As long as nothing's blocked, latency could go up 'a lot' (sub-ms -> ms, maybe 1ms->2ms with batching) without meaningfully impacting overall throughput.
I can definitely see millions of networked memcache calls being a bottleneck, and if the batching adds another ms per req on average, but removes the bottleneck, then they can serve a lot more users at a cost of 1ms per req.
Is there anything in TFA that would support my theory? I don't know. I don't care enough to endure InfoQ. (I did for a Rich Hickey talk once, lo these various months, and yea it were a minor inconvenience).