I am using wolfram desktop, but I assume it’s the the same for Mathematica. It’s a universal binary, so if you right click on the bundle in your applications folder and select “Get Info” there is a little check box to run it using Rosetta. It will run the whole thing using the Intel code path.
You also might be able to launch a single kernel using Rosetta with custom kernel configs and then connect to it from your main kernel to run just your optimization routine, but I have not tried that.
The other option you have is to download the arm Mac version of mosek directly. The new compilers functionality provides ways to access the C library functions directly without having to write your own LibraryLink c code. I was able to get a toy linear optimization problem running like this, but it was a lot of work, but you lose the (in my opinion, immense) benefit of being able to model your problem symbolically and leverage Wolfram’s automatic DCP support, and my problem is conic and would require bridging a bunch more mosek library commands so I gave up on this route. If you were to reimplement the external solver interface like this you’d basically have a version of MosekLink with drop-in library support.
Currently, MosekLink looks like it relies on a precompiled LibraryLink bridge to Mosek, so you can’t just update the mosek c library. I think your best bet is to use Rosetta for now and hope that they update MosekLink soon. The interface is the same, and it’s pretty simple, so presumably they just need to recompile the library link shared object and distribute an arm version of that and mosek in the paclet and it should “just work” (assuming no library link weirdness on Mac arm.)
Hope that helps.