Actually, cray is making a come back. They just had an article on the NY times about that.
One thing crays can do that the clusters can't is deal with really big numbers and problems that can't be divided into smaller parts. Also, crays are easier to program. I guess people are having a hard time programming cluster computers.
Snippet: (article is old enough now that you have to pay to see the whole thing)
Scientists in government, industry and academia are reviving late Seymour Cray's elegant approach to building ultra-fast computers; Burton J Smith, founder of Tera, which bought original Cray Research in 2000, has become industry's leading champion of Cray's design philosophy; Cray's custom machines are known as vector supercomputers and have special hardware intended to handle long strings of numbers in complex scientific computing problems; had been largely replaced by massively parallel processors, designs based on linking thousands of inexpensive, mass-produced microprocessors; such MMP machines are now proving to have limits...