Summary Bullets:

• DeepSeek-R1 Distill Llama and Qwen models became available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart in February 2025.
• DeepSeek-R1 became available as a fully managed, generally available, serverless model in Amazon Bedrock on March 10, 2025.

Amazon Web Services (AWS) just became the first cloud computing provider to offer DeepSeek-R1 as a fully managed, serverless large language model (LLM) in general availability on Amazon Bedrock, a service that provides access to third-party foundation models. The publicly available DeepSeek-R1 series has taken the generative AI (GenAI) market by storm, with its advanced reasoning capabilities and reduced computing costs, made possible thanks to an innovative architecture based on mixture of experts technologies.

This is a logical move by AWS, which has been offering access to foundation models by other providers since early 2023, when Amazon Bedrock was launched. Back then, Amazon started out hosting models by a roster of third-party companies, including startups such as AI21 Labs, Anthropic, and Stability AI, on its cloud platform. It also launched its own proprietary technology, the Amazon Titan foundation models, which was also made available through Amazon Bedrock. The Amazon Titan family has never been a particular strength when compared to foundation models by rivals such as Google Cloud Gemini. For this reason, Amazon chose to focus on the managed service which, being integrated with the rest of the AWS cloud platform, had a massive addressable market. Since then, the company has diversified into another proprietary brand, the Amazon Nova family.

Although Amazon customers needn’t worry about technical setup or maintenance because it is fully managed by AWS, with technology by DeepSeek, the question of security and compliance could potentially be more problematic. According to Amazon, users can rest assured that the inbuilt security features, including data encryption and strict access controls, are sufficient. They include tools to maintain data privacy and regulatory compliance. Regardless, Amazon Bedrock customers are recommended – when using any AI system on the platform – to automatically set up safeguards such as Amazon Bedrock Guardrails, which also can prevent hallucinations.

Many of AWS’ competitors made DeepSeek-R1 available on their platforms weeks ago; Google Cloud, Nvidia, and Microsoft among them. Never one to miss a revenue-generating opportunity, Microsoft went a step further by becoming the first company to implement a per-token pricing structure for DeepSeek-R1, while Google Cloud users pay only for the computing resources they consume, rather than the amount of text the model generates. The latter is a more widespread pricing structure when it comes to open-source AI. By incorporating DeepSeek’s models, Microsoft is following a strategy to diversify its AI offerings and reduce over-reliance on OpenAI’s ChatGPT, focusing on a diversity of internal and third-party AI models to strengthen the Microsoft Copilot suite. The Microsoft Azure AI Foundry currently boasts an array of 1,800 models, and DeepSeek-R1 is also available on GitHub Models free experience.

However, the inclusion of DeepSeek-R1 in Amazon Bedrock as a fully-managed service is significant because it is offering users the opportunity to leverage the technology with built-in security, observability, monitoring, and customizable guardrails. Customers have AWS’ infrastructure and are able to use the model on a serverless pay-per-token basis, helping them scale from experimentation to production. Amazon revealed that thousands of customers have deployed the DeepSeek-R1 model using Amazon Bedrock’s custom model import feature since the system was first hosted on AWS in late January 2025.