XML is far more verbose in an uncompressed form, however you really
should be gzipping your data streams in which case, for typical
models, there is very little difference in payload size over the
network at least for larger data sets.
Historically large data sets in json suffered from performance
degradation (on the client) when performing complex sorts/filters/
queries, areas where xpath and compiled XSLT on the shined. In
recent years JavaScript engines have experienced stunning improvement
while xPath/XSLT performance/feature completeness has stagnated, at
least in webkit browsers (IE and Mozilla fare a bit better).
For some historical perspective see:
http://www.nczonline.net/blog/2008/01/09/is-json-better-than-xml/