Danial Abadi did a nice overview of the issues involved with Exadata and shared disk vs. shared nothing last year.
Oracles Hadoop loader does a nifty trick. If the HDFS nodes are on Oracle's HW, then Oracle can read directly in - faster than parallel block load. I wouldn't be surprised if they played the same trick with XFS files and OSS 'R' on the Big Data Analytics machine. This isn't strategy or gamesmanship - there are advantages to shared disk. There are also massive disadvantages - namely cost - of some of Oracle's approaches. But it certainly makes brute force stuff easier. Engineers have a specific architecture to optomize for. Having been part of optimization efforts in the past, it's not hard to spend way more than the marginal cost of Oracles HW to solve some of these problems. Of course, avoiding them all together with a fully distributed architecture is even better - if your application can take advantage of it. All of this is really about using these tools in conjunction with Oracle - if your app doesn't use/need Oracle, it's not all that relevant.