I agree that there are very little gains to be had by further shrinking the lithography (e.g. going from 8nm process to 5nm to 3nm etc) as we're reaching the limit of how small each transistor can be with current technology. There can still be absolute performance gains (by simply making the die larger, more transistors) but not so much performance-per-watt gains (which depends on the node size among other factors). That is until inevitably a whole new technology comes along (e.g. quantum computing, or computing based around photons rather than electrons) but we're talking decades away for that.
So I'd expect to see performance, power consumption, and cost continue to rise over the next few years, but we are reaching the end of big improvements in terms performance-per-watt as far as the silicon is concerned. Future advances in software/firmware may alleviate those limitations as you suggest.
Also bear in mind that our goals and Nvidia's goals are not the same, and often oppose each other. We want faster hardware, better performance-per-watt and better value, all they want is more profit, they don't particularly care how.