I wonder how big Twitter's infra costs are (capex). Based on the article, they need to build capacity > 2x peak loaf to tolerate one of the DCs failing. But if they instead had 4 and still only tolerated 1 failure, they'd only need capacity > 1.33x peak load.

This doesn't account for the extra overhead associated with extra DCs, but it seems like there's opportunities for major effeciency wins.