Subscribe to watch
Access this and all of the content on our platform by signing up for a 7-day free trial.
Estimating the Energy Use of AI
1 min 46 secs
Key learning objectives:
Outline how much energy AI uses
Understand why energy use varies with AI
Explain why figures are hard to find on AI energy use
Overview:
Popular chatbots operate as black boxes, meaning you know very little about the model itself, where it is running, or what kind of energy powers it. To get around this lack of transparency, researchers have benchmarked open-source models to scientifically estimate their power consumption. These approximations reveal that energy usage varies dramatically depending on the tool.
Subscribe to watch
Access this and all of the content on our platform by signing up for a 7-day free trial.
Summary
Why does energy use vary so dramatically with AI?
Energy requirements vary because different models rely on complex capabilities built during a process called "training". Training involves running a tremendous amount of data and mathematics through a model to create a neural network capable of recognising patterns and generating outputs. Because different types of tasks demand vastly different amounts of computational work, models end up with radically different energy needs. Consequently, the environmental impact of AI cannot be generalised, and broad statements about "AI's footprint" are misleading.
Why are exact energy figures so hard to find?
When you prompt popular chatbots like ChatGPT, Claude, or Gemini, your request travels off to chips in a distant data centre, but the processing happens inside closed, proprietary models. These systems operate as "black boxes", meaning you know very little about the model itself, where it is running, or what kind of energy powers it. Because providers usually do not disclose this information, exact energy figures remain hidden unless the provider explicitly chooses to share them.
How much energy does AI actually use?
To get around this lack of transparency, researchers have benchmarked open-source models to scientifically estimate their power consumption. These approximations reveal that energy usage varies dramatically depending on the tool:
When you prompt popular chatbots like ChatGPT, Claude, or Gemini, your request travels off to chips in a distant data centre, but the processing happens inside closed, proprietary models. These systems operate as "black boxes", meaning you know very little about the model itself, where it is running, or what kind of energy powers it. Because providers usually do not disclose this information, exact energy figures remain hidden unless the provider explicitly chooses to share them.
How much energy does AI actually use?
To get around this lack of transparency, researchers have benchmarked open-source models to scientifically estimate their power consumption. These approximations reveal that energy usage varies dramatically depending on the tool:
- Large Language Models (LLMs): Range from 0.2 to approximately 40 Watt-hours per query
- Video generation models: Consume over 400 Watt-hours
Why does energy use vary so dramatically with AI?
Energy requirements vary because different models rely on complex capabilities built during a process called "training". Training involves running a tremendous amount of data and mathematics through a model to create a neural network capable of recognising patterns and generating outputs. Because different types of tasks demand vastly different amounts of computational work, models end up with radically different energy needs. Consequently, the environmental impact of AI cannot be generalised, and broad statements about "AI's footprint" are misleading.
Subscribe to watch
Access this and all of the content on our platform by signing up for a 7-day free trial.
Boris Gamazaychikov
Boris Gamazaychikov is an AI Sustainability Leader and environmental engineer by training, focused on aligning AI development with climate action. He is the Co-Founder and CEO of Sustainable AI Group, a research and advisory firm helping enterprises understand and act on the environmental impacts of AI. Previously, he served as Head of AI Sustainability at Salesforce, where he built the company's AI sustainability strategy from the ground up and co-created the AI Energy Score - the first standardised benchmark for measuring AI model energy use.
There are no available Videos from "Boris Gamazaychikov"