Compute Comparison
Cost Guides7 min read

GPU Spot vs On-Demand Pricing: When to Use Each (2026)

A practical guide to GPU spot and on-demand pricing — interruption rates, cost savings, workload suitability, and provider-by-provider spot availability.

What is GPU Spot Pricing?

Spot (or preemptible) GPU instances are spare compute capacity offered at discounted rates — typically 40–80% below on-demand prices. The trade-off: the provider can reclaim the instance with short notice (usually 30 seconds to 2 minutes) when demand increases. On-demand instances run until you terminate them, with guaranteed availability at a fixed price.

In 2026, GPU spot pricing has become more sophisticated. RunPod and Vast.ai operate marketplace models where spot prices fluctuate with real-time supply and demand. AWS and Google Cloud use fixed spot discounts with probabilistic interruption. CoreWeave offers "interruptible" instances with 30-second notice and 60–70% discounts. See current spot vs on-demand rates across all providers.

Spot Pricing by Provider

AWS EC2 Spot: H100 (p5.48xlarge) spot saves 60–70% vs on-demand. Interruption rates vary by region and instance type — us-east-1 p4d instances have historically had <5% hourly interruption rates. Google Cloud Preemptible/Spot: 60–91% discount, 24-hour maximum runtime, 30-second preemption notice. Azure Spot: 60–90% discount, variable interruption rates.

Specialist providers: RunPod community cloud spot from $1.89/hr for H100 (vs $2.89/hr on-demand). Vast.ai marketplace spot from $1.60/hr for H100. Lambda Labs does not offer spot — on-demand only. CoreWeave interruptible instances: 60% discount with 30-second notice. Compare RunPod vs CoreWeave pricing side-by-side.

Workloads Suited for Spot

Spot instances are ideal for: (1) LLM training with checkpointing — save model state every 15–30 minutes so interruptions lose minimal progress; (2) batch inference — stateless requests that can be retried on interruption; (3) data preprocessing pipelines — embarrassingly parallel tasks that can be restarted; (4) hyperparameter search — many short independent runs where interruption of one doesn't affect others.

Spot is not suitable for: (1) real-time inference APIs where interruption causes user-facing downtime; (2) long training runs without robust checkpointing; (3) stateful applications that can't tolerate data loss; (4) workloads with strict SLA requirements. For managed inference without interruption risk, compare LLM API providers as an alternative to self-hosted GPU inference.

Checkpointing Strategy for Spot Training

To use spot instances for LLM training safely: (1) checkpoint every 15–30 minutes to persistent storage (S3, GCS, or NFS); (2) use a spot interruption handler to trigger a final checkpoint on the 2-minute warning; (3) use a training framework with automatic checkpoint resumption (PyTorch Lightning, Hugging Face Accelerate, or DeepSpeed); (4) store checkpoints in a different availability zone or region from your compute.

With proper checkpointing, spot training typically costs 40–70% less than on-demand for equivalent training runs, even accounting for occasional restarts. See our guide on cheapest GPU cloud for AI inference for a full provider tier ranking.

Cost Comparison: Spot vs On-Demand

Example: Training a 7B parameter LLM for 100 GPU-hours on H100. On-demand (Lambda Labs): $2.49/hr × 100 = $249. Spot (RunPod): $1.89/hr × 100 = $189, but with ~10% expected overhead from interruptions and restarts = ~$208 effective cost. Savings: ~$41 (16%).

For longer runs (1,000+ GPU-hours), spot savings compound: $2,490 on-demand vs ~$2,080 effective spot cost = $410 saved. The longer the training run, the more valuable robust checkpointing becomes. Track H100 price history to see how spot and on-demand rates have moved over the past 90 days.

Frequently Asked Questions

How often are GPU spot instances interrupted?
Interruption rates vary widely. AWS p4d spot in us-east-1 historically interrupts <5% of the time per hour. RunPod community cloud interruptions are less predictable. Vast.ai marketplace instances can be interrupted when the host reclaims their hardware.
Can I use spot instances for production inference?
Yes, with auto-scaling and load balancing. Deploy multiple spot instances behind a load balancer — when one is interrupted, traffic routes to others. This requires stateless inference and fast instance startup times (<60 seconds).
What is the maximum spot discount available?
Vast.ai and Salad Cloud offer the highest discounts — up to 80–90% below equivalent on-demand rates for older GPU models (RTX 3090, A5000). H100 spot discounts are typically 40–60% below on-demand.
Do spot instances affect training reproducibility?
Spot interruptions don't affect reproducibility if you checkpoint correctly. Resume training from the last checkpoint with the same random seed and data order. Most modern training frameworks handle this automatically.

More Guides