for us the jit/runtime overhead is the real killer for density. we run ~10k gVisor pods and the base memory footprint of the runtime determines how many we can bin-pack per node much more than raw cpu throughput.