Vast.ai or RunPod: renting a GPU for research without overpaying or leaking data

You rent a GPU when you want to fine-tune a model, run an open model on your own text, or train something that will not fit on a laptop. Both services rent by the hour or less. Prices below are from each provider’s pricing page on 9 October 2026 and move daily.

Vast.ai is cheapest. RunPod’s Secure Cloud is safer for your data.

Use Vast.ai for cheap, restartable jobs on public data. Use RunPod’s Secure Cloud when the job has to finish on schedule or the data should only sit in a data center. Put licensed data on neither until you have read the license.

Per hour, 9 October 2026	Vast.ai, from	Vast.ai, median	RunPod, listed
H100 SXM, 80 GB	$1.73	$2.40	$3.99
H100 PCIe, 80 GB	$2.53	$2.53	$2.89
H200, 141 GB	$4.37	$5.00	$5.29

	Vast.ai	RunPod
Who owns the machine	Hosts on a marketplace	Data centers (Secure Cloud) or vetted individuals (Community Cloud)
What you are billed for	GPU, storage and bandwidth per byte	By the second, no fees for data in or out
Cheapest mode	Interruptible, often 50% or more below on-demand	Savings plans for longer use

Vast.ai

How it works. Vast.ai describes itself as a marketplace that connects users with GPU providers worldwide, with prices set by supply and demand. There are three ways to rent:

On-demand: fixed price, guaranteed resources.
Reserved: a discount for paying ahead.
Interruptible: the lowest price, and the machine may be paused. The documentation says these are often 50% or more cheaper than on-demand.

Prices on 9 October 2026. From the pricing page: an H100 SXM from $1.73 an hour with a median of $2.40. An RTX 5090 from $0.21 with a median of $0.61. An RTX 4090 from $0.14 with a median of $0.49.

Watch the bandwidth charge. The cheapest listing is one host, and hosts differ, which is why “from” and “median” sit so far apart. Your bill is GPU plus storage plus bandwidth. The documentation says bandwidth is charged per byte at a rate set by each host, covers both upload and download, and can change the total a lot for data-heavy work. Check it before you move a large dataset.

RunPod

How it works. RunPod has two tiers. Its documentation describes Secure Cloud as running in T3/T4 data centers, and Community Cloud as connecting individual compute providers to users through a vetted peer-to-peer system. Pods are billed by the second with no fees for data moving in or out, and there are savings plans for longer use.

Prices on 9 October 2026. Listed on the pricing page: an A100 80GB at $1.79 an hour, an H100 PCIe at $2.89, an H100 SXM at $3.99 and an H200 at $5.29.

You pay more for a known price. RunPod costs more than Vast.ai’s floor for the same card. In return the price is fixed and the bill has no bandwidth line.

Licensed data usually cannot go on a marketplace GPU

On a marketplace, the machine belongs to someone you cannot identify. For open models and public text, that is fine. For CRSP, Compustat, vendor transcripts or anything under a data license, it usually is not: data licenses commonly restrict where the data may be stored and who may have access to the hardware.

Decide by what goes on the disk.

What is on the disk	Where to run
Public data, open model weights	Cheapest host that is fast enough. Interruptible if the job can resume.
Your own unpublished results or code	A data-center tier. Delete the volume when done.
Licensed vendor data	Your university’s cluster or cloud account, unless the license names the provider as allowed.
Student or personal data	Not a rented marketplace GPU. Ask your institution.

Checkpoint every few minutes to use interruptible pricing

The cheapest hours are the ones that can be taken away. Write checkpoints to disk every few minutes, make the training script resume from the last checkpoint, and keep the data pull scripted so a new machine can rebuild its inputs. A job that restarts cleanly can use interruptible pricing. One that cannot will cost you the run.