GPU rental cost calculator
Estimate what a GPU rental job will cost and see every provider's price ranked cheapest first. Enter a GPU model, the number of GPUs and the hours; the calculator multiplies the per-GPU-hour rate by GPUs and hours across 14 providers and highlights the cheapest. Toggle on-demand vs spot/community, or estimate hours from a training-run GPU-hour figure. Everything runs in your browser on a June 2026 price snapshot.
Data as of June 2026.
Snapshot June 2026 — cloud GPU prices change weekly; verify on the provider's pricing page before relying on a figure.
How the estimate is calculated
The formula is simple and transparent:
job cost = price per GPU per hour x number of GPUs x hours
The per-GPU-hour price comes from each provider's published pricing page (a dated June 2026 snapshot). Hyperscaler (AWS/GCP/Azure) per-GPU prices are derived from their 8-GPU instance prices and include bundled vCPU and RAM; marketplace (Vast.ai) prices are moving medians. The estimate excludes storage, egress, idle time, inter-node networking and minimum billing increments. See the methodology for the full detail.
Frequently asked questions
How does the GPU cost calculator work?
Enter a GPU model, the number of GPUs and the number of hours. The calculator multiplies the per-GPU-hour price by GPUs and hours for every provider that publishes a rate, then ranks them cheapest first and highlights the winner. Switch between on-demand and spot/community pricing. All prices are a dated June 2026 snapshot held in your browser — nothing is sent to a server.
Can it estimate the cost of training a model?
Roughly. Open the 'estimate hours from a training run' helper and enter the total GPU-hours your run needs (from a tokens/throughput estimate). The tool sets hours = GPU-hours / number of GPUs. This is a rough planning figure: real training cost depends on throughput, efficiency (MFU), checkpointing, restarts and data-loading overhead.
What does the estimate exclude?
Storage, data egress, idle time, networking between nodes, minimum billing increments and any committed-use discounts. For multi-node training, inter-GPU bandwidth (InfiniBand) and node count also matter. Treat the result as a starting point and verify on the provider's pricing page.
Why are some providers missing for a GPU?
Not every provider offers every GPU. Where a provider does not publish a price for the chosen GPU, it is omitted from the table for that selection. If no provider lists the chosen pricing mode (e.g. spot), try on-demand.
Keep exploring
Last updated: 2026-06-21