They don't, cause the speed at which we have to implement new features means cutting corners. We write backends, none of which are performance-critical. A lot of faster algos that would take too long to implement in C++ get skipped, so in the end it's slower than partner teams' Java code that does similar things, but again it doesn't matter as long as nobody notices the latency.