I've heard this same question a dozen times a dozen different ways.
First, internal sessions and external sessions are handled similarly, but are not the same. For every internal session, there could be multiple external, and vice versa. What you are talking about isn't logically impossible, but does require a bit of programming and prep. Some devices are being built with the ability to aggregate ethernet connections or wifi for faster network speeds internally on a network, but for external networking, I've only found devices with FAILOVER, which means they only switch when they have to. However, you could use a basic switching (if statements with 0 or 1) in the connection protocol that redirects traffic over a different gateway for each EXTERNAL session. You would then have to keep track of each EXTERNAL SESSION attatched to each INTERNAL SESSION, by wrapping a sub-session inside a main session with a sessionID. You would then have to have a way to check for the site domain or use the sessionID to complete the internal routing (in other words, you'd need to be able to differentiate messages from the client for each external session\connection based on which site they were requesting data from or sending data to so that the pipe for each site would be set, keeping the site from rejecting the pipe). This means you'd have to build a new network protocol into your router, and decide how to differentiate the connections (if only done by site domain, it might be doable from the router, but the client may try to use the ipaddress of the external gate for the website, which negates this; if done on both ends, the client needs to keep track of some identifier, typically a numerical value of binary masking, that allows the send recieve to be matched against the external session being routed on main gate). In other words, you'd have to program your own protocol to handle the networking, and you'd have to decide if you want to build it yourself (making use of both client and host programming), or if you want to build something that's compatible with older existing programming (which would need to be bound by the messaging between host and client, putting more wear on the host but nothing new for the client). If you know your unix, or you know your winserver programming, this can be done with just a bit of time, but it requires a larger amount of resources be allocated for each device being balanced.
For larger corporate style networks, you can mesh the networking, and give each floor or department a dedicated gateway, allowing for multiple ISP's, and never putting too much on any one of them. You can also have Failover handled by a hub that sends changes or redirects to a different gateway when one of them has trouble. This offers some fault tolerance.
