Microkernels are slower than a monolithic kernel like Linux,

Twenty years ago I might have agreed with the basic premise but argued the a properly designed system with a fast zero copy IPC mechanism was worth the small cost.

OTOH, Linux has become massively bloated as it attempts to fit into every nook and cranny. Optimizations that work in one environment simply don't make sense in others. Beyond that computing hardware has changed in 20 years. These days having a bunch of cores is normal, along with scattering them everywhere to manage power/clock or dedicate them for IO is common.

Further a very large number of the highest performing applications are utilizing some kind of kernel bypass. The "standardization" of DPDK despite its shortcomings has brought hundreds of applications out of dark and pointed out the folly of trying to fit everything into a posix view of the world.

Ah the typical microkernel FUD.

UNIX clones, Linux and BSD variants are the surviving monolithic kernels.

On embedded space, most high integrity OSes are microkernel based, Apple and Microsoft OSes are hybrid variants, and as shown at this year's WWDC in a couple of years Apple OSes will only allow for user space drivers.

Even Android now uses a mixed model with userspace drivers (Treble compliant ones) and the standard Linux drivers (Treble legacy mode drivers).

It's not FUD. Measure system call latency on Fuchsia and Linux for a simple read(), and you'll see Fuchsia is slower. It's the context switch that kills you, not the copies. Of course you could use FUSE on Linux and it would be just as slow.

Now go do same with INTEGRITY, QNX, L4 and plenty of others used in production on real time scenarios where Linux is found lacking.

I will, because I am genuinely curious. However the original question was about Google data centers, which do not care about real time. Real time means timely not fast.