# AI Sustainability
The environmental cost of AI. Training a large model can emit as much carbon as five cars over their lifetime. Inference at scale consumes enormous energy. This is not a future problem; it is a current one that scales with adoption.
Key dimensions:
- **Energy consumption per query**: Frontier models consume orders of magnitude more compute than [[Small Language Models (SLMs)]]. A single GPT-4 class query uses roughly 10x the energy of a web search. Choosing the right model for the task matters (see [[AI Model Selection]]).
- **Carbon footprint of training vs inference**: Training gets the headlines, but inference at scale often exceeds training costs. A model trained once but served billions of times accumulates carbon fast.
- **Water usage**: Data centers require massive cooling. A single large training run can consume millions of liters of water. This is a resource cost that rarely appears in cost calculations.
- **E-waste from GPU hardware cycles**: The race for faster hardware creates rapid obsolescence. GPU generations turn over every 1-2 years, generating significant electronic waste.
Practical implications:
- Choose smaller models when they are good enough. Not every task needs a frontier model.
- Use [[Running AI Models Locally]] to reduce cloud compute dependency and gain visibility into actual resource consumption.
- Batch requests and cache results to avoid redundant computation.
- Apply [[AI Quantization]] to reduce compute per inference without proportional quality loss.
- Factor sustainability into [[AI Cost Management]]; energy costs and carbon costs are converging.
Increasingly a compliance and reporting requirement for enterprises. [[AI Governance]] frameworks now include environmental impact assessments. [[Responsible AI]] cannot ignore the environmental dimension.
## References
## Related
- [[AI Model Selection]]
- [[AI Cost Management]]
- [[Running AI Models Locally]]
- [[AI Quantization]]
- [[Small Language Models (SLMs)]]
- [[AI Governance]]
- [[Responsible AI]]