Spot GPU Instances

Up to 50% off – RTX 4090 from only $0.18/hour

Default 1-hour protection, 1-hour advance termination notice, fully compatible with traditional Spot workflows. Best suited for workloads tolerant to interruptions

Deploy Spot Instance View Pricing

GPU Types Available:4090A100H100

节省 50% 的 GPU 成本

为什么选择竞价实例

以极低成本运行工作负载——且不牺牲可靠性

最高节省 50%

Pay half the on-demand rate

内置保护

Guaranteed 1-hour initial runtime

提前通知

1-hour termination alerts via API, email, or console

灵活部署

Combine with on-demand or reserved instances for the right balance of cost and stability

Spot 实例生命周期

保证运行 1 小时，之后可中断，并在自动回收前提前 1 小时通知。Spot 使用闲置容量——成本低，但可能随时被回收。

启动

Instance enters running state

1 小时保护

No interruptions during this period

中断窗口

After protection ends, instance may be reclaimed

终止通知

Sent 1 hour in advance

强制终止

Instance is stopped and released automatically

注意：工作负载应具备中断容忍能力。使用检查点或备份来防止数据丢失。

Ideal Use Cases

Best practices for maximizing Spot

Deep Learning

Short-term training, prototyping, or small-scale inference

TrainingValidationSmall-scale inference

Batch Processing

Cost-effective for overnight, weekend, or large-scale data jobs

Data processingLarge-scale compute

Education & Demos

Classroom labs, technical workshops, and proof-of-concepts

LabsWorkshopsProof-of-concepts

最佳实践

启用检查点

Save training state regularly to persistent storage

设计可重试的工作负载

Break jobs into smaller units with resume/retry

采用混合策略

Run critical workloads on on-demand, offload flexible ones to Spot

快速响应

Use termination notifications to trigger backup or shutdown logic

注意：Spot 最适合灵活的工作负载。对于关键任务应用，请使用按需或预留实例。

APIs 与自动化

通过完整的 API 支持以编程方式管理 Spot 实例

获取价格

Real-time Spot rates

查看可用性

GPU capacity by model

启动实例

Provision Spot with billing_type=spot

处理中断

Subscribe to termination events via API, or email

浏览 API 文档

Start at Half the Cost

Up to 50% lower cost

1-hour guaranteed runtime

1-hour advance notice

Deploy Spot Instance