Even if you prohibit PCRE2, the .NET submissions using out-of-box Regex engine end up being about 4 times faster than Java.
Surprisingly, even though .NET's BigInteger is known for its inefficiency, it ends up being more memory efficient and marginally faster at pidigits than a Java submission that does not use GMP. The implementations are not line-by-line equivalent so may not be perfectly representative of performance of each BigInt implementation.
My point being - if you look at the submissions closer, the data gives much clearer picture and only supports the argument that C# is a very usable language for solving the tasks one would usually reach for C, C++ or Rust instead.
Sure looks like it's written in Java!
Look at all the programming language implementations that provide big integers by calling out to GMP. Why would it be "cheating" when available to all and done openly? Libraries matter.
If you meant BenchmarksGame, then it's the other way around - Java is most competitive where it relies heavily on GC[0], and loses in other areas which require capability to write a low-level implementation[1] that C# provides.
The only places where there are C calls are pidigts[2] and regex-redux[3] benchmarks, in both of which the Java submissions have to import pre-generated or pre-made bindings to GMP and PCRE2 respectively. As do all other languages, with varying degrees of "preparation".
[0]: https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
[1]: https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
[2]: https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
[3]: https://benchmarksgame-team.pages.debian.net/benchmarksgame/...