Discovery and subscriptions will need to be solved, but that’s not insurmountable (ie RSS). CDNs can be plug and play in front of nginx (if P2P torrent isn't your cup of tea), backend storage can be whatever is cheapest.
It's one thing make a service for few 100 users, it's a different beast to serve millions of users. In case of Online Videos, it is not only extremely difficult, it still remains very expensive. There is literally a handful of companies who can financially and technically pull it off at youtube-scale. After more than 10 years, youtube either makes very little profit or still not profitable.
The more popular the content, the cheaper per unit it'll be to serve (based on bandwidth/transit costs and how they scale, and that you most likely can serve it out of RAM with Varnish, a very inexpensive CDN, or peer to peer), but there will still be a cost, and either the producer or the consumers of that content will need to pay for it.
Only a few companies that can do it centrally. Peer to Peer technology was doing it before youtube existed.
I can give you a list of torrents that are dead and never coming back. Other than takedowns, there's some very old YouTube videos still around.
I don't understand why you are mentioning peer to peer technology? What does that have to do with anything? Youtube is not a cute side project where your viewing experience depending on how many peers your video has. Peertube depends on webtorrent, you literally have to host and seed every single video on your platform and duplicate it multiple times by others. Good luck doing it over petabyte scale. Google does something similar, but popular videos are hosted at users-end point on GGC nodes, distributed over 10,000s of servers all over the world.
The super-seeding / swarming approach developed for the BitTorrent protocol is quite interesting. One original seeder just needs to upload to a few leechers. Then this leechers become seeders themselfs and start to upload to a next layer of leechers and so on. This approach scales exponentially meaning you can cover the whole internet in just a few steps. Further more you can start uploading yourself even before you have completed receiving the file.
E.g. if every node just has 10 uploading connections then you just need 8 steps to cover 100 million = 10^8 peers. Remember the traffic on the original seeder (and every other node) is just 10 uploading connections.
1. YouTube Red so I don't have to see ads
2. Queuing and casting from my phone to my TV.
3. Fast loading and buffer-free streaming (as long as your bandwidth is good enough). Note I'm in Australia - Mixer and Twitch often aren't smooth for me even though I'm on 100 megabit fibre. Sydney or Melbourne PoPs will beat Hong Kong and Singapore every time.
I watch Pycon and LCA videos on YouTube because it's easier than going anywhere else.