Thanks for sharing that. It's a reality that is typically avoided by the clickbait YouTubers.
TSMC is investing in much 2nm capability because it's likely their 2nm capability will be used for many years as the standard for processors for many companies.
There are multiple factors in play, The fab equipment ban on Chinese Fab companies is likely going to have a side effect of more of TSMC older business getting eaten away more quickly. There is a very large sizable chunk of TSMC is selling the mature fab processes. If TSMC's competitors can't spend money on new EUV equipment they probably have money to spend money on something else. That something else is TSMC's non 'protected' business.
The second issue is that 2nm 'bake time' complexity is likely going to be in the N3B range, not the N3E. it is just 'better' (more effective) multi-patterning. If getting less finished wafers per month tnen need more machines to hit the same finished throughput rate ( more machines , higher finished wafer costs ). In short, they just need more floor space to do what they did before.
Only gets a bit worse when move to the even larger High-NA EUV machines that are coming ( which TSMC isn't waiting for). To 'print' even smaller you need an even bigger machine than have now. ( somewhat also makes the old facilities less reusable in swapping out old tools for new tools since the tool layout is much different. )
As you noted, the parts of the SoC that deal with the external world is not able to shrink as much as the bit-flipping logic/arithmetic portions.
There are some "backside power" ( PowerVia /Intel ) tricks . But again side effect to make the 'bake time' longer ( and production complexity higher). However, those aren't coming in TSMC N2. A later modifier ( N2_ I forget which letter. ) will pick those up. And the price could creep higher also.
Some progress will come for the other elements , but all three still aren't going to move at the same 'improvement' rate.
It's one reason why I doubt we will see another significant transformation in the small computer industry, as far as processors go.
Intel's Meteor Lake ( multiple chiplet with 3D chip interposer) might be a 1-3 generations 'too soon' for mainstream small computer industry. There are ways around some of the trends. They just don't lead to dramatically lower production costs. The notion that many folks are addicted to ( small computers are going to get cheaper over time).
More than decent change that the 'war' against discrete GPUs in commonly deployed user systems will expand past Apple. Laptops were already 'winning' but some other form factors (not box with slot) will shift the playing field.
And given the current capability of the M3 series chips, I really doubt people will need anything more. "2nm" will not be a revolution but will be a good sales pitch.
The rush to let the computer do the majority of the 'thinking/writing/creating' points otherwise. If the AI/ML push isn't a medium term 'bubble' , the M3 isn't enough. It is enough 'right now' , but down the road if there is enough "local inference' uptick it isn't enough.
And future generation phones are likely going to 'eat into' the laptops in a similar way that laptops' ate into' desktops. ( and in turn how the "attack of the killer microchip' ate into the servers and mainframes the decades before. )
The 'transformation' will be in where the bulk of the customer base went. That "good enough for big majority of basic user's needs" is a threat to the current form factors.
