Happy New Year 2025 to everyone!

It’s time to start a new year. Let’s bet that 2025 will be full of digital adventures and cutting-edge technologies! As in the video below using the latest technologies, I wish everyone a happy new year 2025, marked by a certain responsibility…

Indeed, information technologies have their downside: polluting production, waste and sometimes disproportionate energy consumption. It is therefore everyone’s responsibility to make good use of it and to think before launching poorly worded or useless requests…

Happy New Year to everyone!

Energy and AI: what does ChatGPT o1 Pro say “Optimal Reasoning”?

Below is the result of a query in the pro mode of ChatGPT o1 Pro, “Optimal reasoning” version. The links allowing you to verify the information opposite were also requested from the same version of ChatGPT.

Energy consumption, a major challenge for AI

The energy consumption of artificial intelligence (AI) systems is an increasingly debated topic. The data centers that host and operate the models run 24 hours a day, requiring significant electricity resources for powering and cooling the servers. We therefore assess that infrastructures supporting AI contribute significantly to the digital carbon footprint. Estimates vary, but it is clear that each query performed on an AI like ChatGPT, each image generation via a neural network, or even each online search query has an energy cost. To give orders of magnitude, we generally use kilowatt hours (kWh) or watt hours (Wh), which makes it easier to compare the impact of different technologies.

Compare a ChatGPT query to a Google query

A standard Google search is often cited as consuming around 0.3 Wh*, or 0.0003 kWh. This value includes server operation, data transmission and estimated network cost. In comparison, a request to ChatGPT, which mobilizes a massive linguistic model (transforming several billion parameters), can require between 3 and 15 Wh per request, or 0.003 to 0.015 kWh. The range is wide because it depends on the complexity of the question and the length of the answer. Thus, interacting with a large language model can require up to several dozen times more energy than a simple web search, although the exact figures vary depending on server optimization and network load.

*1 Wh (watt-hour) corresponds to the energy consumed by a device with a power of 1 watt for 1 hour.

ReqDALL-E type image header

Generating images from AI models like DALL-E or Stable Diffusion is even more expensive, because it requires intensive graphics processing: each pixel must be “imagined” by the neural network. For high quality image rendering, we can estimate a consumption of between 15 and 50 Wh, or 0.015 to 0.05 kWh, per generated image. Again, the range depends on the level of detail required, the size of the image, as well as the capability of the hardware used for inference (specialized graphics cards, optimized servers, etc.).

Streaming an HD video of 10 seconds of Sora

Finally, with regard to the broadcast (streaming) of a 10-second HD video like that of “Sora”, we must distinguish the production or initial encoding of the video from its simple online playback. For viewing, estimates are around 0.2 to 0.5 Wh (0.0002 to 0.0005 kWh) per user, taking into account playback, network infrastructure and servers. On the other hand, the generation of such a video by a “text-to-video” AI system would be significantly more demanding, probably between 30 and 100 Wh (0.03 to 0.1 kWh), due to the complexity processing of moving images. In any case, these figures should be taken as orders of magnitude: they illustrate the extent to which AI can consume energy and invite reflection on the optimization of algorithms and infrastructures to contain the carbon footprint of digital technology.

And here are the sources proposed by the same ChatGPT o1 Pro, “Optimal reasoning” always in italics:

Here are some references and sources to illustrate the orders of magnitude mentioned in each of the paragraphs. The numerical values ​​cited may vary depending on the studies, hypotheses and developments in the equipment or infrastructure:

Paragraph 1: energy consumption, a major challenge for AI

1. The Shift Project (2019)‘Lean ICT – Towards Digital Sobriety’:
   • Report detailing the environmental impact of information technology, including AI and data centers.
   • Available online: theshiftproject.org
2. Koomey, J.G. (2011)«Growth in data center electricity use 2005 to 2010»:
   • University study analyzing data center power consumption over several years.
   • Available online: analyticspress.org
3. International Energy Agency (IEA)successive reports on the digital consumption:
   • The IEA regularly offers reports and statistics on energy uses, including those of data centers.
   • See for example: iea.org

Paragraph 2: Compare a ChatGPT query to a Google query

1. Google — Data center energy efficiency data:
   • Google regularly communicates on improving its infrastructure and reducing the carbon footprint of its services.
   • See: sustainability.google
2. Berners-Lee, M. (2010)«How Bad Are Bananas?: The Carbon Footprint of Everything»:
   • Work that cites the average energy consumption of a Google search and other digital activities.
3. OpenAI References or blogs techniques:
   • Although there is not always an official publication precisely quantifying consumption per request, several blog articles and expert interviews suggest a range for ChatGPT or similar GPT models.
   • Examples:
     • openai.com/blog/
     • “GPT-3.5 and GPT-4 energy usage” (unofficial technical articles, conferences, etc.)

Paragraph 3: DALL-E type image query

1. OpenAI — Documentation and announcements around DALL-E:
   • Some blog posts discuss the complexity of image generation and the resources required.
   • See: openai.com/dall-e-2
2. Stability AI (Stable Diffusion):
   • Technical discussions on forums (GitHub, Reddit) discuss the consumption of GPUs (computation, inference) to generate images.
   • Examples: github.com/CompVis/stable-diffusion
3. Industry Reports/GPU Benchmarks:
   • Some manufacturers (NVIDIA, AMD) publish consumption or efficiency data for different types of AI workload.
   • Examples: nvidia.com/en-us/data-center/

Paragraph 4: Streaming a 10-second HD video of Sora

1. IEA (2020)‘Data Centres and Data Transmission Networks’:
   • The report includes estimates on energy consumed during video streaming and data transfer.
   • Available online: iea.org
2. The Shift Project (2019)‘Lean ICT – Towards Digital Sobriety’:
   • Also includes consumption estimates for video streaming and OTT services.
3. Articles or studies specific to streaming:
   • Various academic publications or white papers (e.g. Cisco, Sandvine) offer energy consumption estimates for the distribution of audiovisual content depending on quality (SD, HD, 4K, etc.).
   • Examples:

Noticed: The precise values ​​evolve quickly depending on the optimization of infrastructures (network, data center cooling, GPU efficiency, etc.). The figures indicated in the preceding paragraphs should therefore be considered as orders of magnitude coming from different reports and studies.