Hi The ocr and vote will be in ASM. The ASM diskgroups will actually extend to all the cells across both Exadata machines and hence will be accessible to all servers on the same infiniband network.
It is supported to interconnect Exadata half racks and larger using the included spine switch (the infiniband switch at the bottom of the rack). See appendix G of the latest Exadata Owner's Guide for more information. Basically, you would cable the Exadata racks together and build one large storage grid across the racks. This would create one large RAC cluster, with the 3 default diskgroups (DBFS_DG, DATA, RECO). OCR and voting files would be spread across ASM disks just as they are now.
There are several points to consider about such a configuration:
1- Ib cables can go a 100m or so, so this is not really a stretch cluster, as in any decent data center, 100m apart means 2 rooms.
2- If you setup such a system, you need to configure ASM failure groups yourself, in a way to have the data from the first system mirrored on cells on the second system.
3- if you use normal redundancy, and you loose one rack, you'll end up with the second system having no redundancy at all, and you better rebalance right away to be on the safe side, this means keeping free space to be able to rebalance. If you can afford it, have 3 systems, with high redundancy, so that the los of one keeps the remaining part mirrored.
4- if you do it with 2 machines, better have a 3rd voting on an NFS server somewhere, like you would do with any stretch cluster.
That said, I do not like, nor advise such a setup at all. setting up a DG site for HA is far more easier, useful, robust, .....