Hi Clee -
To answer your original question, all of BE's software-based dedupliction (either done on the client or on the media server) is "in-line". There is no need for scratch space or some other space set aside to hold data before it gets deduped.
On to your Hyper-V question. Straight VMDK files will not see the greatest level of deduplication at this time. I'm not sure what level that will be right now, but it will *not* be 10:1 - probably more like 5:1 or 3:1 if I had to guess.
You will see a much higher level of deduplication if you installed a BE Remote Agent in each guest and do client-side deduplication. If you are purchasing the Agent for Hyper-V then you are entitled to run an agent in each guest attached to a specific host. While you have stated you don't want to do that, it is going to be the best way (currently) to achieve high levels of deduplication. Client deduplication's affect on the host (or, in this case, virtual guest) is greatest in the first few backups.
In closing, while your theoretical approach above will certainly work, I don't believe you will see 90% reduction in backup sizes while backing up VMDK's directly. In the BE 2010 release, you should try client deduplication directly from each virtual guest, in order to achieve that level of deduplication.
Thanks,
Aidan Finley
Sr. Product Manager, Backup Exec