AI's environmental footprint comes from the energy and water used by data centres and the hardware they run on. Teams can reduce it.
Choose Efficient Models
- Use the smallest model that meets quality needs.
- Route simple tasks to small models.
- Use distilled or quantised models.
Efficient Usage
- Cache repeated requests and prompt prefixes.
- Avoid unnecessary long contexts.
- Use batch processing for non-urgent work.
Efficient Training
- Fine-tune rather than train from scratch.
- Stop experiments early when they're clearly not working.
- Reuse and share models.
Infrastructure Choices
- Regions and providers with cleaner electricity.
- Efficient hardware.
- Scheduling flexible workloads when renewable supply is high.
Measure and Report
Estimate the energy and emissions of major workloads. Tools and provider dashboards help, though measurement remains imperfect.
Weigh Benefits
Consider whether an AI feature delivers enough value to justify its footprint, alongside cost and other factors.