Limits

These are the v0.1 customer limits. They can tighten without a docs deploy — treat API errors as authoritative.

​

Runtime

timeout= on @app.function is seconds. The CLI default is 600 unless the decorator or --timeout sets another value. The workload guides use a 30-minute decorator so image pull and a short train still finish.

​

What you get

• One GPU SKU from compute gpu list.
• One process: your function.
• Logs and a JSON return value.
• A receipt after terminate.

​

What you do not get

• A login shell on the machine.
• A machine that stays up between runs.
• A region or instance-template picker.
• Persistent disks.
• A dashboard “Run” button — launch from the CLI.

​

Admission

New creates can be refused when the account already has two active runs, lacks enough prepaid credit, reaches a spend limit, or no provider has capacity. Compute does not have a service-wide one-machine limit: different accounts are not serialized by a single Compute-wide machine slot. Provider capacity is not guaranteed. If a run is refused, send the request id from the error to Support.