FTR, I see about 1-1.5ms improvement (using timeit to do a check_call) out of ~32ms total startup time.

I also don't actually see gdi32 being transitively loaded as claimed in the Twitter thread, even back to 3.8. So presumably there's something else going on to cause that issue.

Either way, this is an easy change with non-zero benefit (at the very least, it gets rid of a dependency we are better off avoiding these days).