Furthermore, I would wager a giant portion of people who have entered the ML space in the last five years started out by using CUDA on their gaming rigs. Throwing away that entrenchment vector seems like a terrible idea.