Published on
reading min
Reduce your consumption with Selectra!
Reduce your electricity bills by contacting one of our advisers:
The same electricity, cheaper! Save by changing their offer with selectra
Compare the offers for free with an advisor
Behind the intelligence of GPT-5 hides an insatiable thirst for energy. A new study by researchers from the University of Rhode Island reveals the colossal environmental bill of the new OPENAI model, while report by the International Energy Agency (AIE) confirms an explosion of global electrical demand drawn by AI. On a scale, the impact is dizzying: the annual consumption of the Chatgpt ecosystem, powered by GPT-5, could equal to that of 1.6 million French households.
Summarize this article with:
GPT-5, up to 10 times more energy-consuming than its predecessor GPT-4O
The technological advance of GPT-5 has a very real and measurable energy cost. According to the study, entitled “How Hungry is ai?”, Only one medium size request (around 1000 words) addressed to the high performance version of GPT-5 consumes on average 19.32 Watthers (WH).
This figure is to be compared to the 1.80 Wh necessary for a similar request on GPT-4O (version of May 2024), that is more than 10 multiplication of energy consumption.
For a simple short question, the gap is just as striking: 9.15 Wh for GPT-5 against only 0.48 Wh for GPT-4O. Each interaction with the new model is therefore structurally more expensive for the planet.
| Model | Short question (~ 300 words) | Average question (~ 1000 words) | Long question (~ 1500 words) |
|---|---|---|---|
| GPT-5 (haute performance) | 9,15 Wh | 19,32 Wh | 29,01 Wh |
| GPT-5 (average performance) | 4,31 Wh | 10,28 Wh | 18,81 Wh |
| GPT-5 (basse performance) | 2,28 Wh | 5,95 Wh | 11,71 Wh |
| GPT-5 mini | 1,82 Wh | 3,78 Wh | – |
| GPT-4O (MAI 2024) | 0,48 Wh | 1,80 Wh | 6,20 Wh |
| GPT-4 Turbo | 1,45 Wh | 4,28 Wh | 7,05 Wh |
| GPT-4 | 1,75 Wh | 5,23 Wh | – |
| GPT-3.5 Turbo | 0,54 Wh | 1,63 Wh | – |
Source: Study “How Hungry is Ai?”, Jegham et al. (2025). Data extracted from the metrics of the study.
If GPT-5 stands out as one of the most powerful models, it is also among the most energy-consuming on the market. An average request on the high performance version of GPT-5 (19.32 Wh) is:
- 67 % more energy -consuming that the most advanced model of its main anthropic rival, Close 4th work (11,53 Wh).
- Almost 3 times more gourmet that the model Grok 4 de xAI (6,54 Wh).
- More than 128 times more consumer that the most sober open-source model in Meta, Call 3.1 8b (0,15 Wh).
The only general public model that surpasses the consumption of GPT-5 is the DeepSeek R1 From the Chinese company Deepseek, which reaches 32.65 Wh for an average request, especially due to less optimized data centers.
What about the consumption of French Mistral AI?
Often presented as a European alternative focused on energy sobriety, the French startup Mistral AI positions itself favorably in this study. Its most efficient model, Mistral Large 2consumes 4,88 Wh For an average request. It is Almost 4 times less than the high performance version of GPT-5 (19.32 WH).
Its smaller models, as Mistral Medium 3 (3.44 Wh), are even more economical. These figures confirm that Mistral offers a significantly less energy -consuming alternative, although its most advanced reasoning capacities are still behind compared to those of the most powerful models in Openai.
Data centers and AI: global electricity consumption explodes
The case of GPT-5 is part of a substantive trend: the explosion of the energy consumption of data centers, driven by artificial intelligence. The World Information Technology Sector (ICT) now represents around 4 % of global electricity consumption. The share of data centers becomes more and more important, reaching approximately 1.5 % of global electricity totalor about 415 Terawattheures (TWH) In 2024 according to the latest report by the International Energy Agency (AIE).
Doped by generative AI, this consumption increases on average 12 % per year since 2017. According to the IEA, it should more than double by 2030 to reach approximately 945 TWh per year, the current consumption of Japan. On this date, the data centers will almost represent 3 % of global electricity consumption.
The annual impact of GPT-5: the equivalent of 1.6 million French households
To measure the concrete impact of these figures, it is possible to project the consumption of the GPT-5 ecosystem on a scale of massive use. Based on a realistic distribution of requests between the different versions of GPT-5, the annual figures remain astronomical:
What are these projections based on?
These figures are a weighted estimate. Rather than assuming that all requests use the most energy -consuming model, we have modeled a more realistic distribution: 10% of requests on the “high performance” model, 30% on the “medium” and 60% on “low performance”. This projection is calculated on the basis of:
- Consumption by request For each version of GPT-5, resulting from the study “How Hungry is Ai?”.
- A hypothesis of 2.5 billion daily queriesa volume of use often cited for the Chatgpt ecosystem.
Why is GPT-5 such an energy chasm?
Several factors explain this massive consumption. First of all, model size. Although Openai no longer communicates on the number of parameters, experts believe that GPT-5 is several times larger than GPT-4. A larger model requires more computing power for each operation.
Secondly, The complexity of tasks. GPT-5 is designed for advanced reasoning tasks (“Thinking Mode”), which extend the calculation time before providing a response. A long and complex request (1500 words) can thus consume up to 29 Wh. It is more than a 10W LED bulb on for almost 3 hours.
Finally, even if GPT-5 uses more recent and effective equipment (NVIDIA H100/H200), the extent of its capacities and its use cancels these efficiency gains, a phenomenon that researchers associate with Jevons paradox.
Find a cheaper supplier!
Get your comparison by phone with Selectra
The same electricity, cheaper! Save by changing their offer with selectra
Compare the offers for free with an advisor
An urgent call for transparency for a sustainable AI
Researchers at the University of Rhode Island and AIE underline a fundamental problem: the lack of communication of tech giants on the environmental impact of their products. Without public and standardized data, it is impossible for users and regulators to make informed decisions.
While artificial intelligence is increasingly integrated into our daily lives, its invisible energy bill continues to climb. The launch of GPT-5 marks a turning point, forcing us to confront the question: are we ready to pay the environmental price of this technological revolution?
Faced with this challenge, France wants to attract low-carbon data centers
In early March 2025, EDF launched two calls for events of interest (friend) in order to offer digital companies fields perfectly suited to the construction of data centers in France. These sites are particularly attractive because they already have a connection to the national electrical network, making it possible to considerably reduce the usually very long installation deadlines.
During the AI summit in Paris last February, President Emmanuel Macron defended EDF’s interest in data centers:
In this world, I have a good friend on the other side of the ocean (Atlantic) who says “Drill, Baby, Drill”. Here there is no need to drill. Just plug, baby, plug! Electricity is available. You can plug. It’s ready
Among the first locations offered by EDF are the site of Montereau-Vallée-de-la-Seine (Seine-et-Marne) as well as the La Maxe and Richemont (Moselle). A fourth site is still confidential, while two others should be identified by 2026, bringing the total to six strategic sites spread over French territory.
Reduce your consumption with Selectra!
Reduce your electricity bills by contacting one of our advisers:
The same electricity, cheaper! Save by changing their offer with selectra
Compare the offers for free with an advisor
