Comment by cardine

I have not found this to be true at all in my field (natural language generation).

We have a 7 figure GPU setup that is running 24/7 at 100% utilization just to handle inference.