Spot GPU Instances:
Up to 50% Off NVIDIA GPUs.
Run batch jobs, training experiments, and non-urgent workloads on discounted spot instances. Same NVIDIA GPUs, same performance, lower price. Instances can be terminated when demand spikes.
Spot instances use idle GPU capacity from our data center network. Same hardware, same data centers, same performance. The only difference is the price tag and the possibility of interruption during demand spikes.
How Spot Instances Work
Idle Capacity
GPUs not currently booked on dedicated instances become available as spot capacity at a significantly discounted rate.
You Deploy
Launch a spot instance just like a dedicated one. Same configuration, same GPU, lower price.
You Run Your Job
Your workload runs at full performance. No throttling. Checkpoint regularly for fault tolerance.
Termination
If capacity is reclaimed, your spot instance is terminated. Checkpoint regularly so you can resume from where you left off.
Built for These Workloads
Any workload that can tolerate interruption is a candidate for spot pricing. Checkpoint your work, and restarts are painless.
Hyperparameter Tuning
Run hundreds of short experiments at a fraction of the cost. Each run is independent, so interruptions only affect one trial.
Data Preprocessing
ETL pipelines and dataset transformations that can be restarted from the last checkpoint. Perfect for spot pricing.
Model Evaluation
Run inference benchmarks and validation passes on cheaper hardware. Results don't change based on instance type.
Research and Prototyping
Explore architectures and ideas without worrying about burning through your GPU budget. Iterate fast, spend less.
Spot vs Dedicated
Mix Spot and Dedicated
Run your critical training loop on a dedicated instance (99.99% SLA) while offloading evaluation jobs to spot. Both are on-demand tiers, so you can mix them across your pipeline to optimize cost without sacrificing reliability where it matters.
Compare Instance Types
Dedicated
On-demand instances with a 99.99% SLA. Non-interruptible. Deploy instantly, pay by the minute.
Production workloads, long-running training, critical inference
Spot
CurrentOn-demand instances at a discount using idle capacity. Same hardware, up to 50% off, interruptible.
Batch jobs, experiments, hyperparameter tuning, flexible workloads
Reserved
Locked capacity with volume pricing. Custom clusters and dedicated support.
Large deployments, multi-month projects, enterprise teams
Spot Instance FAQ
Spot instances let you rent GPUs at a discount by using idle capacity in our data center network. They run on the same hardware as dedicated instances but cost up to 50% less. Both spot and dedicated are on-demand tiers with per-minute billing. The tradeoff: spot instances can be terminated when demand for that GPU model increases, whereas dedicated instances carry a 99.99% SLA.
Savings vary by GPU model and current demand. Typical discounts range from 20% to 50% off the dedicated rate. Check the pricing page for current spot rates on each GPU model.
View pricingSpot instances can be terminated when demand for that capacity is reclaimed. There is no grace period, so design your pipeline with regular checkpointing. Save progress frequently, and restarts pick up from the last checkpoint.
Any workload that handles interruption gracefully: hyperparameter searches, data preprocessing, model evaluation, batch inference, and research experiments. If you can checkpoint and restart, you can save money with spot.
Yes. Both are on-demand tiers, so you can mix them freely. Many teams run critical training jobs on dedicated instances (99.99% SLA) while offloading evaluation, preprocessing, and experimental runs to spot. This hybrid approach gives you reliability where it counts and savings everywhere else.
Halve Your GPU Bill
Same GPUs. Same data centers. Same performance. If your workload can tolerate interruption, there's no reason to pay full price.