The compiler simply interprets this as calling printf with two strings as arguments (but see Zack’s comment).
This happens at compile time (i.e. the compiler does this):
The strings ("%%%s" and "hello") are copied directly into the executable, the compiler leaves them as-is.
This happens at runtime (i.e. the C standard library does this when the app is running):
printf stand for ‘print formatted’. When this function is called, it needs at least one argument. The first argument is the format. The next arguments are “arguments” to this format. They are formatted as specified in the first argument.

About optimization
I have written an example and ran Clang/LLVM with -S:
$ emacs printftest.c $ clang printftest.c -S -o printftest_unopt.s # not optimized $ clang printftest.c -S -O -o printftest_opt.s # optimized: -O flag
C code
#include <stdio.h>
int main() {
printf("%%%s", "hello");
return 0;
}
printftest_unopt.s (not optimized)
; ...
_main:
pushq %rbp
movq %rsp, %rbp
subq $16, %rsp
movl $0, %eax
movl $0, -4(%rbp)
movl %eax, -8(%rbp)
xorb %al, %al
leaq L_.str(%rip), %rdi
leaq L_.str1(%rip), %rsi
callq _printf ; printf called here <----------------
movl %eax, -12(%rbp)
movl -8(%rbp), %eax
addq $16, %rsp
popq %rbp
ret
.section __TEXT,__cstring,cstring_literals
L_.str:
.asciz "%%%s"
L_.str1:
.asciz "hello"
; ...
printftest_opt.s (optimized)
; ...
_main:
pushq %rbp
movq %rsp, %rbp
leaq L_.str(%rip), %rdi
leaq L_.str1(%rip), %rsi
xorb %al, %al
callq _printf ; printf called here <----------------
xorl %eax, %eax
popq %rbp
ret
.section __TEXT,__cstring,cstring_literals
L_.str:
.asciz "%%%s"
L_.str1:
.asciz "hello"
; ...
Conclusion
As you can see (in the __TEXT,__cstring,cstring_literals section and the callq to printf), LLVM (a very, very good compiler) does not optimize printf("%%%s", "hello");. 🙂