Because the LLMs were trained on average software which doesn't give a shit about performance and they'll revert to the mean for sufficiently complicated projects. While they can help with searching for performance improvements, they also are liable to cheat if you aren't reviewing them for correctness of any algorithmic changes. Also you need to be responsible for teaching them the right benchmarks to use, since, again, performance optimization has been niche and not the standard.

Mass refactoring will always compete with mass new product launches and mass new feature production, so it's likely to always lose.