Reza Behforooz, tech lead for the Google Talk service, recently gave a presentation at a Google-sponsored scalability conference in Seattle. In his talk, Reza discussed many of the challenges involved with running a large-scale service, and how Google Talk has addressed these problems. While it doesn't give away all the secrets, the presentation does provide valuable insight into the Google Talk service architecture.
It's also interesting to note that many of the methods used to scale the Google Talk service to millions of online users were the same ones used at AOL to scale up the AIM service. There are a number of minor architectural differences between the two systems (XML vs binary protocol, gatewayed connection to presence server vs direct, for starters), but the essentials are very similar.
You can watch the full talk with Q&A below, or if you'd rather just read the highlights, check out Dare Obasanjo's notes from the conference.