Solar-first AI workloads: what works best?
Some AI workloads shift comfortably into daylight hours. Others cannot. Here is a practical way to tell them apart.
A workload is a good candidate for solar-first scheduling when it is tolerant of delay, has a clear completion state, and can be checkpointed or restarted safely.
Strong candidates include batch embedding and indexing, dataset preparation, image and video generation queues, model evaluation sweeps, and periodic retraining on local data.
Weak candidates include interactive assistants, real-time computer vision on safety-relevant feeds, and anything with a contractual response-time commitment.
The design goal is not to force everything into daylight. It is to move the shiftable share, and to be explicit about which workloads are deliberately always-on so their supply path is planned properly.
- workloads
- policy
- batch