GPU cloud fees: storage, egress and job cost
Compare data transfer, storage and billing rules alongside the GPU rate. The table records provider documentation; "not published" means exactly that, which is itself worth knowing.
GPU job cost calculator
Compare compute, storage and egress for one job. Unknown fees stay unknown. Rates below the form are editable only in your browser.
Choose a configuration and confirm its node size and fees.
Only unambiguous free-transfer or single-rate fee cells are prefilled. Conditional storage classes, free allowances and transfer destinations need a quote; enter applicable rates to complete the estimate. Minimum deposits, taxes, public IPs, licensing, support and regional surcharges remain outside this calculation.
Billable runtime = runtime rounded up to the billing increment. Compute = billable hours × node GPU count × per-GPU rate. Storage = GB × monthly rate × retained hours ÷ 720. Egress = outgoing GB × rate.
GPU provider fee comparison
Verified cells link to their source. Green egress cells = flat free egress (6 of 18 providers); conditional or in-region-only free is not highlighted. Corrections: [email protected]. Licensed under CC BY 4.0 with attribution.
Data egress costs: what leaving the cloud costs
Data egress cost is the fee for moving data out of a cloud: downloading trained model weights, exporting datasets, or streaming results to another provider. Ingress (uploading in) is usually free; egress is where the surprises live. Of the 18 GPU clouds in the matrix, 6 publish flat free egress, 6 charge for it or attach conditions (per-GB rates, included-bandwidth caps, or free only in-region), and 6 do not publish a rate at all. "Not published" is worth treating as a question to ask before you commit, not as free.
| Provider | Data egress cost | Data ingress cost |
|---|---|---|
| RunPod | free | free |
| Lambda | free | free |
| Hyperstack | not published | not published |
| CoreWeave | free | free |
| Nebius | $0.015/GiB Standard Object Storage egress; VM network egress not published | not published |
| DataCrunch | not published | not published |
| Voltage Park | free | free |
| Crusoe | free | free |
| Jarvislabs | free in-region | not published |
| OVHcloud | free except some S3/object storage cases | free |
| Spheron | not published | not published |
| Microsoft Azure | tiered; outbound charged at normal rate | free |
| AWS | $0.09/GB first 10TB after free tier | free |
| Together AI | not published | not published |
| fal | not published | not published |
| Modal | not published | not published |
| Vultr | $0.01/GB over included bandwidth | free |
| Massed Compute | free | not published |
Same verified cells as the full matrix above, sliced to transfer fees; each verified value links to the provider page it was read from. Why it matters: a big checkpoint or dataset leaving a per-GB-egress cloud can cost more than the compute that produced it, and the gap between a free-egress neocloud and a hyperscaler's tiered egress compounds with every export. If your workflow moves large artifacts between clouds regularly, weight the egress column as heavily as the $/hr sticker on the live price table.
Why this matters more than the $/hr
Three patterns to watch. First, egress: moving a trained model or dataset out of a cloud with per-GB egress can cost more than the training run. Second, storage while stopped: some providers keep billing for attached storage when the instance is off. Third, billing increments: per-second billing makes bursty work dramatically cheaper than per-hour billing at the same $/hr rate. The calculator and the live table cover the compute side; this page covers the rest.