Most deployments have more throughput between the proxy and object tier than between their proxy/load-balancing core's and any one client; so the amplification on PUT is not a limiting factor during upload.

Regardless, the consequences of loosing an object between a "successful" upload and an asynchronous offload is too dangerous. The proxy will even go so far as to "hand off" to secondary nodes if one of the primary locations is not responsive in order to ensure full durability immediately upon response to the client.

I should clarify if it wasn't already obvious that the proxy writes all three copies simultaneously - not "one by one". If any single connection drops out mid-stream the remaining two can continue to service the request successfully.

Other storage systems with different goals have taken the approach you've described - DRDB for example has an asynchronous mirroring mode. Recently Swift added support for write-affinity specifically for handling the case of geographically distributed clusters - but you may be able to leverage it to achieve a similar effect if you have use case that requires it:

http://docs.openstack.org/developer/swift/admin_guide.html#write-affinity-and-write-affinity-node-count