In short
Find the jobs that can wait, such as training, fine-tuning, evaluation and batch inference, and start them when a carbon-intensity forecast shows the grid is cleaner. Save regular checkpoints so long jobs can pause when intensity rises and resume when it falls, within a firm deadline. Carbon-aware scheduling cuts carbon, not energy, so do it on top of efficiency work.
Steps
- List AI jobs that can wait hours without harm, with each job's real deadline.
- Get a carbon-intensity forecast for the region where the job runs.
- Start flexible jobs in the cleanest window that still meets the deadline.
- Save checkpoints, including optimiser state, so long jobs can pause and resume.
- Track energy per job as well as carbon, so scheduling doesn't hide growth.
Most training jobs don't need to start the moment you press enter. Moving them to cleaner hours lowers their carbon with no change to the model.
Where do forecasts come from?
- In Great Britain, the National Energy System Operator's Carbon Intensity API forecasts carbon intensity more than 96 hours ahead for 14 regions.
- Electricity Maps provides real-time and forecast data for more than 350 zones worldwide.
- The Green Software Foundation's Carbon Aware SDK is a web API and command-line tool that draws on such data so software can "do more when the electricity comes from clean low-carbon sources and less when it does not".
How do I handle long jobs?
The Green Software Foundation's pattern for AI workloads suggests pausing and resuming long-running training jobs based on carbon-intensity thresholds where feasible. That needs regular checkpoints, including optimiser state, so nothing is lost when a job stops. Measure the checkpoint overhead so the pauses don't cost more than they save.
Common mistakes
- Confusing carbon with energy. Shifting a job doesn't make it use less; a falling carbon figure can hide rising energy. Track both.
- Letting deadlines slip. Make the deadline a hard limit, not a preference.
- Scheduling against yearly averages. You need hourly forecasts for the region where the job actually runs.
- Scheduling tiny jobs. The gain is in big, flexible runs, not every script.
You can also choose a lower-carbon region. See efficient training and inference.
Sources
- Carbon Intensity API, National Energy System Operator (2025)
- Carbon Aware SDK, Green Software Foundation (2025)
- Use carbon-aware scheduling and region selection for AI workloads, Green Software Foundation, Green Software Patterns (2025)
- Electricity Maps data, Electricity Maps (2026)
Related
How do I pick a cloud region with low carbon?
Choose a lower-carbon cloud region for AI work: compare grid carbon intensity and carbon-free energy by region, then weigh latency, data rules and cost.
How do I find the carbon intensity of my electricity grid?
Where to get grid carbon intensity data for AI footprints: regional, hourly sources such as the GB Carbon Intensity API and Electricity Maps.
How do I run AI jobs on spare cloud capacity?
Run interruptible AI work on spot or preemptible capacity: make jobs checkpoint and retry cleanly, and keep work that can't be interrupted elsewhere.
Efficient AI training and inference
Techniques for developers and researchers to cut AI energy: pruning, sparsity, quantisation, distillation, KV cache, batching, mixed precision and scheduling.