News

Leading Cloud Providers Offer DeepSeek-R1

• By John K. Waters
• 02/04/2025

Some of the world's leading cloud service providers are now making the open-source DeepSeek-R1 reasoning model available on their platforms. The Chinese startup's release of the model generated splashy headlines and intense media interest because of its ability to leverage more efficient processing and reduce compute resource consumption, which is a key driver of high AI costs.

Amazon Web Services (AWS), Microsoft, and Google Cloud made the model available to their customers last week, but they have yet to implement the per-token pricing structure used for other AI models such as Meta’s Llama 3.

Instead, DeepSeek-R1 users on these cloud platforms pay only for the computing resources they consume, rather than for the amount of text the model generates. AWS and Google have reported that this approach aligns with existing pricing models for open-source AI.

DeepSeek launched its latest DeepSeek-V3 model in December 2024. It was followed by the release of DeepSeek-R1, DeepSeek-R1-Zero, and DeepSeek-R1-Distill on January 20, 2025. The DeepSeek-R1-Zero model reportedly features 671 billion parameters, and the DeepSeek-R1-Distill lineup offers models ranging from 1.5 billion to 70 billion parameters. On January 27, 2025, the company expanded its portfolio with Janus-Pro-7B, a vision-based AI model.

DeepSeek-R1 is positioned as a cost-efficient alternative to proprietary AI models, particularly for companies with large-scale AI deployments. The model was designed to process information more efficiently, reducing the overall compute burden.

However, cloud providers may ultimately profit more from infrastructure rentals than direct model usage fees, industry watchers have observed. And renting cloud servers for AI workloads often costs more than accessing models via APIs. AWS, for example, charges up to $124 per hour for an AI-optimized cloud server, which translates to nearly $90,000 per month for continuous usage. Microsoft Azure customers do not need to rent dedicated servers for DeepSeek, but they still pay for underlying computing power, leading to variable pricing depending on how efficiently they run the model.

In contrast, companies using Meta’s Llama 3.1 through AWS pay $3 per 1 million tokens, a significantly lower upfront cost for businesses with intermittent AI needs. Tokens represent processed text, with 1,000 tokens equivalent to approximately 750 words, according to AI infrastructure provider Anyscale.

Smaller cloud providers, including Together AI and Fireworks AI, have already implemented fixed per-token pricing for DeepSeek-R1, a structure that could become more common as demand for cost-effective AI models grows.

For businesses seeking the lowest cost, DeepSeek-R1 is available via its parent company’s API at $2.19 per million tokens—three to four times cheaper than some Western cloud providers. However, routing AI workloads through Chinese servers raises data privacy and security concerns, making it a less viable option for many organizations. Sensitive business information could be subject to Chinese government regulations, including potential data sharing under local laws. And many organizations are cautious about sending proprietary or customer data to servers outside their jurisdiction, especially in regions with less stringent privacy protections.

To date, AWS, Microsoft, and Google have not disclosed how many customers are actively using DeepSeek-R1.