That argument doesn't make any sense to me - by the time you are in your dynamic compiler, and you have an IR in something like sea of nodes, what difference does it make in what format the program was represented on disk? The stack part of it has long gone by then.
You can claim that the JVM runs faster than Dalvik though, because it has a better JIT that has had years more to improve.
But I don't know why JVM is being compared against Dalvik when ART is where Android is going in the long run, and has AOT compilation that beats Dalvik's JIT.
The great stack vs. register-based debate is, in my mind, somewhat like the CISC vs. RISC debate—a fun water cooler conversation, but not too relevant in practice. All performance-focused language VMs lower into a (usually SSA-based) IR soon anyway, so the difference only ends up mattering for VMs that are deliberately kept simple for code size, simplicity, or legacy reasons (like Lua, or CPython).
To say that Linux or C is stack-based is nonsensical, as it applies to an instruction set architecture. Virtually all hardware ISAs you're likely to encounter are register-based these days.
The virtual machine itself is a complete abstraction. Modern JVMs do not run a simulation of a stack machine at runtime (after compilation, that is).