That's what we're doing now, just backing everything up through the guest OS. What we're really looking to try to move toward is file-level recovery of the individual guest OS's without starting so many threads to get pulled through the ESX NIC's. I'd rather control the traffic flow at the ESX level instead of possibly kicking off so many clients that might all happen to live on the same ESX box that it diminishes your ability to back the OS's up in the most efficient way possible. Ideally I'd really like to snap the data off (and still be able to do file-level recovery) and then NDMP it to tape.
I believe the new 6.5 Enterprise clients can do this using VCB's but we're also evaluating NetApp OSSV to keep the load as close to hardware as possible.