How do large sites (e.g. Wikipedia) deal with bots that are behind other IP masker? For instance, in my university, everybody searches Wikipedia, giving it a significant load. But, as far as I know, Wikipedia can only know the IP of the university router, so if I set up an "unleashed" bot (with only a small delay between requests), can Wikipedia ban my bot without banning the whole organization? can a site actually ban an IP behind an organizational network?
A site cannot directly ban an IP which is behind NAT. It could act on IPs passed through non-anonymising HTTP proxies - when such a proxy forwards a request on, it typically appends that address to an X-Forwarded-For header, so if access from your private network actually has to go via such a proxy the internal IP could be exposed; however most sites (wikipedia included) wouldn't trust the information in that header anyway because it's easy to spoof to implicate innocent IPs or evade bans.
There are other techniques that attempt to uniquely identify users independently of IP address however. You can interrogate a web browser for a lot of information about it and the system it's running on, such as the user-agent, screen resolution, list of plugins, etc. - see https://github.com/carlo/jquery-browser-fingerprint for an example of this in practice. You could use such fingerprints to control access, though depending on site design you may be able to interact with it without engaging with the fingerprinting process, and even if you can't a bot could provide spurious and randomised data in order to avoid having a consistent fingerprint if you are aware this kind of protection is in place. This method of control also runs the risk of false positives especially when it comes to mobile devices where there will probably be large numbers of clients running identical stock clients on identical stock hardware (most people on a specific model of iPhone running a specific version of iOS, for instance, would probably get the same fingerprint). Fingerprinting like this is normally just used for user tracking rather than to enforce controls but I am aware of places which do use fingerprinting to implement bans when there is concern that an IP block would be too broad, and could be effective against a naive bot.
Generally the IP address isn't sufficient information for a correct ban. So advanced networks work high up the network stack.
A Denial of Service (DoS) attack (which you are worried about creating) is usually handled by rate limiting the initial TCP connection setup. This means legitimate users that are willing to wait will get through whereas those that are just trying to consume server resources are slowed to the point it become harmless. This is where DoS then evolved into a Distributed DoS (DDoS) attack.
Once you have a connection to the server you can make as many requests as you like, the web server administration can configure how many requests to handle.
The web server probably can handle more capacity than your local network gateway anyway, that's probably the limiting factor in your use case. I'd wager your University network admins would come knocking on your door before Wikipedia did.
It is important to be a good Internet citizen so I would add rate limiting code to a bot.
It should also be pointed out that Wikipedia offer data dumps so that trawling the site isn't really necessary.