One in four French adults over 18 now uses generative AI daily. Three-quarters turn to it at least once a week.
Behind this rapid adoption lie questions about energy costs, well documented, but also about its relationship with performance.
It is precisely to this question that the PEReN, the Regulating Digital Regulation Expertise Center, tackled the issue in a technical study conducted in partnership with Arcep and published as part of its 2025 work program.
Result: the conventional wisdom about the energy consumption of AI models is being challenged.
The first belief debunked: the size of an AI model is not a reliable indicator of its energy consumption. Admittedly, the most energy-intensive models are indeed those with the largest number of parameters. But the inverse is not true: some models consume as much, or even less, than much “smaller” ones.
The key lies in the Mixture of Experts architecture (MoE). These models are designed to activate, for each request, only a fraction of their total neural network. The remarkable result measured by the PEReN: MoE models consume on average 45% less than classical (dense) models that have the same total number of parameters. A significant efficiency lever, still largely underutilized in the debate about AI’s footprint.
Quantization is also meaningful: moving to 8-bit mode yields an average gain of 39% in consumption, according to measurements performed on the GENCI Jean Zay supercomputer.
The reasoning comes at a high price
The other major finding concerns so‑called “reasoning” models, trained to break down complex problems into several steps before answering. These models are today at the heart of the commercial strategy of major AI players. But their energy cost is substantial.
PEReN measured that, on average, turning on the reasoning mode increases consumption by 92%. And sometimes considerably more: on code generation tasks (the HumanEval benchmark), the average effect rises to +849% compared with a standard model. In plain terms, asking a model to “think” before answering can multiply its energy bill by nearly ten.
The problem is that this extra consumption does not always translate into an equivalent improvement in quality. On short factual questions, the benefit of reasoning is limited: a +41% performance improvement for +41% more consumption on the SimpleQA benchmark. “A user of an LLM should ensure in advance that such a model is suited to their use case, or that the reasoning mode is disabled before use,” concludes the report.
Specialization, a false good idea?
The study also looks at specialized models—those versions trained specifically for medicine, law, or programming. The conclusion is striking: the specialized models studied do not guarantee better performance on tasks in their domain, at least among open-weight models available today.
Two factors explain this paradox. First, a specialized model may be trained on a task too narrow to generalize well across its entire field of specialization.
Second, and this is the clinching argument, the pace of innovation in generalist models is such that specialized versions, published less frequently, are quickly overtaken by newer generalists. For a user seeking an open-source model, it may therefore be preferable to opt for the most recent generalist.
Good news for users of models capable of handling both text and images. On purely textual tasks, multimodal models show energy use and performance comparable to their 100% text counterparts. The extra cost only appears when images are actually fed to them. And there, consumption can double.
Performance and sobriety are not at odds
Perhaps the study’s most counterintuitive and promising takeaway.
Regardless of specialization, multimodality, or the activation of a reasoning mode, a model’s performance is not a monotonic function of its energy consumption. In other words: a lean model is not necessarily a worse model. And a highly energy‑hungry model is not necessarily the most performant.
PEReN researchers tested 23 models from more than five different providers (Alibaba, Google, Meta, Mistral, Z.ai among others) across six benchmarks representing typical public use. In total, the experimental set generated 0.229 tCO2e in CPU and GPU costs.
On the compared criteria, the gap in consumption between the leanest and the most energy-intensive models reaches a factor of four to five at constant hardware. A gap of similar magnitude, note the authors, to running the same model on two different GPU infrastructures.
What this study does not reveal yet
PEReN is the first to acknowledge the limits of its work. The study is restricted to open-weight models, while major proprietary models such as GPT or Gemini in full version remain opaque.
Moreover, only three specialized models were included—an insufficient number to draw definitive conclusions. Finally, the benchmarks, no matter how rigorous, do not always reflect the realities of everyday interactions with an LLM.
The study also omits the entire lifecycle of AI: chip fabrication, data center construction, water usage, and the initial training of models. These items, in some cases, account for the bulk of the overall environmental footprint.
(1) General knowledge questions, medical reasoning, code generation, creative writing, image understanding, physical reasoning.
