I'm actually surprised if incurring 50% extra hardware costs really is cheaper than the cost of being down. If Netflix is down for a few hours, then it costs them some goodwill, and maybe a few new signups, but is the immediate revenue impact really that great? Most of Netflix's revenue comes from monthly subscriptions, and it's not like their customers have an SLA.
Actually, they do. and Netflix proactively refund customers for downtime. Usually it's pennies on the dollar, but i've had more than refund for sub 30 minute outages which have prohibited me from using the service.
Netflix are very very sensitive to this problem because it's much harder for them to sell against their biggest competitor (local cable) since they rely on the cable to deliver their service. If the service goes down, then the cable company can jump in and say, "You'll never lose the signal on our network" -- blatantly untrue, but it doesn't matter.
When you're disrupting a market, remember that what seem trivial is in fact hugely important when you're fighting huge well-established competition :)
I'd imagine that part of this cost is reputation. The only problem I have ever had with Netflix streaming is when an agreement runs out and the pull something I or my wife regularly watch. (looking at you, "Paint Your Wagon")
I have not had a single service issue with them, ever. They do a better job at reliably providing me with TV shows than the cable company does. That seems to be where they're looking to position themselves, and the reputation for always being there is hard to regain if you lose it.
There isn't a 50% extra hardware cost. You spread systems over three zones and run at the normal utilization levels of 30-60%. If you lose a zone while you are at 60% you will spike to 90% for a while, until you can deploy replacement systems in the remaining zones. Traffic spikes mean that you don't want to run more than 60% busy anyway.
I don't think the cost of expanding to other regions/AZs is necessarily linear such that adding a zone would incur 50% more costs. Going from one zone to two would probably look that way (or even one server to two), but when you start going from two to three or even 10 to 11 then the %change-in-cost starts to decrease.
This is even more true if/when you load balance between zones and aren't just using them as hot backups. As another commenter pointed out, Netflix says they have three zones and only need two to operate.
