Well, no Twitch isn't the largest. NBC/Comcast moved several orders of magnitude more of live video for the last olympics, and those numbers went up substantially this year+.
Live presents interesting challenges yes, but in some ways it is easier than VOD from a scale perspective. In live all users are watching essentially the same content++, so it is nearly perfectly cacheable in memory. That means you don't really need a ton of cache, just a cdn with lots NIC's, ram, and CPUs. You can use mid-tier caches to "fan" out data to edge caches and progressively deload origins. You can take a look at our architecture over here https://www.bizety.com/2015/07/15/deep-dive-comcast-cdn-arch....
The really hard problems with live are actually the surrounding items. If you are placing personalized ads, and 100 million people are watching a stream, your system must support 100million transactions per second. If you have QOS data from clients, you need to support 100 million TPS. Etc, etc.
Building live packagers and segmenters is also a serious bit of work. Especially to "broadcast" standards. This requires the coordination of multiple video streams subject to byzantine failure.
I guess the hardest scale problem though is probably emergency alerts. In this situation all users will tune at almost precisely the same time to a live streaming event. 30million + users can arrive on a single (previously cold) item in +
(sauces so I don't get in trouble, full disclosure I designed and work on the systems that do this):
http://nbcsportsgrouppressbox.com/2012/08/14/ondon-olympics-...
http://www.fiercecable.com/special-report/judges-tally-nbc-s...
++
http://www.cs.umd.edu/~slee/pubs/iptv-sigmet09.pdf