QuickStart is free. Charges apply only when you consume computing resources for training or deployment.
Billing overview
QuickStart provides a no-code interface to train and deploy machine learning models. The service itself — including model selection, parameter configuration, and monitoring dashboards — incurs no charge.
Costs arise from two independent sources:
Training resources — billed through Deep Learning Containers (DLC) when a training job runs
Deployment resources — billed through Elastic Algorithm Service (EAS) while a model service is in the Running state
If you use Object Storage Service (OSS) for data storage or Elastic IP Address (EIP) for network access, those services are billed separately.
What's free
| Item | Free |
|---|---|
| QuickStart service | Always free |
| Model selection and parameter configuration | Always free |
| Monitoring dashboards | Always free |
| System disk (public resources) | Up to 30 GiB |
| System disk (dedicated resources) | Up to 200 GiB |
Billable items
| Stage | Underlying service | What you're billed for | Pricing reference |
|---|---|---|---|
| Model training | Deep Learning Containers (DLC) | CPU, GPU, memory, and disk I/O consumed during training jobs. Billed by instance type and duration. | Billing of DLC |
| Model deployment | Elastic Algorithm Service (EAS) | Instance runtime, system disk usage beyond the free quota, and dedicated gateways (if configured). | Billing of EAS |
Training costs
When you train a model through QuickStart, the system submits an asynchronous training job in DLC and schedules compute instances based on your selected algorithm and dataset size.
Billing formula:
Training cost = (Instance unit price ÷ 60) × Training duration (minutes) × Number of instancesBilling stops when the training job completes, fails, or is manually terminated. Paused jobs do not incur charges.
Three factors drive training costs:
| Factor | Impact |
|---|---|
| Instance type | GPU instances (T4, V100, A100) cost more than CPU-only instances. |
| Training duration | DLC bills by the minute, so longer runs cost more. |
| Parallelism | Distributed training multiplies costs by the number of workers. Use distributed training only when datasets or models are too large for a single instance. |
For detailed pricing by instance type and region, see Billing of DLC.
Start with smaller, cheaper instances to validate your workflow. A failed training run on an A100 GPU instance is expensive — catch configuration errors on CPU instances first.
Deployment costs
When you deploy a model through QuickStart, the system creates a service in EAS and provisions instances to handle inference requests. EAS charges for the entire period the service is in the Running state, regardless of request volume.
Billing formula:
Deployment cost = Instance unit price × Runtime duration (hours)For dedicated resources, billing continues even if no services are deployed on the instance.
Four factors drive deployment costs:
| Factor | Impact |
|---|---|
| Resource type | Public resources use shared infrastructure and cost less. Dedicated resources offer performance isolation but cost more. Use public resources for development and testing; reserve dedicated resources for production workloads. |
| Runtime duration | Charges accumulate while the service is running. |
| System disk | Storage beyond the free quota (30 GiB for public resources, 200 GiB for dedicated resources) incurs additional charges. |
| Dedicated gateway (optional) | Enables custom domain names and advanced access control. Incurs separate charges if configured. |
For detailed pricing by resource type, instance specification, and region, see Billing of EAS.
Stop or delete idle services to avoid unnecessary charges. EAS bills continuously while a service is running, regardless of traffic volume.
Cost estimation example
Scenario: Train an image classification model on 10,000 images, then deploy it as a real-time inference service.
| Cost item | Details | Amount |
|---|---|---|
| Training (one-time) | 1x ecs.gn6i-c4g1.xlarge (Tesla T4 GPU), 2 hours at $0.80/hour | $1.60 |
| Deployment (monthly) | EAS public resources, 2 vCPU + 4 GiB memory, 720 hours at $0.05/hour | $36.00 |
| First month total | $37.60 | |
| Subsequent months (no retraining) | $36.00 |
Actual costs vary by region, instance type, and usage patterns. Use the PAI console cost calculator or refer to the linked billing documentation for precise pricing.
Cost optimization
Stop idle services. EAS charges continuously while a service is running. Stop or delete services when not actively serving traffic.
Start small for training. Validate your workflow on smaller, cheaper instances before scaling to GPU instances. A failed training job on an A100 instance is expensive; catch configuration errors on CPU instances first.
Use public resources for development. Public resources are more cost-effective than dedicated resources. Switch to dedicated resources only for production workloads that require performance isolation.
Avoid unnecessary distributed training. Distributed training multiplies costs by the number of workers. Reserve this for datasets or models that cannot fit on a single instance.
Monitor usage in the PAI console. Track training duration and service runtime. Set up alerts to catch long-running jobs or forgotten deployments before costs accumulate.