If you know that the XML has been generated in a particular way, you can often beat any compliant XML server by employing knowledge of the exact structure.

E.g. you can sometimes skip over chunks of characters without every accessing them, and get speedups of magnitudes over even "just" checking every byte in the input.

There's nothing a faster proper XML parser can do about a custom parser like that.

(Obviously this is a brittle solution and a last resort optimisation, and should be accompanied by ample warnings, but sometimes there are no alternatives)

Surely the only way you could possibly do this is if you controlled both sides - in which case you could achieve the same speedup less brittlely by switching to a binary system like Protocol Buffers?