I think it's more a problem of GPU capacity than costs. Training takes a lot of resources, inference too.