Skip to main content
Select capacity only after you confirm that the provider, region, and workload are enabled for your project. Provider adapter code and catalog entries do not establish public workload availability. This guide does not certify any provider for customer execution.

​
Select capacity

Offers describe observed capacity and price. Review the provider, region, GPU model and count, spot policy, customer price, and observation time. Capacity and quota can change after an offer is listed. A listed offer is not a reservation. Use a current quote where the service requires one; an expired quote is not a valid basis for a new request. Provider capabilities differ. Node creation alone does not prove that container execution, storage, networking, or inference works. Confirm support for your image and workload before creation. Do not infer cross-provider image portability, live migration, automatic fallback, or inference support from an adapter name. Raw compute should identify its source provider and region and show the customer rate and any separate software fee. For a managed workload, confirm the actual placement and any provider or region restriction you need with the service. These are selection requirements, not a promise that every product has these controls available.

​
Handle failure

A capacity error requires a new capacity decision. An authorization or quota error requires correction before another attempt. A timeout leaves the result uncertain: first check the existing operation and resource. Keep the same idempotency key and request content when retrying a supported resource mutation. See operation handling. A provider may accept a request before the connection fails. Do not assume that an HTTP error means no resource exists. If the resource remains in an unknown or failed state, retain its ID and operation ID and ask support to confirm cleanup. A single missing observation is not sufficient evidence of completed cleanup.

​
Finish work

Use the supported resource lifecycle action to stop or delete work. Follow its operation and confirm the resource’s final observed state. Stopping compute and deleting compute are different actions. Check separately whether volumes, snapshots, images, or other retained resources remain and whether they incur charges. Do not assume that closing an SDK connection, a local timeout, or a failed workload removes those resources or stops all charges. Provider lifecycle tests and local SDK fixtures do not prove that a production workload completed or that all retained resources were removed. When cleanup is uncertain, resolve the existing allocation before you request replacement capacity.