<- All terms

Spot instance

Spare cloud GPU capacity sold at a steep discount (often 50-80% off on-demand) that the provider can reclaim with little notice. Good for interruptible work, risky for live serving.

What it is

The provider rents out unused capacity cheaply and takes it back when a full-price customer wants it - sometimes with a 30-120 second warning, sometimes none. Prices float with supply and demand.

Why it matters

Spot is a strong fit for fine-tuning, batch inference and training with checkpointing - anything that can pause and resume. For low-latency live serving it needs fallback capacity and fast cold start handling, which erodes the saving.

Cost & infrastructure impact

On marketplaces like Vast.ai, spot/interruptible rates for a consumer GPU can be a fraction of on-demand - but availability and reliability vary by host.

Related concepts

Use it in Obolith

Last reviewed: 2026-09-01 · evergreen concept