Local training (consumer GPU) has no recurring costs but limits you to QLoRA on -B models. Cloud training (A100, H100) costs more per hour but handles any model size.
Cloud providers:
- Lambda Labs: Simple pricing, good availability
- RunPod: Spot instances, community cloud
- Vast.ai: Marketplace model, cheapest options
Spot instances are -% cheaper but can be interrupted. Use checkpointing.