Mainframes are more expensive per-CPU-unit or per-IOPS. Google has shown that, with the right distributed software, a cluster of cheap machines can be more reliable because you avoid putting all your eggs in the mainframe basket. (The tradeoff is that your software is more complicated and likely introduces new failure scenarios).
So, although mainframes are indeed optimal for today's database, the NewSQL hypothesis is that different database software will be able to run on multiple machines, and so the mainframe will no longer be the optimal solution. This (hypothesized) database would run on a cluster of commodity machines, and should therefore be cheaper and hopefully more reliable. Moreover, just as Google's clusters beat Altavista's mainframe approach, a clustered database should scale to much bigger workloads than the biggest mainframe.
Of course, this relies on a new breed of database.
'Google has shown that, with the right distributed software, a cluster of cheap machines can be more reliable because you avoid putting all your eggs in the mainframe basket.'
'Moreover, just as Google's clusters beat Altavista's mainframe approach, a clustered database should scale to much bigger workloads than the biggest mainframe.'
My one comment on those points is that distributed computing solved Google's problem, and of course applies to what many software companies are trying to do (amazon, yahoo, etc.) But do they have an answer for the set of problems where 100% ACID compliance and high-availability is a requirement?
The ATM machines people use every day to exchange money are not fully ACID compliant, and they do not use "transactions". It's a calculated choice on the design of the system to account for the guarantees in different ways and balance discrepancies that occur after the fact.
If you need full ACID compliance and need to operate at "Google scale", you reorganize the problem so that you don't need the compliance.
The CAP conjecture has certainly been a major barrier to progress in this area ;-)
Oh yes...mainframes that have been developed over decades catering specifically to enterprise needs are far less superior than what a group of 20-something engineers at a social media company came up with.
Excuse the sarcasm, there are certainly places where distributed in-memory data stores on commodity hardware are a fine solution. That does not mean its taking over mainframe computing because that solves a completely different class of problems.