Interestingly, as a tier-2/regional operator, these cache devices are hard to get because they fill a certain role. We have worked with Netflix to try and get the caching device, and it just doesn't do any good if you have less than 3-4gbps of pure Netflix traffic. It does not work because the caches have to ... populate the cache! They do this regularly, and the do it overnight -- but it is an absurd amount of data, especially when there are multiple bitrates. I am told that the cache runs > 1.5gbps to populate, almost nightly. So if you don't push significantly more than that, it is not a cost winner.
As a transit provider/local ISP/bandwidth buyer, 3+gbps is a lot of traffic. We found it mildly more attractive to buy a 10gbps wave to a Netflix-available peering point and peer directly with them than to buy 2+gbps of transit from Level3/Cogent/HE, especially factoring in last mile costs.
Also of note, my own traffic engineering testing shows that Netflix *strongly* prefers Hurricane Electric (as of last fall), then Cogent, then Level3.
There is a really horrible hole between 1gbps and 10gbps of consumption that there isn't a good solution for. Netflix knows about it, but it is a very difficult target to hit -- it may be cheaper to buy transit, or it may not be, but hardware isn't the answer. This same situation exists for all CDNs - limelight, edgecast, akamai, L3.
As usual, peering is the answer. Our customers pay us to bring them Netflix ... so we buy a wave and backhaul it hundreds of miles to satisfy them. It'd be ridiculous for me to charge Netflix when my customers are asking for it!
