Back in September HN had a great discussion on how different DeepSeek providers have very different performance:

https://mmoustafa.com/blog/so-you-want-to-use-openrouter/

And I was speculating that it was likely due to hosting differences like this. It would be fascinating to see if this would improve, for example, tool calling, something that third-party DeepSeek providers never do as well as DeepSeek themselves.