I am not as optimistic that locally run models will be able to compete anytime soon. And even if, the price to run them means you have to buy the compute/gear for a price that is likely equivalent to a lot of 'remote' tokens