Platform For AI:Billing of QuickStart

Billing overview

QuickStart provides a no-code interface to train and deploy machine learning models. The service itself is free, including model selection, parameter configuration, and monitoring dashboards. Charges apply only when you consume computing resources for training jobs or deployment services.

When you use QuickStart, costs are incurred at two stages:

• Training stage: QuickStart creates training jobs in Deep Learning Containers (DLC) and schedules compute instances to run your training workloads. You are billed based on DLC's pay-as-you-go pricing for the instance type and duration.

• Deployment stage: QuickStart creates model services in Elastic Algorithm Service (EAS) and provisions compute resources to handle inference requests. You are billed based on EAS pricing for the resource type, runtime, and optional components like system disks or dedicated gateways.

Billable items

The following table describes the resources that incur charges when you use QuickStart.

Training costs

When you train a model through QuickStart, the system creates a training job in DLC and allocates compute instances based on your selected algorithm and dataset size. Training costs depend on:

• Instance type: GPU-accelerated instances cost more than CPU-only instances. Price varies by GPU model (T4, V100, A100).

• Training duration: DLC bills by the minute. Longer training runs incur higher costs.

• Parallelism: Distributed training across multiple instances multiplies costs by the number of workers.

The billing formula is:

Training cost = (Instance unit price ÷ 60) × Training duration (minutes) × Number of instances

Billing stops when the training job completes, fails, or is manually terminated. Paused jobs do not incur charges.

For detailed pricing by instance type and region, see Billing of DLC.

Deployment costs

When you deploy a model through QuickStart, the system creates a service in EAS and provisions instances to handle inference requests. Deployment costs depend on:

• Resource type: Public resources (shared infrastructure) or dedicated resources (exclusive instances). Dedicated resources offer better performance isolation but cost more.

• Runtime duration: EAS charges for the entire period that your service is in the Running state, regardless of request volume. Stop or delete idle services to avoid unnecessary charges.

• System disk: EAS provides a free disk quota (30 GiB for public resources, 200 GiB for dedicated resources). Exceeding this quota incurs additional storage charges.

• Dedicated gateway (optional): If you require features like custom domain names or advanced access control, you can purchase a dedicated gateway, which incurs separate charges.

The billing formula for computing resources is:

Deployment cost = Instance unit price × Runtime duration (hours)

Billing stops when you stop or delete the service. For dedicated resources, billing continues even if no services are deployed on the instance.

For detailed pricing by resource type, instance specification, and region, see Billing of EAS.

Cost optimization

To minimize costs when using QuickStart, consider the following best practices:

• Stop idle services: Deployed models in EAS are billed continuously while running. Stop or delete services when not actively serving traffic to avoid unnecessary charges.

• Choose appropriate instance types: For training, start with smaller instances to validate workflows before scaling to expensive GPU instances. For deployment, match instance specifications to your expected request volume and latency requirements.

• Use public resources for development: Public resources in EAS are more cost-effective than dedicated resources. Reserve dedicated resources for production workloads that require performance isolation.

• Monitor resource usage: Use PAI console monitoring to track training duration and service runtime. Set up alerts to detect long-running jobs or forgotten deployments.

• Optimize distributed training: Distributed training multiplies costs by the number of workers. Use this feature only when datasets or models are too large for single-instance training.

Cost estimation example

The following example illustrates typical costs for training and deploying a model using QuickStart:

Scenario: Train an image classification model on 10,000 images, then deploy it as a real-time inference service.

• Training cost: Use a single GPU instance (ecs.gn6i-c4g1.xlarge with Tesla T4 GPU) for 2 hours. Assuming DLC pricing of $0.80 per hour, total training cost = $0.80 × 2 = $1.60.

• Deployment cost: Deploy on EAS public resources using 2 vCPU + 4 GiB memory for 30 days (720 hours). Assuming EAS pricing of $0.05 per hour, total deployment cost = $0.05 × 720 = $36.00.

• Total monthly cost: $1.60 (training, one-time) + $36.00 (deployment, recurring) = $37.60 for the first month. Subsequent months cost $36.00 if no retraining is required.

Related information

• Billing of DLC

• Billing of EAS