TSMC can’t do it either without xUV lithography machines made by ASML in the Netherlands.
Furthermore there isn’t anything magical about about the current generation of chips that couldn’t be replicated at at a scale of 12 or 15 or 20 nanometers - it’s just that scaling down to that small allows for a greater density of transistors per wafer and thus increased power efficiency. An AI supercomputer could be built with chips with bigger transistors than 3nm it would just run hotter.
And investing in intel aside, one of Nvidias great competitive moats is CUDA and that’s software not hardware.