Lot of FUD here.
1) Top500 supercomputers are huge clusters. Linux scales well on clusters. HPC clustered workloads run number crunching for loops on a same set of grid points, over and over again. Everything can be run in parallel in the cpu cache. No synching among cpus. Everything run independently on separate cpus. One cluster such as SGI UV3000, serves one scientist at a time, doing 24h calculations. These are scale-out servers. Linux scales excellent on scale-out servers.
Business workloads such as SAP, OLTP databases, etc communicate lots among cpus and synchronize a lot when serving thousands of users. All these thousands of users do different things simultaneously (pay roll, accounting, etc), so you can not cache the workload in a cpu cache. You need to go out to RAM all the time. Typically, RAM is maybe 100 ns or so. This corresponds to a 10 MHz cpu. You remember those? When you do business workloads (serving thousands of users) you scale up to 16 or 32-sockets tops. No more. Clusters are totally useless because when you synchronize among node to node on a network, performance drops much much much much more. You need to have all the cpus on the same bus, i.e. one single huge scale-up server.
SGI explains this well:
The largest scale-up server on the market is a 64-socket Fujitsu SPARC M10-4s, it has 64TB RAM and runs Solaris. The largest Linux scale-up server is the new HP Kraken which has 16 sockets. It is a redesigned Unix Integrity server which scaled up to 64-sockets when sporting Unix/RISC. Other than the HP Kraken, all the rest of the x86 servers are plain vanilla 1-8 sockets. And Linux scales awful on 8 sockets. Why? Because Linux kernel devs dont have access to 8 socket x86 servers. So how can Linux scale on 8-socket servers when no developer can optimize Linux for 8-sockets? Linux kernel devs have 1-2 sockets tops. And on 1-2 sockets Linux scales well. There is no way Oracle can sell large business servers or database servers, if Oracle switches to Linux. Only Unix and Mainframes have large scale-up servers, serving thousands of users.
2) It seems that Linux on Exadata scales to 2-sockets T7 SPARC cpus. This only proves my point. You will never see Linux on large 16 sockets M7 servers, because Linux maxes out on 2-socket SPARC T7 cpus.
3) Oracle has invested tremendously in SPARC. In five years, Oracle has released six SPARC cpus. Each generation has always been twice as fast as the previous (except S7 which is a crippled M7). Intel generations are 10% faster than the previous generation. The worlds fastest cpu is SPARC M7, with 30ish world records:
Even the new coming POWER9 will not be able to beat the M7. T7 is the same cpu as the M7. I doubt Oracle will kill SPARC, as it is 2-3x faster than POWER8 and Intel X86. And we all know that Linux can not drive a 16-socket SPARC, you need Solaris for that. If Oracle kills Solaris, then Oracle is stuck at 1-2 socket SPARC M7 cpus. Nothing larger. And the big money is in large business servers. And Oracle does big money, not tiny money.
4) Fujitsu use SPARC for their large M10-4S business servers. These compete with Mainframes. The new Fujitsu supercomputer will use ARM cpus. Each ARM is very weak, but that is not important because what is important in supercomputers is low watt usage. 1MWh costs 1 million USD year. Also, supercomputers do not run vanilla Linux. They strip out everything of Linux until a minimal skeleton kernel remains. If you can optimize 5% then the entire supercomputer runs 5% faster. IBM Blue Gene which ran Linux, actually used Linux only for I/O to distribute the workload to the nodes, and then each node ran a special light weight OS that can only do number curnching and nothing else.
5) Unix is dwindling. Linux is eating the Unix cake. This is true. However, Oracle engineered systems running Unix or Linux are growing very rapidly. This is also true. Linux can only take you to a few sockets. The largest engineered systems has to be Solaris/SPARC. Oracle has the entire stack: cpus, OS, middleware, database, ERP software. This means Oracle can make huge optimizations, for instance, DAX database accelerators in the cpu, can be utilized by the OS and by the middleware and the database. This means a SPARC M7 is typically 10x faster than x86 on database workloads, because of DAX, OS and middleware. If you own the entire stack, you can do this.
Oracle needs Solaris for the largest workloads. With Linux Oracle is stuck with a couple of sockets. Look at the SAP benchmarks. It's Unix/RISC at the top with the largest workloads. Linux/x86 at the bottom with small workloads.