TL;DR me if you like, but this is important
I'm not convinced the previous commenters really get it. It's nothing to do with Intel, in fact , they stand to gain a lot.
Telcos spent years building robust networks in a highly regulated domain. In exchange for the pain and complexity, they got to have virtual monopolies on their services, and a whole raft of behemoth suppliers arose to feed them: Alcatel-Lucent, Ericsson, NSN, Cisco etc. These guys sold custom gear at high prices and it worked - five nines, six nines etc.
Now, the telcos face a truly existential crisis: over-the-top providers have the money, lack of regulation and smarts to utterly destroy them, and they know it. When you have ubiqitous internet, why would you use a phone at all, compared to gtalk/whatsapp/skype and so on? The OTT guys have leveraged the cloud's economies of scale and the web's development practices to spur product innovation that is orders of magnitude faster than telcos can move. All built around cheapo x86 boxes from white-box vendors.
So one day the telcos get together and say, "we want some of that". The question is not, "why are they doing NFV?" but more "what took them so long?"
However, before NFV hype takes off, virtualizing a function (in the lingo, building a VNF as part of an NFV deployment) is only one small step. The hardest challenges are yet to come, and to be honest, it is too early to say whether the industry can handle them. A couple of examples:
- To deploy VNFs, scale them, and manage their lifecycle, you need an orchestrator. That toy-like GUI you get on Amazon EC2 is fine if you only have one app (function), but not if you have hundreds and you need to chain them together to build your service. There is currently a yawning chasm in the market for orchestrators - everyone agrees you need one, but no one has one.
- virtualization is just dandy when you are operating in the signaling domain: if your host doesn't schedule your VM for a few milliseconds, you will be OK. In the data plane, for things like transporting voice/video media, or transcoding, this sort of delay is disastrous. So you need to find a way to drastically improve networking performance on virtual machines. So far, the leading candidates are SR-IOV and Intel's DPDK, and neither is a slam dunk.
- OpenStack, which the industry is converging on as the standard cloud stack, has weak networking performance, and it's going to take quite some time to improve it. Icehouse in 2014 is just the start.
- sticking an app on a VM is not NFV, although it is a starting point. True NFV requires that your app be able to take advantage of the cloud and do things like scale elastically, store state in a cloud-friendly manner (eg cassandra type storage), etc. There are very few such apps. I think Nominum have a DNS one, and there's an open-source IMS core ("Project Clearwater").