CFD is good at using big machines "efficiently", but the cost of DNS scales as the cube of the Reynolds number which will never be tractable for most engineerin

CFD is good at using big machines "efficiently", but the cost of DNS scales as the cube of the Reynolds number which will never be tractable for most engineering problems. Apart from niche basic research on the edge of tractability, all the effort goes into modeling (RANS, DES, wall, etc.) to deliver statistically calibrated estimates of functionals of interest at feasible cost. Those methods actually don't "scale as well" (though the state of research is ahead of commercial software), but also don't need to because they can solve the problem in less time with less hardware. This situation is actually pretty similar to your DL analogy where more hardware provides diminishing returns for solving the actual problem.