AMD, Apple, NVIDIA and those AWS chips use TSMCs 7nm process node which has a ~50% efficiency advantage over Intel's 14nm (the actual nm isn't comparable though).

Debating which ISA enables higher performance designs has been shown in the last 25 years to be a lot like debating which JIT API will result in faster execution.

No, it's also about system and memory interconnects, caches, various accelerators, core configuration, and indeed process nodes as mentioned above. I don't think there's anything inherently faster in one ISA than the other.

There can be ISA-level differences: Itanium’s VLIW structure meant a lot of memory was storing NOPs filling unused slots unless your work aligned perfectly with the CPU and you had a brilliant compiler.

These days, I think you’re right with the possible exception of density: if you have comparable resources, vector extensions, etc. an ISA which uses memory more efficiently is going to get a little more out of its cache and memory bandwidth, but that’s probably not going to be huge.

Then I fail to understand how is Intel's perf/watt behind some ARM vendors. They probably have the best talent in the industry.