How long will that $4.40 rate persist? Until we know more about the real unit economics it will be damn near impossible to rely on steady inference costs or make them predictable at the enterprise level. Gonna be a wild ride for awhile.
Deepseek's 0.86 or whatever is likely subsidized but alternate providers offer it for a price comparable to glm-5.2.
They have published tons of articles dedicated to performance and efficiency engineering. Feel free to have a look...
Does ananyone outside deepseek have a working code for the v4 compressed attention mechanism?
Has any other provider managed to bypass CUDA and program the compute engines in their native assembly language to get 10% more performance out of them?
There is your answer.