Guidance on AWS DeepSeek AI Pricing and deployment options

As a newcomer to AWS AI services, you have several options for deploying DeepSeek models on AWS, each with different features and cost considerations. Here's an overview of the deployment options and associated pricing:

1. Amazon Bedrock Marketplace: This is a straightforward option for quickly integrating pre-trained DeepSeek-R1 models through APIs. Pricing is based on usage, and you only pay for the compute resources you consume.

2. Amazon SageMaker JumpStart: This option allows you to deploy and manage DeepSeek-R1 models with just a few clicks. It's suitable for those who want a balance between ease of use and customization. Pricing is based on the underlying EC2 instances used for deployment.

3. Amazon Bedrock Custom Model Import: This option is for deploying DeepSeek-R1-Distill models. It offers more flexibility and control. Pricing is based on the number of active model copies and their duration of activity. Factors affecting cost include model architecture, context length, region, and compute unit version.

4. Amazon EC2 with AWS Trainium/Inferentia: This option allows you to deploy DeepSeek-R1-Distill models on specific hardware for optimal price-performance. Costs are based on EC2 instance pricing and vary depending on the instance type and usage duration.

When considering costs, keep in mind:

• Larger models like DeepSeek-R1-Distill-Llama-70B offer better performance but at a higher cost.
• Smaller versions, such as the 8B model, might be sufficient for many applications at a lower cost.
• Consider implementing monitoring with Amazon CloudWatch to track usage patterns and performance.
• Use AWS Cost Explorer to monitor and manage your expenses.

To optimize costs:

• Start with a lower concurrency quota and scale up based on actual usage.
• Evaluate your use case requirements carefully to choose the right model size.
• Consider using Amazon SageMaker Neo to optimize model performance and efficiency, potentially reducing compute costs.
• Implement auto-scaling for your endpoints to dynamically adjust compute resources based on incoming traffic, helping you pay only for the resources you're actively using.

Remember that pricing can vary based on factors such as region, specific services used, and your overall AWS usage. It's recommended to consult the AWS pricing pages for the most up-to-date and detailed pricing information for your specific deployment scenario.
Sources
DeepSeek-R1 models now available on AWS | AWS News Blog
Deploy DeepSeek-R1 distilled Llama models with Amazon Bedrock Custom Model Import | AWS Machine Learning Blog
Deploy models for inference - Amazon SageMaker AI
Deploy DeepSeek-R1 Distilled Llama models in Amazon Bedrock | AWS Machine Learning Blog

I am wondering why there is no token-based pay-per-use pricing for DeepSeek models. is AWS considering such an option?