The true nature of GPT-5: eight times more power consumption

GPT-5 impresses with its power, but behind this technological advance lies an alarming reality: its power consumption could shatter all records. According to estimates, it would consume nearly 8.6 times more energy than its predecessor GPT-4, reaching almost 18.35 Wh per query. With a colossal demand of 2.5 billion daily queries, OpenAI’s AI could soon require as much electricity as two to three operating nuclear reactors. In other words, this breakthrough could transform the global energy landscape, raising questions about the sustainability of our AI consumption on such a vast scale.

GPT-5, OpenAI’s latest model, seems to impress everyone with its incredible power, but one figure particularly stands out: its power consumption, which is said to be up to eight times higher than that of its predecessor, GPT-4. As technology advances by leaps and bounds, it is becoming crucial to examine the environmental impact of these increasingly energy-intensive artificial intelligences.

Figures that speak for themselves

Estimates, although controversial, suggest that GPT-5 consumes approximately 18.35 Wh per query. This places the model in a category of its own, ranking it among the most energy-intensive, just behind OpenAI’s o3 and Deepseek’s R1 model. By comparison, GPT-4 required only 2.12 Wh per query. With a massive volume of 2.5 billion daily queries, energy consumption is at an alarming level, approaching 45 GWh per day.

Consequences of such consumption

To illustrate, this demand would correspond to the energy production of two to three nuclear reactors operating continuously. Such a reality raises ethical and practical questions about the sustainability of large-scale artificial intelligence. In a world already plagued by ecological crises, this voracious appetite can no longer be ignored.

The Assumptions Behind These Figures

OpenAI has not specified precisely the infrastructure on which GPT-5 runs, leading researchers to make assumptions. It is likely that the model relies on cutting-edge systems such as the Nvidia DGX H100 or DGX H200, hosted on Microsoft Azure. These calculations take into account not only the central processing units, but also all the ancillary components, such as storage and cooling systems. A Hybrid Architecture and Its ImpactGPT-5 uses an architecture called

« mixed experts »

which activates different parameters depending on the complexity of the query. While this system can reduce power consumption for simple tasks, it can become a real energy drain for more complex requests. According to recent studies, some difficult queries could exceed 40 Wh. An Uncertain Future Current estimates, however, are subject to rapid change. If OpenAI opts for more powerful hardware like Blackwell, the dynamics between computing performance and power consumption could change considerably. AI data centers in the United States have already seen a significant increase in their electricity bills, and the future looks set to be even more burdensome in terms of energy consumption.

Pragmatic Responses from Developers

This excess consumption could directly influence the strategic choices of artificial intelligence users and developers. The occasional activation of more energy-efficient models, such as GPT-40 **, could potentially reduce environmental impact while still providing a satisfactory level of performance. This change of direction, although considered, may be necessary to reconcile innovation and environmental responsibility.**There is much speculation about the impact of AI and future energy choices. Between the need for performance and the ethical imperative of sustainability, the debate remains open and fascinating. To learn more about these issues, discover the analyses on the energy impact of artificial intelligence in this article, or explore the European gems in this field here.

/limpact-energetique-de-lintelligence-artificielle-et-du-bitcoin-un-comparatif-des-consommateurs-delectricite/ /les-pepites-europeennes-de-lintelligence-artificielle-selon-le-figaro-magazine/

Notez cet article