This makes the assumption that their bottleneck is the overhead of HTTP. This assumption is common on HN, but (in my experience) is a rare problem to have for most companies.
It looks like they do a good job of tackling performance problems as they see them, even at the architectural level. My guess is they didn't tackle this because it wasn't a problem worth prioritizing compared to the compounded shitshow of a monolithic Rails app.