The same reason that pie charts are not very effective - it's relatively hard to make comparisons between angles instead of lengths like in a bar chart. It's probably even harder with a doughnut.
That's a compelling article and I think I'm going to be a lot more cautious about using pie charts in the future, but the thin-doughnut design they're using here looks OK to me.
When you're summarizing coarse relative size across arbitrary categories, bar charts and tables give your selection of categories too much weight: everyone looks at the biggest bar or the biggest number and it's really hard to measure a long tail distribution visually.
If the message you're trying to send is, "nothing but X and Y matter, everything else combined is tiny", I still can't imagine a better visualization than a pie (or a doughnut).
You're absolutely right. Everyone harps on pie charts for being awful for element-to-element comparisons. This mainly stems from Tufte's work and experiments from Cleveland and McGill.
However, in part-to-whole relationship tasks, pie charts can outperform bar charts, since in bar charts there is no true "whole".
Here is a PDF article by Stephen Few about pie charts: http://www.perceptualedge.com/articles/08-21-07.pdf