These numbers match independent direct measures: 0.00004 kWh for 400 tokens on Llama 3.3 70B on a H100 node. We do not know the amount of energy required to train these models, which was estimated at a little above 500,000 kWh for GPT-4, about 18 hours of a Boeing 737 in flight.
Energy consumption metrics for Llama 3.3 and GPT-4 training
By
–