The argument made by Nvidia's Steve Scott is fairly well reasoned with a discussion of inherent performance boundries with Intel's MIC and/or hybrid computing in general. Overall though his message really just underscores several fundamental points about high performance computing:
1). All computations and algorithms are different with respect to the degree of inherent parallelism and required interprocess communications.
2). No highly parallel computer/design will deliver ideal performance, or be the most energy efficient on all problems
3). The key to high performance computing is to match the right architecture to the problem to be solved. Some architectures trade off ultimate performance on a narrow class of probolems so as to deliver very good performance on a wider class of problems.
4). Intel's past efforts in the supercomputer industry have taken the top spot in absolute speed (ie; Intel Supercomputer Systems Division) so they know how to do this. However, Intel has always put more emphasis on how to bring "supercomputer like" performance to the mainstream market that leverages their manufacturing strength.