United StatesRunPod01
Why consider
An integrated path across GPU development, Pods and serverless inference.
Watch for
Separate stopped-resource, persistent-storage, idle endpoint and transfer costs from compute.
Official-source GPU decision tool
Compare the rate—then price the completed GPU job.
Use the dated price and VRAM index below, then turn the selected rate into a workload estimate that includes setup, retries and retained storage.
Completed-job estimate
$19.00
$1.90 per successful workload hour
This exposes setup, retries and retained storage that an hourly GPU headline omits.
- Billable GPU time
- 12.00 hours
- Compute subtotal
- $9.00
- Retry overhead
- 1.00 hours
- Retained storage
- $10.00
- Other costs
- $0.00
Compare the same completed job
Return to the full API-versus-GPU decision with this completed-workload estimate, or compare the two active GPU routes directly.
Operational-cost signalSetup and retry time remain below 25% of productive runtime
Marketplace price flexibility may deserve more weight, but individual host and recovery risk still matter.
United StatesVast.ai02
Why consider
Price-sensitive, recoverable workloads that can compare individual marketplace offers.
Watch for
The lowest hourly listing is not automatically the lowest completed-job cost; verify host and recovery conditions.
Sponsored affiliate links · Re-enter each live offer into the same calculation
Taxes, currency conversion, network transfer, request pricing and your labor are excluded unless entered above. This tool runs only in your browser.
Official-source decision index
GPU cloud price, VRAM, region and billing index
Compare dated provider-published rates and billing boundaries before entering the same workload in the completed-job calculator. This is a research snapshot—not a live inventory feed or performance ranking.
12official price observations
5provider models
Aug 30, 2026Current research window
12 matching price rows
| Actual billing unit | Region evidence | Official source | ||||
|---|---|---|---|---|---|---|
| RunPod | RTX A5000×1 | 24 GB | $0.27 | Compute and storage billed per second | 31 global regions stated on the Cloud GPU page; exact GPU inventory varies by location. | Open sourceAug 30, 2026 |
| RunPod | A40×1 | 48 GB | $0.44 | Compute and storage billed per second | 31 global regions stated on the Cloud GPU page; exact GPU inventory varies by location. | Open sourceAug 30, 2026 |
| RunPod | A100 PCIe×1 | 80 GB | $1.39 | Compute and storage billed per second | 31 global regions stated on the Cloud GPU page; exact GPU inventory varies by location. | Open sourceAug 30, 2026 |
| RunPod | H200×1 | 141 GB | $4.59 | Compute and storage billed per second | 31 global regions stated on the Cloud GPU page; exact GPU inventory varies by location. | Open sourceAug 30, 2026 |
| Thunder Compute | RTX A6000×1 | 48 GB | $0.35 | GPU instances billed per minute; the page displays hourly equivalents | North America offering; confirm the exact deployment location and available configuration in the console. | Open sourceAug 30, 2026 |
| Thunder Compute | L40×1 | 48 GB | $0.79 | GPU instances billed per minute; the page displays hourly equivalents | North America offering; confirm the exact deployment location and available configuration in the console. | Open sourceAug 30, 2026 |
| Thunder Compute | A100 80GB×1 | 80 GB | $1.09 | GPU instances billed per minute; the page displays hourly equivalents | North America offering; confirm the exact deployment location and available configuration in the console. | Open sourceAug 30, 2026 |
| Thunder Compute | H100 PCIe×1 | 80 GB | $3.20 | GPU instances billed per minute; the page displays hourly equivalents | North America offering; confirm the exact deployment location and available configuration in the console. | Open sourceAug 30, 2026 |
| Massed Compute | A30×1 | 24 GB | $0.35 | Hourly on-demand catalog; confirm minimum billing and teardown behavior at launch | The pricing table does not attach a deployment location to every row; confirm the selected location before launch. | Open sourceAug 30, 2026 |
| Massed Compute | RTX A6000 (ALT config)×1 | 48 GB | $0.55 | Hourly on-demand catalog; confirm minimum billing and teardown behavior at launch | The pricing table does not attach a deployment location to every row; confirm the selected location before launch. | Open sourceAug 30, 2026 |
| Massed Compute | A100 80GB×1 | 80 GB | $1.35 | Hourly on-demand catalog; confirm minimum billing and teardown behavior at launch | The pricing table does not attach a deployment location to every row; confirm the selected location before launch. | Open sourceAug 30, 2026 |
| Massed Compute | H100 80GB×1 | 80 GB | $2.73 | Hourly on-demand catalog; confirm minimum billing and teardown behavior at launch | The pricing table does not attach a deployment location to every row; confirm the selected location before launch. | Open sourceAug 30, 2026 |
Use a displayed rate in the calculator above. The hourly display is not completed-job cost; add CPU, RAM, storage, retries and idle time where they are not already bundled.
Before checkout
Costs outside the GPU headline
Storage, stopped-resource behavior, transfer and offer type can reverse an hourly-price comparison. These notes come from current official provider pages.
Published catalog
RunPod
Representative single-GPU rates displayed on the official Cloud GPU page. Recheck the Community/Secure selection, region, stock and complete configuration.
- Operating company
- United States
- Service-region evidence
- 31 global regions stated on the Cloud GPU page; exact GPU inventory varies by location.
- Storage lifecycle
- Container and running volume disk: $0.10/GB/month; stopped volume disk: $0.20/GB/month; network volume under 1TB: $0.07/GB/month.
- Transfer
- Official billing documentation states no data-transfer fee.
Published catalog
Thunder Compute
Single-GPU base rates from the official pricing table. CPU, RAM, disk and snapshot choices can change the complete charge.
- Operating company
- United States
- Service-region evidence
- North America offering; confirm the exact deployment location and available configuration in the console.
- Storage lifecycle
- First 100GB included while running; additional running storage is listed at $0.03 per 100GB/hour; snapshots at $0.05/GB/month.
- Transfer
- Official pricing FAQ states no data-egress charge.
Published catalog
Massed Compute
Representative single-GPU on-demand configurations from the official catalog. Bundled vCPU, RAM and storage differ by row.
- Operating company
- United States
- Service-region evidence
- The pricing table does not attach a deployment location to every row; confirm the selected location before launch.
- Storage lifecycle
- Each listed configuration includes a stated storage allocation; confirm persistence and retained-storage terms for the selected product.
- Transfer
- The official pricing page states no bandwidth overcharges.
Live marketplace—no static rate copied
Vast.ai
No static price is copied into this snapshot because hosts set live marketplace rates. Use the official offer search for the exact GPU, host and region.
- Operating company
- United States
- Service-region evidence
- Global marketplace; country, datacenter and host vary by exact offer.
- Storage lifecycle
- Storage continues while an online instance exists, including stopped states; the rate varies by host.
- Transfer
- Upload and download prices vary by individual host and are charged separately.
Quote required—no public comparable SKU
Cudo Compute
Kept as a neutral research boundary. It is not presented as a directly comparable self-service catalog without a current public SKU and price.
- Operating company
- United Kingdom
- Service-region evidence
- Confirm deployment region, capacity and topology in the provider quote.
- Storage lifecycle
- Confirm storage, retention and snapshot terms in the proposal.
- Transfer
- Confirm transfer charges in the proposal.
RunPod, Vast.ai and Thunder Compute buttons are sponsored referral links; SmarterBuyLab may earn credit or commission. Massed Compute and Cudo Compute remain neutral research entries here. Provider order is not based on commission.
Research boundary: SmarterBuyLab did not run GPU, network, speed, uptime, availability or reliability tests. Rates can change after the observation date, and a listed GPU may be unavailable in your region.
Compute the work that gets billed—not only the work you wanted.
The estimate multiplies the hourly rate by successful runtime, setup or idle time and expected rerun hours. It then adds retained storage and other known costs. Effective cost per productive hour divides the complete estimate by successful workload hours.
Use one workload definition across every provider.
Keep the same model, precision, dataset, context, batch, output target, checkpoint plan and successful runtime. Change only the provider-specific inputs. A marketplace listing and a managed platform are not fairly compared when one estimate omits setup or recovery work.
Check the line items this calculator cannot discover.
- Network transfer, public IP or endpoint charges
- Volume, snapshot and object-storage billing after shutdown
- Taxes, currency conversion and minimum billing units
- Request, cold-start or provisioned-worker terms for serverless products
- The value of engineering time and failed experiments
Frequently asked questions
Why is completed-job cost better than hourly GPU price?
A completed-job estimate includes setup, idle time, retries and retained storage. These costs can reverse the apparent ranking shown by hourly price alone.
How should I estimate retry overhead?
Use results from a representative short run when possible. For a new workload, model more than one scenario rather than treating a guess as a guarantee.
Does this compare serverless GPU pricing?
It models time-based GPU rental. Serverless services may charge by request, execution time, provisioned workers or other units, so translate the live terms into a matched completed-workload estimate separately.