Skip to main content
Serverless AI does not have its own quotas and pricing; the service applies the quotas and pricing of Compute.

Pricing

Billing of Devlabs, endpoints and jobs depends on whether the deployment is running or stopped:
  • While a Devlab, job or endpoint is running, you are billed for its computing resources and storage under Compute VM and disk pricing.
  • While a Devlab is stopped, you are billed for the entire container disk but not for computing resources.
  • While an endpoint is stopped, you are not billed for computing resources or storage. Jobs cannot be stopped.
If you mount an Object Storage bucket or a shared filesystem to a Serverless AI resource, separate billing for the bucket and filesystem applies, including when this resource is stopped.

Quotas

When you create Devlabs, endpoints or jobs, they count toward Compute quotas. If the quotas are exceeded, resource provisioning fails with the QuotaExceeded error. Before you create a Serverless AI deployment, check the quotas required for Serverless AI and increase them, if required. Devlabs, endpoints or jobs are based on the same resources, so you need the same quotas for them. When you calculate what values for the quotas you need, consider that each Serverless AI deployment includes the following resources:
  • Virtual machine (VM), a regular or preemptible one.
  • GPUs if you use a GPU platform. The platform availability depends on the region.
  • vCPUs. You need a dedicated quota for vCPUs only if you run a VM without any GPUs.
  • Network SSD disk.
  • Public IPv4 address if the deployment configuration includes it.
  • Allocations that represent public and private IP addresses.
  • (Optional) Shared filesystem or bucket if you mount an additional volume to the deployment.
To run multiple Serverless AI deployments, multiply each resource by the number of these deployments.

Quotas to check

Required quotas: Optional quotas that apply when you mount a volume: You also need quotas for size of objects in buckets or for shared filesystem capacity. However, they depend on the size of the files that you’re going to store.

Example

You create an endpoint:
In this case, you need at least 1 VM, 1 GPU for the gpu-l40s-a platform, 1 disk, 100 GiB for the Network SSD disk capacity, 1 public address, 2 allocations and 1 bucket. You don’t need dedicated quotas for RAM and vCPUs because Compute automatically provides them together with GPU quotas. If you add the --preemptible parameter to the command, you’d create an endpoint that runs on a preemptible VM. In this case, you need different quotas: 1 preemptible VM, 1 disk, 100 GiB for the Network SSD disk capacity, 1 public address, 2 allocations and 1 bucket.

See also

For more information about the quotas and pricing, see: