I totally missed the fact that FileIO is exclusive to one thread
That's not actually true. The kernel is totally free to re-order I/O requests. And if you use BeginRead()/EndRead() to do asynchronous I/O, it's even very explicit that file I/O is threaded, because different worker threads will call your IAsyncResult completion function for different requests.
However, that's not what causes problems with multi-threading loading. The problems are:
1) The content manager uses a dictionary from string to asset, to return the same object the next time you ask for it, without going back to disk. This dictionary is not thread safe. I don't think the content manager wraps it in locks, either.
2) The content manager has to create graphics resources when it loads textures and models. This calls into the graphics device. While the graphics device is somewhat thread safe for each individual call, depending on what the specifics of your content are (especially for custom content types), this may be a problem. Graphics device resets happening in the middle come to mind (only a PC problem, though).
3) If you're loading from a DVD, the seek time is terrible. You really want to order all your loads to be sequential. While there may be multiple I/O requests outstanding, in the end, there is only a single laser reading the disk. For hard disks, depending on what your disk geometry looks like, you may get parallelism out of the disk(!) but that's only possible on PC, not on Xbox, which has a very low-end single hard drive set-up.
So, putting I/O onto a separate thread to not block the main thread is a good idea, assuming you can manage the thread problems of the content manager dictionary and the graphics device. Spreading I/O across multiple threads is only a good idea if you have a studly I/O subsystem (which only happens on PCs, and only some PCs at that). However, if there is significant processing required as part of creating the asset (again, typically with custom assets), putting that processing onto separate threads will help parallelism.