What reduces GPU spending?
Choose a configuration that fits, measure useful work per billed hour and stop resources you no longer need. The potential savings depend on your workload and provider terms. NVGPU does not promise a percentage reduction or assume that renting always beats ownership.
Build a budget with the full instance
Compute cost equals full-instance rate × billed hours. Add download, setup, checkpoint, retry and idle time to your estimate. Then add storage, data transfer, licenses, support and taxes where applicable. The homepage calculator accepts your own budget for extras rather than assuming a universal storage or egress rate.
Record the source version next to each quoted rate. Match GPU count, VRAM, hardware variant, region and pricing type on the comparison page.
Check where time is being spent
Low GPU utilization can result from input loading, CPU preprocessing, synchronization or a small workload. Profile before upgrading. Quantization, batching and activation checkpointing can change memory use, quality and speed in different directions; validate the result instead of assuming a smaller representation always runs faster.
Use the PyTorch profiler to distinguish CPU and CUDA work. Treat the profile as evidence for a specific software version and workload.
Separate spot from committed pricing
Spot is interruptible capacity; a commitment is a contractual purchase. Compare them separately. Test restoration from persistent checkpoints before a long spot run. The discounted rate may be outweighed by retries or waiting. AWS Spot documentation explains that interruption notices and behavior have limits.
For reservations or savings plans, estimate paid-but-unused hours and read cancellation and transfer terms. Our catalog does not include negotiated or commitment quotes.
Check what still bills after stopping
Stopping a workload, stopping a VM and deleting a VM are different operations. Confirm which stops compute charges and which retains storage. RunPod's pricing documentation describes storage charges on stopped Pods. Lambda's billing documentation distinguishes instance and filesystem charges.
Back up needed data before terminating resources. Automate an idle timeout only after confirming how your provider persists data and resumes billing.
When is a different region cheaper?
A lower listed compute rate in another region is only one input. Add data movement, latency, required data residency, service availability and quota constraints. Compare total completed-job cost under the same assumptions; do not apply a generic regional savings percentage.
Compare renting with ownership if sustained utilization makes a local system plausible.