It's all three of the things you mention: peer to peer, stateless, and on-demand. It also uses UDP, so it's not opening huge numbers of TCP connections. The only time it uses TCP is if you can't use UDP -- as a fallback mode.

I'm not sure why that wouldn't scale. Even if you had millions of users on a network you'd only be connecting to those with whom you're communicating.

It is also as you mention a star of stars network. The supernodes are at very high bandwidth sites and their number and size can be increased on short order. They only relay data if P2P/NAT-t fails, which only happens for about 1-2% of users. Otherwise they just shepherd NAT-t. They're geographically distributed for high performance: Singapore, Tokyo, San Francisco, New York, Amsterdam, and (soon) Sydney. If a supernode fails it takes 10-30 seconds to fail over.

I have plans to further decentralize and add automatic promotion of nodes at some point in the future, but that's a hard problem that requires more study.

Edit: this isn't just another VPN. This is the result of over four years of work, including a huge amount of research into networks and cryptography. It's a completely new system.