Training and running large AI models consumes significant energy — raising important environmental and equity questions.

The Numbers

• Training GPT-3 consumed approximately 1,287 MWh — equivalent to flying 550 times across the US
• A single ChatGPT query uses roughly 10x the energy of a Google search
• AI data centers are driving unprecedented increases in electricity demand

Mitigation Approaches

• Smaller, more efficient models (distillation)
• Renewable energy-powered data centers
• Model sharing instead of retraining from scratch
• Efficient inference techniques (quantization, pruning)

Reference:

ML Carbon Impact Calculator

https://mlco2.github.io/impact/