models: a more sustainable alternative?
A large part of the consumption of generative artificial intelligence is therefore due to the resources used to train the models. So, are companies able to limit resource usage while preserving performance? In this regard, DeepSeek seems to have led the way. While the trend was towards increasingly heavy models, the Chinese company managed to make a name for itself thanks to lighter models, which often display performance close to that provided by ChatGPT.
To achieve such a feat, DeepSeek combined several strategies:
- Rigorous sorting of training data: DeepSeek favored structured, high-quality datasets (clean code, coherent texts, technical documentation), which allowed the model to learn more efficiently with less data.
- An architecture optimized for efficiency: The model uses a tuned structure (e.g., local or semi-local attention), which reduces computational complexity while maintaining strong performance on targeted tasks.
- Using advanced compression techniques: DeepSeek uses distillation, where a small model learns to mimic a larger one; pruning, which removes unnecessary parts of the neural network; and quantization, which reduces the precision of calculations (for example, from 32 to 8 bits) to gain speed and lightness.
- Specialization on specific tasks: DeepSeek focuses on areas such as logical reasoning, code generation or question answering, which allows it to optimize performance without aiming for too much versatility.
Lightweight models thus provide a relevant path, but environmental gains still face performance limitations. The launch of GPT-5, expected this year, will reveal whether the power gap observed between GPT-3 and GPT-4 remains the norm, or whether another approach may emerge; in which case we can say “thank you” to OpenAI.
