It's possible that the demands on the model are growing faster than it gets efficient though. A lot of the improvement in output quality/metrics come from more

It's possible that the demands on the model are growing faster than it gets efficient though. A lot of the improvement in output quality/metrics come from more churning and passes. You could argue that a RTX5090 needs less power than a 1996 3Dfx Voodoo 1 to render GLQuake at 320x200/60fps, but it's rarely being used for that.