If the traffic is write-heavy, like a forum or blog with active commenting, you'll need a beefy database to handle all of the concurrent writes. Same for the interactive game.
For read-heavy sites like the video site, you would have to use a CDN to keep that sustained video transfer rate and to reduce latency for users. For the product listings or the news site, you could get away with a lighter server setup and use static caching and CDN content mirroring to ease the load on your servers
It would also depend on whether this traffic is regular throughout the year, or does it spike at predictable times (think about an Apple product blog after WWDC or a coupon blog on Black Friday). If you have regular, predictable traffic, you can buy or lease dedicated servers...for traffic to spikes, look into cloud based servers like Amazon EC2 where you can scale up and down at will.
At my current job (a media company that operates 850 radio stations in the US), we just have a few loadbalanced frontend servers with squid and memcache and a few database servers behind Akamai Edgecast CDN mirroring. All of the traffic is read-only (articles, transcripts, news), so no actual users ever hit our servers, except to sign up for newsletters. Show audio and video are streamed by Akamai or another provider.
Also check your SQL, often a lot of cpu cycles are wasted on shoddy database interaction.
1) You want to move static data off your system onto another site like s3.
2) You want to do whatever you can async. ie. if somebody shares data, you don't have to make it available immediately. Just place it in a queue and have a second system go through and deal with it.
You pretty much have to be using a non-blocking event-based server to achieve this level of concurrency in an efficient way. Forking/threading processes will never perform anywhere near as well.
Some non-blocking event-based frameworks: Node.js/Twisted/Tornado/AnyEvent/Libevent
You probably want to create a program that all clients connect to and acts as a coordinator for the whole system. Any actual heavy lifting can be done on other servers (CPU/IO intensive tasks, etc). Each web client can establish one connection to your "coordinator" server process and additionally make whatever HTTP requests are necessary to save/fetch data, etc.
This was historically known as "The C10K problem" as in concurrent 10,000 connections. Modern hardware and epoll/kqueue make 10K pretty easy in many cases.
