Understanding GPU Pricing: Spot vs On-Demand
On-Demand Instances
On-demand instances are the standard option. You request a GPU, you get it, and it stays yours until you terminate it. Pricing is fixed and predictable.
**Best for:** Production inference endpoints, long-running training jobs where interruption is costly.
Spot Instances
Spot instances use excess provider capacity at a significant discount — often 50-70% cheaper than on-demand. The catch: they can be preempted (terminated) with little notice when demand increases.
**Best for:** Fault-tolerant training jobs, batch processing, experiments.
The Preemption Problem
The biggest risk with spot instances is losing your work. If a training job gets preempted after 6 hours, you might lose all progress since the last checkpoint.
How Pulserun Handles Spot
Pulserun's spot recovery system polls for preemptions every 30 seconds. When a preemption is detected:
- We immediately find the cheapest replacement instance
- Auto-provision the new instance
- Restore from the latest checkpoint
- Resume your workload
Most users experience less than 2 minutes of downtime, and with proper checkpointing, zero data loss.
The Best of Both Worlds
With Pulserun's Autopilot, you can run on spot instances and get near-on-demand reliability. Set a budget, enable spot recovery, and let automation handle the rest.