Especially the 'backend' languages that do all the heavy lifting for domain-specific software. Just in my vertical of choice, financial software, there are literally billions of lines of Java and .NET code powering critical systems. The code is the documentation, and there's little appetite to rewrite all that at enormous cost and risk.

Perhaps AI will get reliable enough to pour through these double-digit million LOC codebases and convert them flawlessly, but that looks like it's decades off at this point.