I am constantly amazed at how the efforts of the HPC crowd are not being backported to the rest of business. Things like infiniband, external PCIe, heterogeneous computing (GPU + CPU), RAM-only storage, etc. All of these things that were originally tested, designed and implemented in the HPC realm are now the watchwords of the day for high performance business platforms.

From a software realm, I think there needs to be a radical shift in how people perceive the problems that they write software to solve. We have single systems with 40 cores and 80 threads, and yet, there isn't really a workload that could take advantage of all of that, especially when usually your access to the storage solution backing it will end up single threaded (for example, linux and irqd).