Description
the supply is constrained and thus the prices rise. So going back to DeepSeek, the KV cache only needs a fourth of the HBM and then they also need 1/8 of the SSD storage. Well, they're able to serve many more tokens without necessarily having to pay for all that memory. The…