I'm using it to build a world server for a multiplayer game. A lot of people suggest Scala/Java, Erlang, or Node.js for this, but I'm using Go. Go has a good, efficient memory model that lets me simulate a heck of a lot more of my world per megabyte of RAM. In today's world of hosting, that translates into a huge impact on how many dollars per user/hour we need to pay to run our game. That can have a direct translation into what kinds of business models can, and cannot, work for a game design. I've actually written rudimentary versions of my world server in all of the languages I mentioned above, and Go is by far the best in terms of what I can do with memory.
Go makes it very easy to communicate between nodes, almost as easy as what you can do in Erlang. Doozer is a library in Go for coordination of the activities of many machines, a lot like Zookeeper. Between Doozer and Netchan, Go is a really solid choice for making a system like a world simulator that needs to be distributed among many computers.
Go is also better than Node.js for things that require manipulating binary data. For example, I have some communications back and forth where the server is in Go, the client is in Python, and they talk using Protocol Buffers. Javascript doesn't do binary (at all? not very well?) like Go or Java can. There is an experimental protocol buffer library in Javascript, but in general Javascript needs to kind of bend over backwards to work with binary data. It's possible to do it, but it's not natural or efficient.
In general, if you care about memory usage, working with binary data, tons of simultaneous threads/actors, building distributed systems, then Go is a great choice, especially if you want to do more than one of the above.
Used it for my Master's thesis research. It turned out pretty well; Go hits a sweet spot between systems programming and higher-level languages. Anecdotally, it performed about 2/3 as fast as C in stress tests, which was easily worth the productivity boost (especially GC).
I chose it because the native code interface is so nice, so it's trivial to use libraries like ZeroMQ without the overhead you'd see out of something like JNI.
I'm using it to do http traffic analysis on captured packets... about 2.5 million packets a second at peak. This is to provide realtime charts for our web service which didn't have realtime reporting hooks.
I am also using it for HTTP request reconstruction and retransmission in real-time to enable live-data and full load testing of pre-production code.
soon to be launch a vertical content repository, purely GO code (leveldb inspired database with fulltext, ranking and highly scalable, crawler, webserver, load balancer, etc.) with only dependency being libxml2.
Used it experimentally to port a python tool that does some log aggregation with several options (read backwards, query on fields, etc) and returns the results in JSON.
A straight line by line port (ie less idiomatic Go) runs at about twice the speed of Python and with the same memory footprint. With the right fine tuning should be even better.
It was also very easy to do -- i.e code translated quite easily from Python to Go.
It looks intriguing but I can't think of any obvious immediate applications (I'm sure there are plenty, I just don't see them).