They invented a hybrid attention design that drastically reduces the amount of memory needed for the KV cache at inference time. Like, dividing it by 10. And memory is a large part of the cost of inference.

The American models are eating their cost big time, in return they get user data to train on and a massive reality distortion field that can theoretically be exploited later. It costs less for DeepSeek to eat their cost, and maybe the value of that user data is worth it now? Maybe there is some Chinese VC getting involved to try and boost DeepSeek with a little reality distortion field they can attempt to exploit later? I don't know, but all these seem plausible to me.

easy: users pay the difference with their data