What consumes energy inside an LLM, and how much?_

I read a lot of figures on the footprint of AI, enough to see there is a real subject. But I did not want to do green AI on a black box. I wanted to understand how an LLM works, and be able to explain it to someone who is not an engineer.

This course is what I understood. Written partly with AI. The figures I checked at source, with their date and their measurement boundary.

Seventeen chapters, over 3 hr 30 min.

The seventeen chapters

• 01What happens when you send a prompt to an AI?What goes on inside an LLM when you send a prompt, and where the electricity goes at each of the six steps.19 min read
• 02Why count tokens instead of requests?The unit that makes an AI footprint auditable, and the three token classes a single counter always gets wrong by a factor of 5 to 10.9 min read
• 03Does French cost more than English?What a language costs in tokens, why the corpus rather than the grammar decides it, and what it changes for a multilingual service.10 min read
• 04Why does an output token cost 35 times an input token?The two physical regimes of an inference, the measured coefficients behind the factor of 35, and the levers that follow from them.12 min read
• 05Why does long context cost so much?Where the cost of a long context actually sits, and why two independent providers landed on the same x2 pricing tier.11 min read · technical
• 06Why does a model's parameter count no longer tell you its cost?Read the only number that predicts inference compute, and see why MoE cuts calculation while quietly inflating the hardware you have to keep loaded.12 min read · technical
• 07Does training or inference weigh more in an AI model's footprint?Gives you the crossover threshold, the reason server fill moves it by a factor of 30, and the five circulating figures that measure five different things.17 min read · technical
• 08How much energy does an LLM's reasoning mode use?Sorts six published multipliers by scope so you stop stacking them, and shows how to measure the effort setting on your own tasks in an afternoon.13 min read
• 09Why does an AI agent use up to 1,000 times a single request?Explains the quadratic growth behind agentic cost, what prompt caching does and does not solve, and which three usage-export figures rebuild the footprint.14 min read
• 10How much electricity do AI data centres really use?Pins down the IEA figure and its boundary, separates AI from the rest of the sector, and arms you against gigawatt numbers that are announcements, not load.13 min read
• 11What does a figure of 0.24 Wh per request actually cover?The only per-line breakdown a provider has published, what each omitted item costs you, and the questions to ask before accepting any per-request number.11 min read
• 12How do you build an emission factor per token?The five-line formula with its unit checks, the value and range of every parameter, and the one omission that makes it underestimate agentic use.12 min read · technical
• 13Which grid emission factor should you use for a data centre?The three decisions that pick one value out of eight, when the ADEME factor is mandatory in France, and why "our servers are in PJM" is not an answer.15 min read · technical
• 14Why are there two carbon numbers for the same electricity?What each scope 2 method actually records, the Microsoft case that proves market-based tracks purchases rather than consumption, and which number to steer by.12 min read · technical
• 15Water and PUE in an AI data centre: what are the published figures worth?The one question to ask about any water figure, why a hyperscale PUE is not comparable to a survey average, and the defensible default for an uninstrumented site.15 min read · technical
• 16How do you take apart a comparison between two AI providers?A checklist to run before accepting any provider comparison, plus what the EU AI Act does and does not require of providers.13 min read
• 17How do you cut the carbon footprint of AI? Ten levers ranked by impactTen quantified levers with who controls each one, the fake lever you will be offered, and why efficiency alone never brings the total down.19 min read

Version française