We have been looking at SAP HANA - their in memory column store database. The demonstrated performance for a 6TB database was phenomenal - for reading... but we've a lot of questions re read/write and getting those log entries to disk just as quickly. This has been designed from the ground up for in-memory optimisation, including the application sitting on top.
This may be a way to go for reporting, analytics, planning and heavy business planning processes, but moving an ERP system into memory even on highly resilient IBM POWER7 architecture is a scary concept.
Another question for these systems is 'how long does it take to load 6TB+ into memory on restart' or DR?
Microsoft Research is also working on Trinity, an in-memory Hypergraph database which looks very nice as an in-memory, up to 100 node, object store - backed onto SQL for data integrity, storage and reporting. Its code is similar to MSAGL from some time back for visual graphing, so should be very accessible when stacked atop MS SQL.