[adam_gyroscope]

I worked on YouTube transcoding about 12 years ago. First, the scale is mostly reused - what’s doing transcoding now is doing a different compute job later. Transcoding was also done for most videos only on idle compute. Second, Google had 300k+ caches around the world, in many surprising places (buses, cruise ships) as well as many thousands of other larger but not full data center locations; get the content as close as possible to the user. (I imagine now all transcoding from the mezzanine format is done in real time on an edge GPU for all but the most popular platforms and content). Tl;dr: build out a huge amount of infrastructure to serve ads very quickly and you can piggyback video serving on that at little marginal cost.