The main reason is that nginx is much better and faster at handling certain things than Python:
* handling HTTPS and serving plain old HTTP to the application server so Python doesn't have to worry about it
* doing the gzipping of content before it goes out
* routing requests to different places/ports based on various elements matched in the URL or HTTP headers
* virtual hosts, i.e., "Host" header matching and routing things to the right place based on that
* various request sanitization, like setting client_max_body_size, ignore_invalid_headers, timeouts, etc.
Historically we've also had multiple types of application servers, some Python and some C++, and nginx routes requests to the right app server (based mainly on URL prefix).
We also use nginx to do GeoIP with the GeoIP nginx module (though arguably that would be just as simple in Python).
Edit: Note that we don't use it because our "application server is slow" (it's not). Also, I know some people use nginx to serve static content, because it's usually much faster/better than say Python at doing that -- we serve static content via Amazon S3 and a CDN, so that's a non-issue for us.
In most cases reverse proxies are useful as application firewalls, where you control what passes and what doesn't in an application-independent way (i.e. your systems administrators can do this without the need to touch applications - which in many enterprise settings can't be easily touched by developers, much less by operations).
This is why I have yet to use nginx, and stay with good-old apache. Apache is extremely configurable (not only is mod_rewrite very powerful, but you can insert your own request mangling scripts for that weird edge case).
Apache is also good enough for most cases. For applications with a few thousand users hanging on the site all day, apache can handle it just fine in a 5-year old, low-end Celeron rack server, with just 5% CPU usage even with all connections being SSL for both Internet and application server traffic.
Caching and load-balancing are nice things to have, but not the reason for having a reverse proxy in most cases.
I guess a TCP load balancer could do them based on IP as well, but IPs don't identify sessions as well as cookies can.
Also, there are many that use nginx as the load balancer (including me). Would a TCP load balancer be better? Possibly but it also comes at an additional cost whereas nginx is free and far more flexible with routing HTTP content.
I'm not even going to summarize all the other things nginx can do via modules (Lua, memcached, etc -- http://wiki.nginx.org/Modules).
Why not just have your load balancers (which can operate cheaply at the TCP layer) throw traffic directly at your application servers?
If you need caching, that's cheap to do, too. If you need static file serving, can't you another load balancer end-point that points directly at static content servers, or make your application servers faster?
Is nginx primarily useful for slow application server runtimes that can't keep up with what nginx can do?