If only it were that simple ! Clojure has no provision for compile time type checks, unlike Common Lisp running in SBCL. On the other hand Common Lisp running on JVM (see Armed Bear Common Lisp) could theoretically achieve those results because, but if we are including calls to host systems then Embedable CL could better it since it runs on top of C. The optimized Java in the benchmarks you quote in fact looks more like C or Cpp code. Im not informed enough to make a judgement if idiomatic Java even looks like that. Idiomatic SBCL (as opposed to just idiomatic Common Lisp) code does look like the versions achieving best results, because people often use SBCL for compiler optimization

> The benchmarks I could find had Java beating SBCL on most benchmarks,

Most of the benchmarks are actually quite similar - ie within the margin of error. Most of the times when Java did beat SBCL by a somewhat more significant margin I noticed that Java's submission used multi-threading whereas that particular submission for SBCL didn't, so the difference isn't so much as Java vs SBCL, but algorithm a vs algorithm b.

Also, as far as I can remember, it is worth mentioning that there was some controversy over some SBCL submissions. One of the submitters had solutions which saw SBCL beating C, Cpp, Zig, and Rust on some of the numerically intensive benchmarks, however their submissions were rejected later for using a SIMD library which at the time was not yet part of SBCL, but now it is.

The point is that if everyone used more efficient languages, a very large reduction in energy costs could be expected (not just in data centers by the way, every device takes power, and the ridiculously inefficient apps people are running probably cost a lot of energy). It's the same argument as to why you should have a ecofriendly house or whatever: you try to do your part and hope the others will do too. You may be cynic enough to think that makes no difference, but you're just blind to your own unjustified prejudices.

I guess we are interested in real life application performance, not in low-level benchmarks or benchmark games.