I occasionally get gzipped data back from Wikipedia (even though I explicitly set Accept-Encoding in the proxy) - I think it's due to their intermediate caches not respecting the Vary headers. I plan on adding gzip support soon to rectify this, but until then, you can get around it by adding ?foo to the end of the custom URL.
Curious detail: any idea why searching for "Islam" seems to make things go haywire?
http://www.jasondavies.com/wordcloud/#http%3A%2F%2Fen.wikipe...
Perhaps it's to do with the Arabic script corrupting the SVG somehow?