[mutkach]

There's also an estimation of how much a KV cache grows with each subsequent token. That would be roughly ~MBs/token. I think that would be the bottleneck