What is the difference between spot and on-demand GPU pricing?
On-demand means you rent a GPU at a fixed rate and keep it until you stop it. Spot, also called interruptible or preemptible, rents spare capacity at a discount but the provider can reclaim it with little warning. Spot is often much cheaper, so the choice comes down to whether your job can survive being interrupted.
Spot pricing exists because providers have idle GPUs and would rather sell them cheap than leave them dark. When demand rises or someone pays more, your spot instance can be taken back in seconds. That is fine for batch inference, data preprocessing, or training that checkpoints regularly, since you just resume from the last save.
On-demand is what you want when an interruption costs you something real: a live inference endpoint, an interactive session you are actively using, or a deadline where a restart would hurt. You pay for the certainty that the GPU stays yours until you release it.
A useful pattern is to mix the two. Run the interruptible bulk of a workload on spot, and keep a small on-demand instance for the parts that must stay up. Always compare the firm rate against the spot rate for the same card before deciding, since the gap varies by provider and by how tight capacity is that day.
Related questions
- What is the difference between secure cloud and community cloud?
- Why are GPU prices different between providers?
- What is the cheapest cloud GPU right now?
Numbers on this page come from today's verified snapshot. Full table on the homepage; method in the methodology.