If it could output 1k tokens per second but needed 4 seconds to produce the first batch of 4k, would that not be viable?