And it's both cumulative across all those languages (see above), cheap/amortized (if you can do a good multilingual NMT for 50 languages, how hard can 50+1 languages be?), and many of those languages are likely to grow both in terms of sheer population and in GDP. (Think about South Asian or African countries like Indonesia or Nigeria.) The question isn't why are FB & Google investing so much in powerful multilingual models which handle hundreds of languages, but why aren't other entities as well?