The software will run on inexpensive dumb hardware from whitelabel manufacturers such as Quanta computers. The secret sauce in hardware that gave Cisco big margins is increasingly replaced by software that is good enough (e.g. You can use SDN controller and a cheap switch hardware to replace very expensive Cisco stuff). The trend in Enterprises and telcos is to use cloud (AWS, GCE, Azure,...) and open source (aka Facebook Open Compute) so the hardware spending has slowed down. (in tier 1 carrier I deal with, all new deployments must be on Openstack that uses whitelabel hardware purchased in bulk). EMC, Oracle, Cisco, HP are facing the impacts of such decisions.
White box switches are all great in the data center if we are talking about a shitload of layer 2 and TRILL, but somebody still has to manufacture the core routers capable of advanced IP layer 3 features at line rate 400-1000Gbps bus backplane per slot. You're not going to whitebox replace an ASR9010 with 3rd generation line cards.
It's a much smaller market than the old "everything with more than one network port should be Cisco" mantra, though. Yeah, Cisco won't disappear any time soon, but their market is shrinking.
And they won't be trusted or used by large carriers for serious POPs for many years. New manufacturers have tried to show up on the market and sell routers that are designed to be highly redundant and offer a full suite of layer 3 feature set to handle full global v6/v6 routing tables. They've either failed or been acquired.
Major ISP backbones are very risk averse, when you're trying to run a six nines availability IP services POP. The routing software needs to be rock solid and mature when you're looking at a twin pair of core routers in a node that might be handling 200-300 Gbps of traffic and critical edge peering/transit connections.
My counterargument is that Tomahawk LC (the third generation LCs referenced by parent) are just NP5 and Mellanox just picked up EzChip. I imagine you'll see whitebox competitors targeting large swaths that use ASR9K for the datapath resources it provides (perhaps Mellanox will even build this box).
Others, such as in my particular space, were previously forced to ASR9K or similar for scale reasons (because scaling full tables is hard when you need capacity too), but Broadcom's Jericho chips hit the scale mark with added TCAM, kicking us from $200K-ish/PoP for barely meeting needs to $60K/PoP way over provisioned in the case of NCS5501 or Arista 7280R. Of course, there are caveats to how this works out, but good enough is good enough!
Customer mix is important for margins. SPs, for example, are notorious for getting bottom dollar deals for their scale and ability to play the other side. And sure, they need the feature set of ASR9K. But when the rest of the mix (higher margin) looks around, they see that the scale promises are unnecessary for their operations and move on to lower cost boxes.
Substitute Python for bash and I get the same sense from Kubernetes.. I'm not sure it's fair to say "before the problem is well understood... But maybe. Whatever Google had going on internally Kubernetes is very much figuring it out along the way at almost every level. If you are wanting to use on AWS then you have kops, kube-up, hyperkube, protokube; a dizzying, un-approachable and interconnected ecosystem of launch scripts/programs. Very strange interop decisions with AWS, but hard to fault because it's darn near one person trying to manage all AWS integrations.
AWS, Azure or GCE....or if you insist on your own hw, just what spit and bailing wire you minimally require... at least you will own and understand it.
What's good about it? It's a bunch of different dumpster fires created by a bunch of committees across a bunch of companies which mostly compete with each other, and as such have no real incentive or motivation to have it all work well together even if they had the ability to, and the ability alone is seriously in question.
I tried to use OpenStack 4 years in our new PaaS product. I ended up pulling it and using the regular, well baked tools built into our managed hardware. I see much hasn't changed in those intervening years.
OpenStack is open source. It can be forked, modified, whatever. Let me know when you manage to build your own AWS cloud with Amazon's software and then I'll entertain the thought that they are comparable.
