FAQ

How am I charged for a GPU?

Per second, at the offer's hourly price, for as long as the instance is running. Provisioning, stopping and teardown are unmetered, so the pull time before your workload comes up costs nothing. Charges settle to your organization's balance about once a minute — see How billing works.

Does a stopped instance still cost money?

Yes. Stopping ends the GPU charge but keeps your disk and your slot on that machine, and the disk bills per GB per hour until the instance is destroyed. A large disk parked for a week can cost more than the GPU hours that filled it. Read Stop vs destroy before you leave anything overnight.

What happens when my balance runs out?

At zero, your running instances are stopped automatically and your disks survive. If the balance keeps falling and reaches −$5, instances are destroyed and the disks go with them. Top up before the second threshold — the details are in Auto-stop and auto-destroy.

Can I pay by invoice instead of prepaying?

No. Superheat is prepaid only: you buy credits, and instances spend them. Packs are $10, $25 and $100, or you can name any amount between $5 and $1,000, and the credits land as soon as the payment confirms. See Adding credits.

Why can't I add credits?

Billing is admin-only. If you joined an organization as a member you can deploy, connect and destroy instances, but the Add credits button is not yours — the balance is shared, and only an admin tops it up. Ask an admin, or check your role under Roles and permissions.

Is the price on a card per GPU or for the whole slice?

For the whole slice. A 4x offer at $9.60 per hour costs $9.60 per hour in total, not four times that. See Offers and slices.

The offer sold out before I could deploy. What happened?

Someone took the last unit first. A card shows how many of that shape are free, but the number is a snapshot — if two deploys race for the last one, the second is refused with OFFER_UNAVAILABLE and you are sent back to the marketplace. Pick another offer — Marketplace explains how to filter down to equivalent hardware quickly.

Note that a card showing 1 of 2 has not sold out: someone else holds one GPU on that machine and the other is still yours to take.

What happens to my data when I destroy an instance?

The disk is deleted permanently. There is no snapshot, no trash and no undo, and the GPUs return to the machine's pool once teardown finishes — so every shape that needs them goes back on sale. Copy anything you want to keep off the instance first — Disk and storage shows how.

Why isn't my SSH port 22?

Because your container's port 22 is published on a port from the host machine's own range, so every instance on a machine gets a distinct one. Use the connect string on the instance's Connect tab exactly as written; it already carries the right port, user and host. See Connect over SSH.

I added an SSH key after deploying and still cannot connect.

The key you select on the deploy page is the one written into the instance. Adding a key to your account afterward does not reach an instance that already exists — attach the key at deploy time, or destroy and redeploy with it selected. Add keys before you deploy, as described in Add a key.

Why does my browser distrust JupyterLab, and where is the token?

JupyterLab is served over HTTPS with a self-signed certificate, so the browser shows a warning the first time. The access token is minted by the platform and never returned by the API, so always open the notebook with the Open JupyterLab button on the instance page — that link already contains the token. See JupyterLab.

Can I run my own Docker image?

Yes. Create a private template pointing at your image, including one in a private registry with credentials attached. The ssh and jupyter launch modes need an image built from the Superheat base image, since a plain upstream image has no SSH server or notebook; args and vm run any image. Start at Create a template.

My onstart script did not run. Why?

Check the launch mode first: onstart applies to the ssh, jupyter and vm modes, and an args template runs its image's own entrypoint instead. The script also runs only once, on the instance's first start, so a stop and start will not repeat it. A failing script leaves the instance up so you can log in and fix it — see onstart.

Can I resize the disk after deploying?

No. Disk size is fixed when you deploy, anywhere from 10 GB up to the machine's capacity, and the only way to change it is to destroy the instance and deploy a new one. Size it once, with the storage charge in mind, using Cost estimate.

Can I automate deploys with an API key?

Yes. An shk_ key can browse offers, read templates, and create, stop, start and destroy instances in the organization it was created in. It always acts as a member, so it cannot buy credits or touch account-level settings such as your SSH keys — add those from the console first. See Authentication.