Breadcrumb
AI Energy Use Fact Sheet
AI has many benefits for productivity and learning; however, one concern many have raised is the environmental impact and energy demands of LLM models. Being a good steward of our environment is important, and while more answers are still needed, it’s important to put things in perspective.
It’s important to put the energy consumption from AI in context of other daily activities.
TikTok
1 hour of TikTok ≈
58.8 Watt Hours →
58 LLM chats
Gaming
1 hour of PS5 ≈
209 Watt Hours →
209 LLM
Driving
1 mile EV driving ≈
333 Watt Hours →
333 LLM chats
Flight
1 mile of flying ≈
576 Watt Hours →
576 LLM chats
Other Facts
The often-cited statistic: “LLMs take 10x the energy to run compared to a Google search” does not have empirical backing — it’s a guess someone made that stuck.
The major AI vendors have not released data on the energy consumption of their AI models, and the available data on Google Search’s energy use comes from a decade-old press release.
A large amount of energy is expended in training AI models; however, using these models is much less energy intensive.
All model training is energy intensive, but so are many forms of high-performance research computing — such as climate modeling, genomics, and physics simulations.
Energy consumption varies greatly depending on the model you are using.
The website ml.energy provides data on energy use for open models. The largest listed is a LLaMA 3, 140B parameter model that uses roughly 1 watt hour per use.
