Servers aren't bought to run SPECfp_rate, and they're not even bought to run workloads that even resemble SPECfp_rate.
You are not helping yourself with such nonsensical absolute
claims like this.
My current employer just bought half a dozen single socket
NGMA based systems for circuit simulation. When we run
corner simulations we often spin off 32 or more simultaneous
Eldo or Hspice runs across the server farm. In my experience
over the last decade circuit simulation performance tracks
SPECfp with a high degree of correlation across various x86
and RISC platforms. Corner sim runs are exactly analogous
to SPECfp_rate - multiple copies of FP intensive programs
running in parallel and competing for memory bandwidth.
So I have a first hand example that blows your claim right out
of the water. Please be careful to note that I am not saying
that my example was more than a niche compared to more
mainstream IT infrastructure, web application etc server uses.
But then again I am not the one making absolutist claims. ;-)
BTW, our current simulation servers are mostly Opteron based.
Market Data 
Markets 
AI
