The large SOTA models have hit very diminishing returns on further scaling, I think. So you’d rather double the number of models you can run in parallel.