GPU Pricing
Find Cheapest GPU prices across multiple providers and deploy instantly
Forgetting to stop the box is the real cost
Idle auto-stop is opt-in — turn it on per deployment with a storage bucket selected, and it's off unless you do. Idle past 30 minutes gets you a warning; idle past 60 saves your environment to that bucket and stops the box, on the providers where we can issue a remote stop. On the handful that don't support one, idle auto-stop can't step in and you'll need to close the deployment yourself.
What the price includes — and how billing works
The number in the table is a per-GPU hourly rate, but you're not billed in hourly chunks. Metering runs per second while the box is active: billing starts once the box is actually ready (not from the moment you hit deploy, and not while it's provisioning), and stops immediately when a stop request goes out — yours, or idle auto-stop's — rather than waiting for the box to finish tearing down in the background.
There's no separate storage charge. Snapshotting your environment — packages, weights, config — so it survives a stop or a move to a different provider is part of what you're paying the GPU rate for, not a metered add-on.
See how idle auto-stop works for the full mechanics of protecting a box you forgot to close.
Worked examples
These use the live per-GPU rates from the table above, so the dollar figures move with the market. Each assumes the box runs for the stated duration with nothing else billed — no storage fee, no setup fee.
You're metered per second the box is active, not rounded up to the next hour — these totals are exact for the stated duration, not an estimate.
Pricing FAQ
What does the price per hour actually include?
The rate on the GPU, nothing else. There's no separate charge for storage — no per-GB or per-snapshot fee — and no setup fee. Your environment (packages, weights, config) is what survives a stop or a provider switch; that's part of the platform, not a line item.
When does billing actually start and stop?
Billing starts once your box is up and ready to use — not from the moment you click deploy and not while it's still provisioning. You're metered per second it's active, and stopping it (manually, or via idle auto-stop) stops the meter immediately when the stop goes out, not once the box has finished shutting down in the background.
What happens if I forget to stop it?
Turn on idle auto-stop for the deployment (it's opt-in, off unless you switch it on, and needs a storage bucket selected) and 30 minutes of idle GPU gets you a warning, 60 minutes saves your environment to that bucket and stops the box. That save-and-stop works the same way on every provider — it's a create-and-terminate call, not a provider-specific stop API. Either way it only protects you from your own idle box — it can't recover one a provider force-terminates out from under you (spot reclaim, node failure); that's only ever as good as your last snapshot.
Can I pause a box instead of losing it?
Pause snapshots your environment and tears the box down, then billing stops. The guarantee that matters is that your environment is captured and can be restored onto a fresh box on any provider — Resume is the one-click path to that, renting a fresh matching box and restoring your snapshot onto it, so it isn't the same physical machine coming back to life. That doesn't depend on which provider you're on — it needs only a create call and a terminate call, which every provider we support implements. The real requirement is having automated backups turned on for the deployment (a storage bucket selected).
What GPU models are available for rent?
H100, A100, H200, B200, AMD MI300X, RTX 4090/5090, and other high-performance GPUs for rent across multiple providers and regions — see the live table above for current availability.
