I think it would be less bandwidth. Considering the overhead of all the individual HTTP headers both in request and response, you could reduce that significantly by batching up log files and uploading them at regular intervals, perhaps once a day? GA seems to perform batch updates at that interval in any case so it might not even put a big cramp in their style from that perspective.

However, there are a bunch more issues with this. To allow this, they'd have to be able to parse all of the most popular Web server log formats. This, in and of itself, is not that hard until you account for people customizing their logging format via measures like the LogFormat directive in Apache. As well, customizing the log format is a virtual necessity, as GA requires more information than is in the standard log formats of most web servers. Another issue is that the use of a JavaScript snippet on the page prevents non-browser user agents from appearing, clearing GA's logs of spiders and other such cruft that detracts from determining human user behavior (what they are apparently most interested in for the purposes of optimizing AdWords and AdSense).

But the biggest problem for them, I think, is that with this method they are no longer able to set first-party cookies on your user's browsers and thus (more) reliably track individual (read: human) users. So, in effect, they would waste time with this, in that they'd still have to make you put a JS snippet on the page to set cookies that you'd then have to modify your web server logging format to include in every hit. Urchin worked like this back when it was still alive and you could run it on your own servers. So, all in all, it seems like something Google would just be wasting their time in doing given what they want to get out of your data.