Spot Instances
Spot instances are GPU compute resources offered at variable, discounted pricing with the caveat that the provider can reclaim (pre-empt) the capacity with minimal notice — typically 30 seconds to 2 minutes. Spot pricing reflects real-time supply and demand dynamics: when a provider has unused GPU capacity, spot prices drop to attract workloads; when demand is high, spot instances may be unavailable or priced near on-demand rates. Spot discounts for GPU instances typically range from 40-70% off on-demand pricing.
Spot instances are best suited for fault-tolerant workloads: large-scale training with checkpointing, batch inference, hyperparameter sweeps, and data preprocessing. Workloads must be architected for interruption — checkpoint saving at regular intervals, automatic job resumption, and graceful shutdown handling. Some providers offer "spot-like" products with longer interruption guarantees (24 hours) at smaller discounts. The spot market for GPUs is less liquid than for CPU instances, and availability can be unpredictable, particularly for high-end GPUs like H100 and B200.
Spot pricing data is a leading indicator of supply-demand dynamics in the GPU market. We track spot rates across providers as part of our pricing intelligence, using them to inform utilisation assumptions in due diligence and to benchmark operator revenue projections.
This glossary is maintained by Disintermediate as a reference for GPU infrastructure professionals, investors, and operators. Each entry reflects terminology as used in active advisory engagements and market intelligence work.