Meanwhile,
Google finally reveals how really:
1 km by electric car is equivalent to more than 600 IA chatbot requests. Furthermore, Here is what we can deduce from the results published by Google.
While the debate on the environmental impact of artificial intelligence swells, Google comes out of its silence with precise figures on the energy consumption of its Gemini models. In addition,
The Mountain View company announces that answering a question via Gemini consumes 0.24 watt-hour energy. In addition, To give you a concrete idea: it is the equivalent of watching television for less than nine seconds. Therefore, A figure that may seem derisory. For example, but which takes on another dimension when you know google finally reveals how really that millions of users question these systems every day.
- 8 seconds of television (as announced by Google)
- 30 seconds of LED lighting (ampoule 8W)
- 2 minutes of smartphone load standby
- 1 minute of operating a digital alarm clock
- 15 seconds of microwave minimal power
- 45 seconds of use of an electric toothbrush
- One hour on Netflix = 150-200 Wh = 625 to 830 IA requests
- An electric car route (1 km) = 150-200 Wh = Over 600 IA requests
. Similarly,
A traditional Google search consumes approximately 0.2 to 0.3 watt-hour. Similarly, So an IA Gemini request (0.24 Wh) would almost consume the same as a classic web research. Moreover, It is rather surprising when you think of all the computing power required to generate a complex response to. Furthermore, display a list of links.
Spectacular progress … Similarly, according to Google – Google finally reveals how really
Google google finally reveals how really engineers claim impressive efficiency gains over the last twelve months. Their gemini model would now consume 33 times less energy per request than in May 2024. while reducing its carbon emissions by 44 times. These improvements result from technical optimizations at all levels: clearer algorithms, more effective tailor-made fleas, and ultra-optimized data centers.
« Our latest generation Ironwood tpus are 30 times more energy efficient than our first specialized processors »explain Amin Vahdat and Jeff Dean, respectively vice-president of Google Cloud and chief scientist at Deepmind. This race for efficiency is based on advanced engineering techniques such as the “mixture-of-experts”. which activates only a small part of the model for each request.
Google would stand out by offering a more complete calculation approach than its rivals. google finally reveals how really Where most estimates have only the consumption of fleas in operation. the Californian firm incorporates all the elements of the chain: servers in standby (necessary to guarantee availability), consumption of conventional processors and memory, plus the complete infrastructure of data centers.
This “understanding” method leads to figures 2.4 times higher than the usual calculations. If Google had just measured its active processors only, a Gemini request would only have consumed 0.10 Watt-hour. “” Many current AI energy consumption calculations represent only theoretical efficiency rather than real operational efficiency on a large. scale ”underline the researchers.
Self-declared data to be taken with care
These announcements are timely for Google. while the energy consumption of AI becomes an issue for public opinion and regulators. But we must keep in mind that these figures come directly from the company, without independent validation. Google communicates what it wants to reveal on its own google finally reveals how really systems.
Moreover, the absence of comparable data at Openai, Anthropic or Meta makes any perspective difficult. By publishing its detailed methodology. Google seems to want to impose its measurement standards on industry, while demonstrating its supposed technological advance. A strategy that allows it to control the narrative around the environmental impact of the AI.
This selective transparency raises legitimate questions. Why these revelations now? How to ensure that the data reflect the reality of use rather than an optimized use case? The 0.24 watt-hour announced correspond to a “median” request in May 2025. without specifying the variability according to the types of questions or the length of the answers.
The Californian company remains discreet about its total consumption linked to AI, preferring to communicate on effectiveness google finally reveals how really by request. An approach which, if it testifies to real technical progress, says nothing about the explosion of the volume of use. Because even with an efficiency multiplied by 33. the total impact can increase if the number of requests explodes even more quickly. Besides, it looks like Nvidia’s communication for its graphics cards.
Further reading: New Cytal Lite Poc helmet – Documentary: a filmmaker confronts conspirators with the reality of war in Ukraine – International study: estrogens protect the kidneys of women – How does electro deposit do so much the price of the Redmi Note 14 of the Xiaomi brand? – Recognized for its quality, this XGIMI video projector is at canon prices during the sales.
