I agree. Also, if you gave tons of compute power and DNS using Navier Stokes, it may feel that with CFD provides an answer; although it is approximate, it is a decent one. To me, the issue is every single time we do practical CFD of a whole city or an automobile etc. , we need specialized models that fit the flow regime. Even with petascale compute power that most supercomputers in the world are equipped with, once we simulate Navier Stokes + a turbulence model + some custom boundary condition + other microscale model, the approach is not generalizable and needs deep knowledge.