You are correct that the JVM is ubiquitous in server software but not nearly as much as you assume and there has been a shift away from it in recent years for reasons having nothing to do with Oracle etc. I've been designing high-performance server engines since the 1990s and have watched the architectures and tool chains evolve.
Five years ago most new server engine development was done on the JVM. Since then there has been a shift toward new development being done in C++ such that the vast majority of new server engines I know about are being developed in C++ (mostly C++11) across a diverse range of companies. The reasons are practical and reflect the evolution of hardware.
The short version is that on current hardware C++ is much more efficient, both in terms of throughput per core and can achieve integer factor improvements in absolute performance relative to the JVM. Most of the differences are boil down to two things. First, server performance tends to be bound by memory performance and the JVM is quite a bit worse than what is easily achievable in C++. Second, an optimal high-performance server engine design in C++ is difficult to express within the JVM so basic design of the engine kernel tends to be less efficient as well.
In large-scale systems, that starts to add up in terms of power and hardware consumption and companies are more sensitive to this than they used to be. C++ currently offers significantly better characteristics and using less hardware to do it.
I thought I knew most of what is hot and fresh on the software development front, but I've never heard of any "server engines." Mind explaining what it is? Virtualization containers perhaps?
A "server engine" (usually called a kernel actually) is like a userspace operating system kernel that is purpose-built for a class of server workloads and attempts to optimally manage system resources for that workload. The server application is built on top of that kernel. When people refer to e.g. "database kernels", this is what they are referring to. It reimplements the operating system resource management services through the OS APIs.
Writing an excellent server kernel requires a high level of technical ability and quite a bit of low-level code since you have to reimplement most of the operating system resource management services most developers take for granted. Few pieces of open source server software are built on userspace kernels and the ones that do like PostgreSQL are (currently) only partial kernels that still rely on the OS to do significant things a full kernel would reimplement. A properly designed database kernel, for example, is at least 100k LoC of low-level C/C++ and that is before you actually write the server application that sits on top of it. I've designed and written kernels for both database engines and network engines, it is not trivial.
So why would you want to go through all this effort instead of writing to the standard POSIX APIs directly? Because performance and scalability. For example, a well-designed disk buffer and scheduler for a database kernel can easily triple the I/O throughput possible with a highly tuned POSIX implementation. A lot of locking and blocking, both explicit and implicit in the OS, become unnecessary. Certain kinds of distributed system problems become easier to solve because you can adaptively schedule data flows at a fine-grained level. One of the reasons Oracle and DB2 scale so well and get such high throughput is that they are built on highly optimized userspace kernels.
Virtualization actually degrades the performance of high-performance server systems in part because the hypervisor acts as a primitive operating system underneath the operating system the server kernel can see. This does not impact non-kernel based servers as much.
In the really large server-side systems we're talking about here (banks, defense, etc.) both developer productivity and maximum speed rank way, way below other factors, such as having long-lived processes.
Java's not a clusterfuck? I'll take the clusterfuck that's not controlled by a creepy asshole who thinks the NSA is essential and says stupid shit like "Who's ever heard of government misusing information?"
Five years ago most new server engine development was done on the JVM. Since then there has been a shift toward new development being done in C++ such that the vast majority of new server engines I know about are being developed in C++ (mostly C++11) across a diverse range of companies. The reasons are practical and reflect the evolution of hardware.
The short version is that on current hardware C++ is much more efficient, both in terms of throughput per core and can achieve integer factor improvements in absolute performance relative to the JVM. Most of the differences are boil down to two things. First, server performance tends to be bound by memory performance and the JVM is quite a bit worse than what is easily achievable in C++. Second, an optimal high-performance server engine design in C++ is difficult to express within the JVM so basic design of the engine kernel tends to be less efficient as well.
In large-scale systems, that starts to add up in terms of power and hardware consumption and companies are more sensitive to this than they used to be. C++ currently offers significantly better characteristics and using less hardware to do it.