Vast.ai pricing explained: how the marketplace charges in 2026

18 October 2026

This post contains affiliate links. If you sign up via our links we may earn a small commission at no extra cost to you. We run ComputeClarity.

This page explains how Vast.ai prices GPU rental. Unlike a provider with a fixed price list, Vast.ai is a marketplace, so there is no single rate to quote. We give no hourly prices here because Vast.ai's pricing page loads its live prices dynamically and we have no saved source for a figure. What we can set out, from Vast.ai's own pages read on 18 October 2026, is how the pricing works. We run ComputeClarity and earn a commission if you sign up to Vast.ai through our link.

Why there is no single price

Vast.ai's pricing page says prices are set by supply and demand across 40+ data centres, and lists 68+ GPU types, from RTX 3060 to B200. Hosts list machines and the rate you pay depends on the listing you choose. Two offers for the same GPU can differ in price, location, reliability and bandwidth. Compare live offers in the Vast.ai search rather than relying on a number in an article, including ours.

The three ways to rent

Vast.ai's pricing documentation lists three instance types.

On demand: fixed pricing with guaranteed resources, at high priority. The pricing page tags it as the best fit for production.

Interruptible: the lowest cost, but it may be paused because it runs at low priority. The docs say it is often 50%+ cheaper than on demand, which is Vast.ai's own claim, so check the actual offers. It suits batch jobs that save checkpoints and can restart.

Reserved: a discount for pre-payment, with Vast.ai stating up to 50% off on 1, 3 or 6 month terms. The pricing page says to contact sales and shows no figures.

How it bills

Vast.ai's docs say GPU rental is charged per second while the instance is running. The pricing page states there are no minimum hours and no rounding up.

Storage keeps billing

This is the cost that catches people out. Vast.ai's docs say storage is billed continuously while your instance exists, regardless of running state, and that charges continue when an instance is stopped. The docs add that stopped instances typically cost more per GB than running ones, and that rates vary by host. To stop storage billing you must delete the instance, so copy off anything you need first.

Bandwidth

Per the docs, charges apply per byte transferred, for both upload and download, and the rates are shown when you select an instance. Bandwidth varies by host, so check it for each offer, especially if you move large datasets or model weights. Our guide to hidden fees in a GPU cloud quote covers what to ask.

What happens when credit runs out

You must add credits before starting any instance. When the balance runs out, the docs say instances are stopped automatically and the GPU is released. With a saved card, the card is charged to restore the balance and your instance and data stay available. Without a saved payment method, instances and data are scheduled for deletion unless you add credits promptly, and no deadline is given. Keep a buffer and copy important data elsewhere.

Choosing an offer

Decide first whether the job can survive an interruption. If not, use on demand. Compare the total cost, which is the GPU rate plus storage and bandwidth, not only the headline rate. Check host details, reliability and location before you rent, and run a short test before a long job.

To compare Vast.ai with a provider that has a fixed list, read RunPod pricing explained and RunPod vs Vast.ai. For other options, read Vast.ai alternatives and GPU cloud providers compared. You can also compare stored rates in the ComputeClarity calculator.