Not to mention there are more ways to tell apart browsers than user agent sniffing.
Then I read that IE doesn't like data uris larger than 32k, so I recompressed my fonts with zopfli, but one was still ~35k. I did more digging and discovered that the 32k limit is for IE8; IE9's limit is 4 gig, and everything is fine.
Some recent articles have seriously questioned the wisdom of doing this.
I have to admit I've not read all the articles in full so the question may have been answered. I see Pete did write in a comment to  "I haven't tried data URId fonts or SVGs but those are great ideas for follow on tests"
That is completely false. A CSS file isn't less cacheable if it embeds inline fonts or images.
I'm on a really slow connection and I've seen this maybe once or twice so I'm wondering how big of a problem it really is.
Server: Here you go! Oh BTW you'll almost certainly need the prettyfont.ttf file too. Accept?
"Can I use the fonts in any publication, even embedded in the file
Yes. You may use them like most other fonts, but unlike some fonts you may include an embedded subset of the fonts in your document. Such use does not require you to include this license or other files (listed in OFL condition 2), nor does it require any type of acknowledgement within the publication."
(a) You may Use the Licensed Content to design and develop Your Published Websites or webpages (and must Use any Kit We require for such purposes) and You may reference or encode a link to selected Licensed Fonts within Your Published Website design so that when others view or interact with Your Published Websites, they will see Your content displayed with the Licensed Fonts as You intended; and
(b) You may only Use Licensed Fonts within Your Published Websites as described in this Section 3.1.2.